SRB 136 Aug13 Grana

BAYESIAN INVERSION METHODS
FOR SEISMIC RESERVOIR CHARACTERIZATION

AND TIME-LAPSE STUDIES
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF GEOPHYSICS
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Dario Grana
August 2013
c Copyright by Dario Grana 2013

All Rights Reserved
ii
I certify that I have read this dissertation and that, in my opinion, it
is fully adequate in scope and quality as a dissertation for the degree
of Doctor of Philosophy.
(Gary Mavko) Principal Adviser

(Tapan Mukerji)

(Jack Dvorkin)
Approved for the University Committee on Graduate Studies
iii
Abstract
This dissertation addresses mathematical methodologies for seismic reservoir charac-

terization and time-lapse studies. Generally the main goal of reservoir modeling is to
provide 3-dimensional models of the main properties in the reservoir in order to per-
form fluid flow simulations. These properties generally include rock properties, such
as porosity and lithology; fluid properties, such as water and hydrocarbon saturations;
and dynamic properties, such as pressure and permeability. None of these properties
can be directly measured in the subsurface, therefore reservoir properties must be
estimated from other measurements. In petroleum geophysics we generally have two
kinds of measured data: well log data and seismic data. Well log data contain high
resolution information about elastic and petrophysical properties, but they can only
sample few locations of the reservoir. On the other side, seismic data cover the whole
reservoir but the resolution is lower than well log data. Electromagnetic data are
sometimes acquired in addition to seismic data to improve the reservoir description
but the resolution is still limited. In order to obtain suitable models of the reservoir,
we have to combine these two sources of information, wells and seismic, and integrate
physical relations (rock physics and seismic modeling) with mathematical methodolo-
gies (inverse theory and probability and statistics). In particular by using a Bayesian
approach to seismic and rock physics inversion we aim to obtain reservoir models of
rock and fluid properties and the associated uncertainty. Since the resolution and the
quality of seismic data are generally not ideal, uncertainty quantification plays a key
role in reservoir modeling.
This thesis includes three innovative methodologies for seismic reservoir character-
ization: the first method is a Bayesian inversion methodology suitable for reservoirs
iv
in exploration phases with a limited number of wells, the second method is a Bayesian
sampling methodology that can provide multiple reservoir models honoring the given
seismic dataset, the third one is a stochastic inversion methodology that provides
high-detailed models suitable for reservoirs with a large number of wells. The key
innovation in all these methods is the use of new statistical tools to describe the mul-
timodal behavior of rock and properties in the reservoir and the direct integration of
the rock physics model. The main principle of these methodologies is then extended
to time-lapse studies to invert time-lapse seismic data and improve the reservoir de-
scription in terms of changes in rock and dynamic properties. The novelty of this
method is the simultaneous inversion of the pre-production base seismic survey and
repeated monitor surveys. This dissertation contributes to both deterministic and
statistical seismic-based reservoir characterization. Complementary, I investigated
velocity-pressure transforms to determine analytical physical models to describe the
pressure effect on elastic properties and integrate these models in time-lapse reservoir
studies. Finally I also developed a statistical methodology to integrate rock physics
models in formation evaluation analysis and log-facies classification.
All the proposed probabilistic reservoir-characterization techniques can predict
reservoir models with multiple properties (static and dynamic) and the associated
uncertainty. Multiple models can then be derived to run multiple scenarios and
the corresponding risk analysis. All the methodologies were tested using synthetic
data and applied to real case datasets. In the future, these methodologies could
be integrated with history matching techniques to develop statistical methodologies
for seismic history matching and improve reservoir description and monitoring by
simultaneously matching seismic data and production data.
v
Acknowledgements
First, I want to express my thankfulness to my advisers: Gary Mavko for his support,
trust and for the enjoyable conversations; Jack Dvorkin for all the help during my
time here; and Tapan Mukerji for his collaboration which allowed me to expand my
knowledge in many different domains of geophysics and reservoir engineering. I am
really grateful to Gary, Jack and Tapan for the outstanding help and friendship, and
for their helpful comments and challenging questions. It was an honor and privilege
to have them all as advisers.
Many thanks go to the SRB program and all its sponsors for the funding and
support during these three years. Thanks to this generous support, I have been
allowed to choose this research topic and complete my Ph.D. at Stanford. A special
thank goes to all the SRB members: Tiziana for her constant support as a scientist
and as a friend; Amos for the interesting conversations; Adam, Amrita and Anthony
for sharing the office; and all the other students and postdocs that I have met over
my time at Stanford: Piyapa, Danica, Adam Tew, Nishank, Yu, Sabrina, Ammar,
Yuki, Chisato, Humberto, Prianka, Sissel, Fabian, Stephanie and Jane. Thanks to all
the staff members in Geophysics, particularly Fuad, Nancy and Stephanie. It was a
pleasure to work with all of them.
There are a lot of people that I would like to thank for the time we spent together:
Kevin, who was clearly born in the wrong country, for sharing a lot of enjoyable time
in my unofficial office, Randi and Sarah who were wonderful unofficial officemates,
Denys and Adam who always have funny story to tell, Jason, Andreas and Gader
for the afternoon study sessions, Amrita for being such a good listener, Shahar and
Ksenia for the international students solidarity. A great thank goes to the volleyball
vi
group and all the volleyball players and friends I have met: Alec, Guillaume, Sean,
Alice, Jack, Ryan, Alex, Anthony, Shandor, Kaipo, the chinese team and many others.
Finally, thanks to Maurizio for his friendship.
The most important acknowledgement goes to my mom. She is the one without
whom nothing of what I did in my life would have been possible. Thanks to the rest
of my family especially to my nieces with whom I return to be a ten year old kid. A
special thank goes to Ernesto Della Rossa, an amazing mathematician, great scientist
and wonderful person. I want to thank my friend Valentina a special person for me
who makes our friendship unique. I finally want to thank all my friends from Italy,
who I have constantly been in touch with, at least through my ”fairy tales from the
west coast”.
There are other many people that would deserve many acknowledgements, and
there is not enough space here to thank all of them, but the nice thing is that if I
close my eyes I have a memory for each of you guys, and when I reopen my eyes,
most of you are still around me, and this is a great satisfaction. These three years in
California have been an amazing experience.
vii
Contents
Abstract iv
Acknowledgements vi
1 Introduction 1
1.1 Motivation and objectives . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Chapter description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Inversion methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Bayesian approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Probabilistic reservoir properties estimation 10

2.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Probabilistic approach to petrophysical inversion . . . . . . . . . . . . 14
2.3.1 Statistical rock physics modeling . . . . . . . . . . . . . . . . 14
2.3.2 Probabilistic upscaling . . . . . . . . . . . . . . . . . . . . . . 19
2.3.3 From seismic to petrophysics inversion . . . . . . . . . . . . . 20
2.4 Real case application . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1 Methodology implementation . . . . . . . . . . . . . . . . . . 22
2.4.2 Data application . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
viii
3 Sequential Gaussian mixture simulation 50
3.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3 Theory: Linearized Gaussian Mixture Inversion . . . . . . . . . . . . 53
3.4 Theory: Sequential approach . . . . . . . . . . . . . . . . . . . . . . . 56
3.4.1 Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4 Stochastic inversion for facies modeling 67

4.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3.1 Geostatistical methods . . . . . . . . . . . . . . . . . . . . . . 73
4.3.2 Geophysical forward model . . . . . . . . . . . . . . . . . . . . 76
4.3.3 Stochastic optimization algorithm . . . . . . . . . . . . . . . . 79
4.3.4 Secondary information . . . . . . . . . . . . . . . . . . . . . . 84
4.4 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.4.1 Synthetic case . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.4.2 Real case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5 Statistical methods for log evaluation 113

5.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.3.1 Quantitative Log Interpretation . . . . . . . . . . . . . . . . . 117
5.3.2 Rock physics modeling . . . . . . . . . . . . . . . . . . . . . . 126
5.3.3 Log-facies classification . . . . . . . . . . . . . . . . . . . . . . 130
5.4 Application: first example . . . . . . . . . . . . . . . . . . . . . . . . 133
5.5 Application: second example . . . . . . . . . . . . . . . . . . . . . . . 146
ix
5.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
5.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
6 Pressure dependence of elastic properties 159

6.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.3 Physical model and application to lab data . . . . . . . . . . . . . . . 161
6.4 Application to log data and sensitivity . . . . . . . . . . . . . . . . . 168
6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
7 Bayesian time-lapse inversion 173

7.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
7.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
7.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
7.4 Application: synthetic case . . . . . . . . . . . . . . . . . . . . . . . . 178
7.5 Application: real case . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
7.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
8 Final remarks 198
References 200
x
List of Tables
2.1 Rock physics model parameters: density ρ, bulk modulus K and shear
modulus µ of the matrix components. . . . . . . . . . . . . . . . . . . 28
2.2 Correlation coefficients between estimated petrophysical properties and
real data at well A location. . . . . . . . . . . . . . . . . . . . . . . . 37
2.3 Contingency analysis of petrophysical properties estimation at well A
location (f = Absolute frequency; R = Reconstruction rate; r = Recog-
nition rate; E = Estimation index). . . . . . . . . . . . . . . . . . . . 39
4.1 Confusion matrix of the reference case (T stands for true facies, C
stands for classified facies). . . . . . . . . . . . . . . . . . . . . . . . . 94
4.2 One-way sensitivity analysis of the synthetic inversion test: the first
column show the different cases, in the second column we report the
main diagonal of the confusion matrices of the different cases, in the
third column we show the average of the elements of the main diagonal
(sum of the trace of the matrix normalized by the number of facies). . 95
4.3 Mean values of petro-elastic properties in the different facies. Values
of porosity and clay content have been estimated from well log data,
elastic properties values have been computed by rock physics model. . 98
4.4 Confusion matrix of the inversion results obtained by stochastic inver-
sion at well 2 location (T stands for true facies, C stands for classified
facies). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.1 Standard deviations associated to log measurements. . . . . . . . . . 140
xi
List of Figures
2.1 Flowchart of the probabilistic petrophysical properties estimation. . . 13

2.2 Petrophysical curves derived from log interpretation of well A. From
left to right: P-impedance (well log in blue and rock physics model pre-
dictions in red), S-impedance (well log in blue and rock physics model
predictions in red), effective porosity, clay content, water saturation
and litho-fluid classification (oil sand in yellow, water sand in brown,
shale in green) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 Well log data distribution and litho-fluid classification of well A. (top)
P-impedance versus effective porosity color coded by litho-fluid class.
(bottom) S-impedance versus P-impedance color coded by litho-fluid
class (oil sand in yellow, water sand in brown, shale in green). . . . . 27
2.4 Rock physics model calibration on well A: velocity at wet conditions
(obtained performing a fluid substitution on well log velocities) versus
effective porosity, color coded by clay content. The curves are from the
stiff-sand model for a mixture of wet clay and sand, with clay content
equal to (from bottom to top) 0.9, 0.7, 0.5, 0.3 and 0.1. . . . . . . . . 28
2.5 Prior distribution of petrophysical variables. (top) 2D marginal dis-
tribution of effective porosity and clay content with the associated
1D marginal distributions. (bottom) 2D marginal distribution of clay
content and water saturation with the associated 1D marginal distri-
butions. The black crosses represent petrophysical data of well A, the
background color is the joint probability. . . . . . . . . . . . . . . . . 29
xii
2.6 Experimental and fitted variograms of effective porosity for the three
litho-fluid classes. From left to right: variogram of effective porosity
in shale, water sand and oil sand (color lines are the experimental
variograms and dotted lines are the fitted variograms). . . . . . . . . 30
2.7 Petrophysical properties estimation at well A location: effective poros-
ity, clay content and water saturation probability distributions ob-
tained with GMM. (top) Petrophysical properties estimation condi-
tioned by high resolution impedances, extracted from P (R|mf ). (bot-
tom) Petrophysical properties estimation conditioned by upscaled data,
extracted from P (R|mc). The background color is the conditional
probability. Black lines are the actual petrophysical curves, red dotted
lines represent P10, median and P90. . . . . . . . . . . . . . . . . . . 34
2.8 Petrophysical properties estimation conditioned by synthetic seismic
data at well A location: effective porosity, clay content and water
saturation probability distributions extracted from P (R|Sz ) computed
with GMM (top) and KDE (bottom). The background color is the
conditional probability. Black lines are the actual petrophysical curves,
red dotted lines represent P10, median and P90. . . . . . . . . . . . . 36
2.9 Litho-fluid probabilistic classification conditioned by synthetic seismic
at well A location. (top) From left to right: probabilities of litho-
fluid classes P (πz |Sz ) based on petrophysical inversion, MAP of the
probability and actual litho-fluid classes. (bottom) Some realizations
obtained with a Markov chain approach (oil sand in yellow, water sand
in brown, shale in green). . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.10 2D seismic sections passing through well A (on the right) and well B
(on the left): (top) angle stack 20o ; (bottom) angle stack 44o . . . . . . 40
xiii
2.11 Prior model and posterior distributions at well locations. From left to
right: P-impedance of well A, S-impedance of well A, P-impedance of
well B, S-impedance of well B. Blue curves are the actual logs, green
curves represent the upscaled data, black curves are the prior model,
red curves represent the inverted values. Dotted lines represent the
P10 and P90. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.12 2D sections of inverted impedances: (top) inverted P-impedance; (bot-
tom) inverted S-impedance. . . . . . . . . . . . . . . . . . . . . . . . 42
2.13 Probability distributions of effective porosity at wells locations (well
A on the right, well B on the left) and at an intermediate location
between the two wells. Black lines are the actual effective porosity
curves. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.14 Estimation of effective porosity (top), clay content (middle), and water
saturation (bottom) in the 2D section obtained from the mode of the
posterior distributions. . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.15 Litho-fluid probabilistic classification conditioned by real seismic at
well A location. From top left to bottom right: probabilities of litho-
fluid classes P (πz |Sz ) based on petrophysical inversion, six realizations
obtained with a Markov chain approach, and actual litho-fluid classi-
fication (oil sand in yellow, water sand in brown, shale in green). . . . 45
2.16 Estimation of effective porosity along two lines extracted from the 3D
volume. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.17 Isoprobability surface of 70% probability of oil sand litho-fluid class.
The background slices represent two 2D sections of probability of oil
sand occurrence. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
xiv
3.1 Conditional realizations of porosity and reservoir facies obtained by
SGMixSim. The prior distribution of porosity and the hard data values
are shown on top. The second and third rows show three realizations of
porosity and facies (grey is shale, yellow is sand). The fourth row shows
the posterior distribution of facies and the ensemble average of 100
realizations of facies and porosity. The last row shows the comparison
of SGMixSim results with and without post-processing. . . . . . . . . 62
3.2 Linearized sequential inversion with Gaussian Mixture models for the
estimation of porosity map from acoustic impedance values. On top
we show the true porosity map and the acoustic impedance map; on
the bottom we show the inverted porosity and the estimated facies map. 63
3.3 Sequential Gaussian Mixture inversion of seismic data (ensemble of 50
realizations). From left to right: acoustic impedance logs and seismo-
grams (actual model in red, realization 1 in blue, inverted realizations
in grey, dashed line represents low frequency model), inverted facies
profile corresponding to realization 1, maximum a posteriori of 50 in-
verted facies profiles and actual facies classification (sand in yellow,
shale in grey). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4 Application of linearized sequential inversion with Gaussian Mixture
models to a reservoir layer. The conditioning data is P-wave velocity
(top left). Two realizations of porosity and facies are shown: realiza-
tion 1 corresponds to a prior proportion of 30% of sand, realization 2
corresponds to 40% of sand. The histograms of the conditioning data
and the posterior distribution of porosity (realization 2) are shown for
comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1 Workflow of stochastic inversion. . . . . . . . . . . . . . . . . . . . . 72

4.2 Flowchart of PPM algorithm. . . . . . . . . . . . . . . . . . . . . . . 82
xv
4.3 Synthetic well log dataset (well A), from left to right: effective porosity,
volume of clay, water saturation, P-wave and S-wave velocity, density
and facies profile (green represents shale, brown represents silty-sand,
and yellow represents sand). . . . . . . . . . . . . . . . . . . . . . . . 88
4.4 (Top) Rock physics crossplots of well dataset: clay content versus effec-
tive porosity (top left), P-wave velocity versus effective porosity (top
right), S-wave velocity versus P-wave velocity (mid left), and VP /VS
ratio versus P-impedance (mid right), color coded by facies classifica-
tion (green represents shale, brown represents silty-sand, and yellow
represents sand). (Bottom) Joint probability of petrophysical proper-
ties distribution: conditional probability contours color coded by facies
(bottom left), and joint probability surface (bottom right). . . . . . . 89
4.5 Variograms of porosity estimated at the well location, from top to
bottom: variogram of porosity in shale, silty-sand, and sand. . . . . . 90
4.6 Stochastic inversion results at well location, from left to right: ac-
tual facies classification, initial realization, and partial results of the
optimization loop after 3, 10 and 25 iterations classification (green rep-
resents shale, brown represents silty-sand, and yellow represents sand).
The last result (right plot) is the optimized model according to the
fixed tolerance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.7 Synthetic seismograms (red) corresponding to the optimized model of
Figure 4.6 compared to input seismic traces (black). From left to right:
near, mid and far stack, corresponding to the incident angles of 12o ,
24o , and 36o. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.8 Set of 10 different realizations obtained by stochastic inversion. We
show 10 optimized models (obtained from 10 different runs) compared
to the actual classification (green represents shale, brown represents
silty-sand, and yellow represents sand). . . . . . . . . . . . . . . . . . 93
xvi
4.9 Real case application: well log dataset from well 2 (calibration well).
From left to right: P-wave and S-wave velocity, effective porosity,
clay content, water saturation, and actual facies classification (shale
in green, silty-sand in brown, stiff sand in light brown, soft sand in
yellow). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.10 Rock physics model: (left) P-wave velocity versus effective porosity;
(right) S-wave velocity versus effective porosity, color coded by clay
content. Black lines represent constant-cement sand model for different
clay contents (from top to bottom: 0%, 25%, 50%, 75%, and 100%). . 98
4.11 Marginal probability density functions of effective porosity and clay
content conditioned by facies classification. The pdfs of petrophysical
properties are used to distribute rock properties within the reservoir
model at each iteration of stochastic inversion (shale in green, silty-
sand in brown, stiff sand in light brown, soft sand in yellow). . . . . . 99
4.12 Inversion results at well 2 location with synthetic seismic data, from
left to right: actual facies classification, upscaled facies profile, seismic
facies probability, initial model, optimized model after 50 iterations
(shale in green, silty-sand in brown, stiff sand in light brown, soft sand
in yellow). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.13 Multiple realizations with different tolerance conditions: (top) 25 sim-
ulations obtained with a small tolerance; (bottom) 25 simulations ob-
tained with a larger tolerance (shale in green, silty-sand in brown, stiff
sand in light brown, soft sand in yellow). For each set of simulations
we plot the ensemble average (e-type) and we compare the results with
the upscaled facies classification at well 2 location. . . . . . . . . . . . 103
4.14 Inversion results at well 2 location with real seismic data, from left to
right: actual facies classification, upscaled facies profile, seismic facies
probability, initial model, optimized model after 50 iterations (shale in
green, silty-sand in brown, stiff sand in light brown, soft sand in yellow).104
xvii
4.15 (Top) Convergence plot of stochastic inversion results with and without
low frequency information (blue and red symbols respectively) as a
function of iteration number. (Bottom) Boxplots of 25 runs consisting
of 50 iterations with and without low frequency information (left and
right respectively). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.16 Inversion results at well 3 (left plots) and well 5 (right plots) loca-
tions. We compare the actual classification with the optimized model
obtained by stochastic inversion (shale in green, silty-sand in brown,
stiff sand in light brown, soft sand in yellow). . . . . . . . . . . . . . 106
4.17 2D seismic section passing through wells 2 and 3: (top) near angle
stack corresponding to 8o , (bottom) far angle stack corresponding to
26o . The black line represents the top horizon, in time domain, corre-
sponding to the interpreted top of the reservoir. . . . . . . . . . . . . 107
4.18 Inversion results along the 2D section shown in Figure 4.17. On the
left we show the optimized model of reservoir facies (top left), porosity
(mid left) and P-wave velocity (bottom left), obtained by stochastic
inversion. On the right we show the corresponding synthetic seismic
sections, near (top right) and far (top left) and the maximum a pos-
teriori (MAP) of seismic facies probability (converted in depth and
mapped in the geocellular grid) used as secondary information in the
inversion and obtained by multi-step inversion (shale in green, silty-
sand in brown, stiff sand in light brown, soft sand in yellow). . . . . . 108
5.1 Flowchart of the probabilistic petrophysical properties estimation. . . 118

5.2 Schematic representation of petro-elastic uncertainty estimation and
log-facies classification through Monte Carlo simulations of petrophys-
ical and elastic properties. . . . . . . . . . . . . . . . . . . . . . . . . 123
5.3 Well log dataset interval of well A, from left to right: P-wave and
S-wave velocity; density; neutron porosity; gamma ray. . . . . . . . . 134
xviii
5.4 Petrophysical curves performed in Quantitative Log Interpretation,
from left to right: porosity (total porosity in blue, effective porosity in
red); volumetric fraction curves (clay in green, quartz in yellow, and
silt in brown); water saturation; and finally the cumulative volumetric
display (shale in orange, silt in brown, quartz in yellow, water in blue,
and oil in green; red dashed line represents 1 minus porosity and it
separates solid and fluid phase). . . . . . . . . . . . . . . . . . . . . . 135
5.5 Petrophysical outputs comparison (red curves represent our results;
blue curves represent standard commercial software outputs). . . . . . 136
5.6 Calibration of the rock physics model, from left to right P-wave and S-
wave velocity, and VP /VS ratio (black curves represent the actual sonic
log, blue dashed curves represent the predicted rock physics model).
The rock physics model has been calibrated in wet condition and then
applied to the well scenario by Gassmann fluid substitution. . . . . . 137
5.7 Dendogram associated to log-facies classification. A dendrogram con-
sists of many U-shaped lines connecting objects in a hierarchical tree.
The stem of each U represents the distance between the two objects
being connected. Red clusters refer to the connecting histories of the
three recognized facies: low concentration turbidite (LCT), mid con-
centration turbidite (MCT), high concentration turbidite (HCT). . . 138
5.8 Log-facies classification performed at well location: LCT in green,
MCT in brown, and HCT in yellow. Log-facies classification is de-
rived by using petrophysical curves (porosity and clay content) and
velocity data (VP /VS ratio). On the right we show two crossplots in
petrophysical (top right) and petroelastic (bottom right) domain, color
coded by facies classification. . . . . . . . . . . . . . . . . . . . . . . . 139
5.9 Set of 100 realizations of petrophysical curves (gray curves), from left to
right: effective porosity, volume of clay, and volume of quartz (volume
of silt is computed by difference 1 minus the sum of effective porosity,
clay, and quartz). The pointwise median curve (P50) is displayed in red.141
xix
5.10 Set of 100 realizations of elastic curves (gray curves), from left to right:
P-wave and S-wave velocity. The pointwise median curve (P50) is
displayed in red. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.11 Posterior probability density functions of porosity (left side) and P-
wave velocity (right side). For two depth locations z1 and z2 we also
show the histogram of the simulated values. . . . . . . . . . . . . . . 143
5.12 Set of 100 realizations of facies (top left) at well location and estimated
most likely facies profile (top right). On the bottom 5 realizations are
shown: LCT in green, MCT in brown, and HCT in yellow. . . . . . . 144
5.13 Posterior probability distribution of facies estimated by Monte Carlo
simulation (left), most likely facies profile (middle) and associated en-
tropy function (right): LCT in green, MCT in brown, and HCT in yel-
low. On the right we show two crossplots in petrophysical (top right)
and petroelastic (bottom right) domain, color coded by the associated
entropy given by the probabilistic facies classification. . . . . . . . . . 145
5.14 Well log dataset and preliminary petrophysical curves at well location,
from left to right: P-wave and S-wave velocity; density; porosity (total
porosity in blue, effective porosity in red); volumetric fraction curves
(clay in green, quartz in yellow, and muscovite in brown); and water
saturation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.15 Preliminary facies classification: a set of 8 facies has been identified
in this reservoir: 1) marine silty-shale; 2) pro-delta; 3) flood plain; 4)
mouth bar; 5) distributary channel; 6) crevasse splay; 7) tidal deltaic
lobes; 8) tight (left). Grouped histogram of effective porosity (top
right) and clay content (bottom right) as a function of facies classification.148
5.16 Rock physics crossplots: (left) P-wave velocity versus effective porosity;
(right) S-wave velocity versus effective porosity color coded by volume
of clay. Black lines represent stiff sand model for different clay contents
(from top to bottom: 10%, 20%, 30%, 40%, 50% and 60%). . . . . . . 149
xx
5.17 Set of 50 realizations of petrophysical and elastic curves (gray curves),
from left to right: effective porosity, volume of clay, and volume of
quartz (volume of silt is computed by difference 1 minus the sum of
effective porosity, clay, and quartz), P-wave velocity and S-wave veloc-
ity predictions. The pointwise median curve (P50) is displayed in red.
The dashed blue line represents sonic log data. . . . . . . . . . . . . . 150
5.18 Set of 50 realizations of facies (left) at well location, etype (ensemble
average) and estimated most likely facies profile (top right). Color
codes are the same as in Figure 5.15. . . . . . . . . . . . . . . . . . . 151
5.19 Facies probability estimated from ensemble and extracted statistics,
from left to right: probability of facies (first three plots); entropy,
maximum a posteriori of facies distribution and new facies classification
performed by Markov chain integrated approach. Color codes are the
same as in Figure 5.15. . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.20 Comparison of two different classifications. On the left (first two plots)
we show the 8-facies classification: maximum a posteriori of facies and
Markov chain classification (color codes are the same as in Figure 5.15).
On the right we show the simplified 3-facies classification (namely seis-
mic facies classification, last three plots): posterior probability distri-
bution, entropy, and maximum a posteriori of facies (shale in green,
silty-sand in brown, sand in yellow). . . . . . . . . . . . . . . . . . . . 153
6.1 Han’s dataset: dry-rock bulk modulus versus effective pressure color
coded by porosity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.2 Han’s dataset: dry-rock bulk modulus versus effective pressure. The
color represents porosity (left) and a linear combination of porosity
and clay content(right). . . . . . . . . . . . . . . . . . . . . . . . . . . 162
6.3 Eberhart-Phillips relation: on the left each sample has been fitted in-
dependently, on the right all samples have been fitted all together. . . 163
6.4 MacBeth relation (bulk modulus): each sample has been fitted inde-
pendently. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
xxi
6.5 Sensitivity analysis on fitted parameters on MacBeth relation: (top)
fitted parameters versus porosity; (bottom) fitted parameters versus
linear combination of porosity and clay content. . . . . . . . . . . . . 165
6.6 Self similarity concept: (left) measured bulk moduli versus a linear
combination of porosity and clay content color coded by effective pres-
sure; (right) bulk modulus at 50 MPa versus a linear combination of
porosity and clay content and linear fitting. . . . . . . . . . . . . . . 166
6.7 Modified MacBeth relation (bulk modulus). . . . . . . . . . . . . . . 166
6.8 Modified MacBeth relation (shear modulus). . . . . . . . . . . . . . . 167
6.9 Modified MacBeth relation (set of five samples): bulk modulus (left)
and shear modulus (right). . . . . . . . . . . . . . . . . . . . . . . . . 167
6.10 Well log data: from left to right P-wave velocity, S-wave velocity, den-
sity (well log in blue and rock physics model in red), and volumetric
fractions (volume of quartz in yellow, clay in green, silt in black, effec-
tive porosity in red, and water saturation in blue). . . . . . . . . . . . 169
6.11 P-wave velocities computed through the rock physics model in 8 dif-
ferent scenarios (in-situ condition in red, and production scenarios in
black). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
6.12 S-wave velocities computed through the rock physics model in 8 dif-
ferent production scenarios (in-situ condition in red, and production
scenarios in black). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
6.13 Density computed through the rock physics model in 8 different pro-
duction scenarios (in-situ condition in red, and production scenarios in
black). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
6.14 Synthetic seismograms in 8 different production scenarios (short offset). 171
6.15 Synthetic seismograms in 8 different production scenarios (long offset). 172
xxii
7.1 Combined Bayesian inversion results for elastic properties and elas-
tic properties relative changes for a three-layer wedge model: esti-
mated P-impedance relative change, S-impedance relative change and
P-impedance base model (black curves represent the prior model, blue
curves the actual model, solid red curves represent the mean values
and dotted red curves percentiles represent P10 and P90). . . . . . . 179
7.2 Combined Bayesian inversion results for elastic properties and elastic
properties relative changes for a real well log dataset: estimated poste-
rior distributions of P-impedance relative change, S-impedance relative
change and P-impedance base model (black curves represent the actual
model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
7.3 Highly informative prior distribution for pressure and saturation changes.181
7.4 Combined Bayesian inversion results for elastic properties and elastic
properties relative changes for a real well log dataset : estimated poste-
rior distributions of P-impedance relative change, S-impedance relative
change and P-impedance base model (black curves represent the actual
model). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.5 Poorly informative prior distribution for pressure and saturation changes.182
7.6 Combined Bayesian inversion results for elastic properties and elas-
tic properties relative changes for a real well log dataset assuming a
poorly informative prior distribution: estimated posterior distributions
of water saturation change, effective pressure change and porosity base
model (black curves represent the actual model. . . . . . . . . . . . . 182
7.7 Fluid saturations and effective pressure before (top) and after produc-
tion (mid). Bottom left, effective porosity; bottom right, net-to-gross. 183
7.8 Base (left) and repeated (right) seismic survey (perfect data): from
top to bottom near (10o ), mid (20o ), and far (30o ). . . . . . . . . . . 184
7.9 Time-lapse seismic differences: from top to bottom near (10o), mid
(20o ), and far (30o ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
xxiii
7.10 Mean values of inverted elastic properties and elastic properties rel-
ative changes: from top to bottom P-impedance relative change, S-
impedance relative change, and P-impedance estimation from base sur-
vey. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
7.11 Mean values of inverted elastic properties and elastic properties ab-
solute changes: from top to bottom P-impedance absolute change,
S-impedance absolute change, and P-impedance estimation from base
survey. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.12 Mean values of inverted reservoir properties and dynamic properties
relative changes: from top to bottom gas saturation change, oil satu-
ration change, effective pressure change, and porosity estimation. . . . 189
7.13 Examples of point-wise posterior distributions of reservoir properties
and dynamic properties relative changes from top-left to bottom-right
gas saturation change, oil saturation change, effective pressure change,
and porosity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
7.14 Main structure of the geocellular model (blue, top layer in red) and
seismic survey geometry (black). . . . . . . . . . . . . . . . . . . . . . 191
7.15 Active cells of the geocellular model (blue, top layer in red). Seismic
survey geometry is shown in black for comparison. . . . . . . . . . . . 191
7.16 Well log data (well E2): from left to right P-wave velocity, S-wave
velocity, density, clay content, porosity and water saturation. . . . . . 192
7.17 Well log data (well E2): crossplot of P-wave velocity versus porosity
(left) and S-wave velocity versus porosity (right) color coded by clay
content. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
7.18 Seismic data differences in 2003: from left to right, near, mid and far. 193
7.19 Seismic data differences in 2006: from left to right, near, mid and far. 193
7.20 Predicted gas saturation in 2003. . . . . . . . . . . . . . . . . . . . . 194
7.21 Predicted gas saturation in 2006. . . . . . . . . . . . . . . . . . . . . 194
7.22 Predicted fluid pressure in 2003. . . . . . . . . . . . . . . . . . . . . . 195
7.23 Predicted fluid pressure in 2006. . . . . . . . . . . . . . . . . . . . . . 195
xxiv
Chapter 1
Introduction
1.1 Motivation and objectives

The main goal of this thesis is to provide mathematical methodologies to model rock
properties in the subsurface. When we want to create a subsurface model we generally
cannot measure the rock properties we are interested in, at least not exhaustively in
space, but we can measure other properties that are related, directly or indirectly, to
the properties of interest. If we know the physical relations that link the measured
properties (data) with the properties we want to estimate (model ), we can estimate
the model from the data using a mathematical tool known as inverse theory. The set
of physical relations that relates the model to the data is called forward physical model,
and the problem of estimating the model from the data is called inverse problem.
Most of the modeling problems in earth sciences and geological engineering are
inverse problems. For example in reservoir modeling we generally want to estimate
a 3D model of porosity and hydrocarbon saturation, but we have only few wells
where we measure these properties directly, however we can measure seismic data
that are related to the rock properties through seismic and rock physics models.
Similarly, in near surface geophysics we want to estimate the spatial distribution of
water saturation but we measure electromagnetic data or resistivity data; in mining
engineering we are interested in mineral grades but the only source of information
available away from the well locations is seismic. These are all different examples of
1
CHAPTER 1. INTRODUCTION 2
geophysical inverse problems, where we measure some data d, we know the forward
model f that links the model m to the data d and we want to estimate the model m as
a solution of an inverse problem. Most of the methods presented in this thesis can be
applied to different domains, however I will focus on seismic reservoir characterization
problems.
In seismic reservoir characterization we are generally interested in modeling the
porosity of the rocks, the fluid saturations (gas, oil, and water) and lithological in-
formation (for example, sand content and clay content). However these properties
cannot be directly measured in the subsurface away from the well locations. The
main source of information is given by seismic data: traveltime and seismic ampli-
tudes of elastic waves. Seismic data provide information about the elastic contrasts
at the interface between adjacent layers in the subsurface: the amplitudes depend
on the elastic properties (velocities and densities of the upper and lower layer) and
the elastic properties depend on the rock and fluid properties. Estimating reservoir
properties given seismic data is clearly a non-trivial inverse problem. This problem
is generally split in two sub-problems: in the first step, generally called elastic in-
version, we estimate elastic attributes (such as, P- and S-wave velocities, or P- and
S-impedances) from seismic data, whereas in the second step, called petrophysical
property estimation, we estimate rock properties from elastic attributes. Most part of
geophysical research focuses on the first subproblem. Seismic data are very complex
to acquire since tools are generally located on the surface and the target can be few
kilometers deep in the subsurface. Moreover seismic is also difficult to be processed
and interpreted as well, since the data are uncertain and noisy. Elastic attributes
obtained as a solution of the first part of the inverse problem, can be used by geolo-
gists to qualitatively or quantitatively interpret lithologies and fluids presence in the
subsurface. This interpretation is clearly uncertain especially when few well data are
available in the area. Physical relations between elastic properties and rock properties
are one of the main matters of interest in rock physics. Generally, the forward model
is well known, but the solution of the inverse problem is not unique, which calls for
probabilistic approaches. By combining the results of elastic inversion and petro-
physical property estimation we aim to build 3-dimensional models of the reservoir in
terms of rock properties: porosity, lithology, fluid saturation, and possibly effective
pressure. These models can be subsequently used by reservoir engineer as input for
the reservoir fluid flow simulator.
So far in literature, few complete methodologies for solving the corresponding
inverse problem have been proposed in order to obtain reservoir models of rock prop-
erties from seismic data. The inverse problem is quite complex and in the past
reservoir engineers preferred to use statistical relations and simulations to populate
their models rather than estimated models obtained as a solution of the inverse prob-
lem. The most challenging aspects of the inverse problem are: quality of the seismic
data, non-uniqueness of the solution, uncertainty of the measurements, approxima-
tions in the physical models, and changes of scale and domain from seismic data to
the reservoir model. The solution of an inverse problem should not be limited to
a deterministic estimation of the solution which is supposed to be optimal in terms
of best-fitting of the given dataset, but it should also include the evaluation of the
uncertainty of the inversion results. Uncertainty evaluation necessarily requires a
probabilistic approach to inverse problems, in which the inversion results include the
most likely model and the associated uncertainty, or in other words the estimation of
the posterior probability of the inverted parameters. Furthermore, in seismic reservoir
characterization, the number of variables that we want to estimate from seismic data
is generally greater than the number of input variables: for example in an oil clastic
reservoir we generally want to estimate porosity, clay and sand content, water and
hydrocarbon saturation, but we generally have a limited number of angle stacks in the
seismic dataset (generally three partial stacks, representing near, mid, and far offset)
and the solution could be not unique. In addition to that, seismic data are generally
noisy and even in optimistic conditions the signal to noise ratio is quite low. In other
words the inverse problem is underdetermined with noisy data, and the deterministic
approach is generally not suitable. The natural choice for many geophysical inverse
problems is the Bayesian framework, where we combine geological prior knowledge
with the information contained in measured data (Tarantola (2005)). In the following
chapters we are going to present some methodologies to solve this inverse problem in
different situations (exploration stage with few wells or development stage with more
wells).
1.2 Chapter description

In Chapter 1, I introduce the inverse problem and review the main contributions in
literature. In this chapter I also provide the mathematical background of probabil-
ity and statistics of the Bayesian approach. In Chapter 2 we present a probabilistic
approach to reservoir properties estimation based on Bayes rule and Gaussian mix-
ture models. The method aims to estimate at each location of the seismic grid the
posterior probability of reservoir properties conditioned by elastic properties. The
Bayesian inversion accounts for different sources of uncertainty: noise, scale and reso-
lution of seismic data, approximations of the physical models, natural variability and
heterogeneity of the rocks. The final reservoir model is then obtained by taking a
statistical estimator of the posterior probability. In Chapter 3 we introduce a sequen-
tial approach for Bayesian linear inverse problems in the Gaussian mixture case. The
goal of this method is to provide multiple high-detailed models in terms of reservoir
properties. The method is based on the same Bayesian approach presented in Chapter
2, but instead of summarizing the probabilistic information in a single statistical es-
timator at each grid location, the aim of this method is to sample from the pointwise
posterior distributions, or in other words to draw a random sample according to the
estimated posterior distribution. Random sampling in a 3-dimensional space is not
straightforward: in fact, if we sample independently at two adjacent locations (i.e., we
draw a random sample for the first location and then the second random sample for
the second location independently of the value we drew at the first location), we may
obtain a very high value at the first location and a very low value at the second loca-
tion, or vice versa. Such independent random sampling ignores the spatial continuity
expected in the variations of the property, according to depositional and sedimento-
logical laws. In the sequential Bayesian approach, I account for a spatial statistics
model to describe the spatial correlation of the rock properties within the reservoir.
In Chapter 4 we propose a stochastic inversion methodology based on the proba-
bility perturbation method to obtain high detailed reservoir models that match the
input seismic dataset. This method is based on a stochastic optimization technique,

therefore the computational cost is generally higher than the methods presented in
Chapters 2 and 3, however it allows us to include more complicated spatial statis-
tics models and inter-dependence relations between rock properties. In Chapter 5 we
explore a Monte Carlo approach for log evaluation to integrate formation evaluation
analysis and rock physics modeling and to quantify the uncertainty in well log data.
In Chapters 6 and 7 we extend the Bayesian inversion method proposed in Chap-
ter 2 to time-lapse seismic data to estimate pressure and saturation changes during
reservoir production from time-lapse seismic data. In Chapter 6 we introduce a new
rock physics model to relate velocity changes to effective pressure changes whereas in
Chapter 7 we present the Bayesian inversion workflow. The novelty of this method is
the quantitative use of rock physics models in the inversion workflow and the simul-
taneous inversion of the base survey and the seismic differences (between repeated
seismic surveys and base survey). The simultaneous inversion allows us to account for
the correlation between rock properties in the static model (pre-production model)
and changes in dynamic properties (during production).
1.3 Inversion methods

Seismic signals depend on the elastic properties of the media through which they
propagate. Elastic properties of a porous medium depend on intrinsic rock properties
(lithology, porosity, etc.), fluid pressures and fluid saturations. Different inversion
methodologies have been proposed for reservoir properties characterization to infer
rock and fluid properties from seismic signals. Generally, rock properties (porosity
for example) are better-resolved than pressures and saturations. Inverse theory is a
very wide topic of mathematics which includes analytical and numerical approaches
to inverse problems. Many excellent books on geophysical inverse theory have been
published, the most related to exploration geophysics being Tarantola (2005) and
Oliver (2008). However the final goal of this thesis is not to give an exhaustive
overview of the mathematical methods to solve inverse problems, but to describe
new methodologies to infer reservoir properties that include characteristics of the
rocks (lithology, fluid, porosity). For this purpose, Doyen (2007) is a good reference
which contains the description of many mathematical tools used in this thesis. The
forward models used in this thesis are extensively described in Mavko et al. (2009)
and Avseth et al. (2001), whereas the geostatistical methods are presented in Doyen
(2007), Deutsch and Journel (1992), and Goovaerts (1997).
There are different methods to solve inverse problems in reservoir characteriza-
tion including deterministic and probabilistic methods, optimization and sampling
techniques. Following the classification proposed by Bosch et al. (2010), the available
methods can be divided in two main families: 1) hierarchical or multi-step approaches
(e.g., Grana and Della Rossa (2010)), and 2) simultaneous stochastic workflows (e.g.,
González et al. (2008)). In the hierarchical approach, first seismic data are inverted,
deterministically or stochastically, into elastic properties; then rock physics mod-
els transform those elastic properties to the reservoir properties of interest. The
simultaneous workflow aims to estimate simultaneously elastic parameters and the
reservoir properties, guaranteeing consistency between these properties and seismic
data. Stochastic sampling methods and geostatistical algorithms allow represent-
ing the natural variability and heterogeneity and incorporating subseismic scales of
heterogeneities but they generally require higher computational times compared to
sequential approaches.
In hierarchical inversion, seismic elastic inversion is performed on seismic data to
arrive at volumes of P- and S-wave impedances, deterministic or probabilistic (for
example Bayesian elastic inversion); then, a rock physics model is established at a
well to link the elastic properties to porosity, lithology, and fluid saturation; following
these two steps, the rock physics model is used to arrive at the spatial distributions
of rock properties from seismic-derived elastic properties; independently, well data
are used (via, e.g., cluster analysis) to derive a facies classification based on the
reservoir properties rather than elastic properties; finally, this classification is fed
into the probabilistic volume of rock properties obtained in the previous step. If a
probabilistic approach is applied at each step of the inversion process (Grana and
Della Rossa (2010)), the result is a set of volumes of facies probabilities. The natural
approach for all these steps is the Bayesian approach in order to provide full posterior
distributions and quantify the uncertainty. These computations are fast, but the
resolution of the resulting most probable facies volume is the same as of the seismic
data.
Stochastic inversion (also called geostatistical inversion) was introduced in the
90s by Bortoli et al. (1992) and Haas and Dubrule (1994). In stochastic simultaneous
inversion, high-resolution models of subsurface properties (facies and rock properties)
are generated; then relevant rock physics transforms are applied to these volumes to
generate the corresponding volumes of the elastic properties; synthetic seismic traces
are computed on these volumes; finally, the so-obtained synthetic seismic is compared
to real seismic to evaluate the mismatch and an optimization algorithm (deterministic
or stochastic) is applied to determine the most likely reservoir model. This approach
provides a high-resolution model of the subsurface (which still is verified by the low-
resolution seismic data) but is computationally expensive.
Some hybrid methods have been presented by first using the output of the first
technique, namely the spatial probability of rock properties, as a secondary informa-
tion for a high-resolution geostatistical simulation of the facies with spatial variograms
a priori derived at the well(s) or anticipated geological occurrences verified on anal-
ogous reservoirs. The aim of these methods is to integrate the flexibility of Bayesian
inversion of seismic data with advanced geostatistical techniques for detailed reservoir
characterization.
Traditionally reservoir models are used to describe static reservoir properties such
as porosity and lithology and possibly fluid saturations. With the emerge of the
concept of time-lapse (or 4D) earth models, we can integrate static and dynamic
reservoir properties in the reservoir modeling workflow. These models are driven by
the need to integrate time-lapse seismic and reservoir fluid flow simulation predictions
in a unique framework to improve reservoir monitoring and production forecast.
1.4 Bayesian approach

The Bayesian approach is a probabilistic method that allows us to combine a priori
information about a model and data measurements. Suppose that we are interested
in some property of a rock that we cannot measure in the subsurface, for example
porosity, and that for the same rock we can measure another property, for exam-
ple P-wave velocity. From geological, sedimentological or depositional information,
geologists can formulate hypotheses about the porosity of the rock: for example if
the rock is a sandstone, then they can suppose that the porosity is between 10%
and 40%. This information is called prior information and it summarizes the prior
knowledge that we have before looking at the data. In addition to this information, we
can generally measure some properties that are physically related to the property we
want to estimate, and we can generally establish a physical model between these two
properties. In our example if we can measure velocity and we know the rock physics
model for that rock, then we can infer some information about porosity: for instance
if the velocity is high, then the porosity will be low since porosity and velocity are
anti-correlated. This information is called likelihood function, since it links the data
to the model. Both information, the prior information about the rock and the rock
physics likelihood function are uncertain. If we can express these two information
in probability distributions, Bayes’ rule allows integrating these two information in a
single probabilistic information called posterior distribution.
Given an event A, for which we have some prior information, and another event
B that is somehow related to A, then the posterior probability P (A|B) of the event
A given the outcome of the event B is given by:
P (B|A)P (A)
P (A|B) = (1.1)
P (B)
where P (A) is the prior probability of A, P (B) is the probability of B, and P (B|A)
is the conditional probability of B|A.
Intuitively, we can think about Bayes’ rule as a method to reduce the uncertainty
of our prior information when new relevant data are available. Suppose for example
that we know that a rock sample is a sandstone, and from geological information we
know that its porosity is generally between 10% and 40%. Velocity measurements
provide additional data to improve the estimation of porosity. Suppose that in our
example we measured a relative high velocity, which means that the rock is quite
stiff and as a consequence porosity is quite low. These qualitative relations can be
translated into physical-mathematical models which are used to build the likelihood
function (the conditional probability P (B|A)). The product of the two probabilities
allows us to reduce the uncertainty in the prior information. A nice feature of the
Bayesian approach is that when the prior distribution and the likelihood function are
Gaussian, then the posterior distribution is again Gaussian, and the solution of the
inverse problem can be analytically derived (see Chapter 3).
Chapter 2
Probabilistic reservoir properties

estimation
2.1 Abstract
We propose a joint estimation of petrophysical properties which combines statistical
rock physics and Bayesian seismic inversion. Since elastic attributes are correlated
with petrophysical variables (effective porosity, clay content, and water saturation)
and this physical link is associated with uncertainties, then the petrophysical proper-
ties estimation from seismic data can be seen as a Bayesian inversion problem. The
purpose of this work is to present a strategy for estimating the probability distribu-
tions of petrophysical parameters and litho-fluid classes from seismic. The estimation
of reservoir properties and the associated uncertainty is performed in three steps:
linearized seismic inversion to estimate the probabilities of elastic parameters, prob-
abilistic upscaling to include the scale changes effect, and petrophysical inversion
to estimate the probabilities of petrophysical variables and litho-fluid classes. Rock
physics equations provide the link between reservoir properties and velocities, while
linearized seismic modeling connects velocities and density to seismic amplitude. We
adopt a full Bayesian approach to propagate uncertainty from seismic to petrophysics
in an integrated framework which takes into account different sources of uncertainty:
10
CHAPTER 2. PROBABILISTIC RESERVOIR PROPERTIES ESTIMATION 11
the heterogeneity of the real data, the approximation of physical models, measure-
ment errors and scale changes. The methodology has been tested, as a feasibility
step, on real well data and synthetic seismic to show the reliable propagation of the
uncertainty through the three different steps and to compare two different statistical
approaches, parametric and non parametric. The application to a real reservoir study,
including two wells data and partially stacked seismic volumes, has provided as main
result the probability densities of petrophysical properties and litho-fluid classes and
it demonstrates the applicability of the proposed inversion methodology.
2.2 Introduction
In reservoir characterization studies constrained by seismic data, statistical rock
physics is normally used to combine statistical techniques with physical equations
to generate different petroelastic scenarios. The goal of statistical rock physics is to
predict the probability of petrophysical variables when velocities (or impedances) and
density are assigned, and to capture the heterogeneity and complexity of the rocks
and the uncertainty associated with theoretical relations.
The use of statistics in rock physics is becoming more and more frequent: in the
typical statistical rock physics workflow (Avseth et al. (2005) and Doyen (2007)), de-
terministic models are firstly established in order to build physical relations between
elastic properties and reservoir attributes; then, probabilistic petroelastic transfor-
mations are determined, combining these relations with Monte Carlo simulations, to
include the uncertainty associated to the real data (measurement errors and natural
heterogeneity of the rocks) and to the degree of accuracy of the model itself.
The traditional Bayesian framework (Tarantola (2005)) used for uncertainty eval-
uation in elastic inversion (Buland and Omre (2003)) has recently been adopted for
problems of litho-fluid prediction from seismic data (Larsen et al. (2006), Gunning
and Glinsky (2007), and Buland et al. (2008)).
Statistical rock physics has been introduced in Mukerji et al. (2001a) and Eidsvik
et al. (2004a) to estimate reservoir parameters from prestack seismic data and to
evaluate the associated uncertainty. Petrophysical seismic inversion methods based
on statistical relations between elastic and rock properties have been presented in
Bornard et al. (2005) and Coleou et al. (2005). Subsequently, stochastic rock physics
models have been used in Bachrach (2006) for a joint estimation of porosity and sat-
uration and in Sengupta and Bachrach (2007) for pay volume uncertainty evaluation,
while Spikes et al. (2008) developed a probabilistic seismic inversion to constrain reser-
voir properties estimation with well data and seismic. To infer litho-fluid classification
from seismic data, Larsen et al. (2006) proposed an integrated litho-fluid inversion
methodology based on a Markov chain model, while Gallop (2006) presented an ap-
proach based on mixture distributions for facies estimations. Geostatistical methods
in seismic inversion have been introduced in Doyen (1988), Bortoli et al. (1992), and
Haas and Dubrule (1994) and good reviews can be found in Dubrule (2003) and Doyen
(2007). New approaches integrating advanced geostatistical techniques have recently
been proposed in González et al. (2008), and Bosch et al. (2009).
We present a method to integrate statistical rock physics and Bayesian elastic
inversion to compute the probability distributions of the petrophysical properties.
Similar approaches have already been presented, with some assumptions and limita-
tions about the form of the probability distributions, the size of the data and the
type of dependencies considered. By means of more general parametric distributions,
such as Gaussian Mixture Models (GMMs) (Hastie and Tibshirani (1996)), or non
parametric statistical techniques such as Kernel Density Estimation (KDE) (Silver-
man (1986)), some limitations can be overcome by our approach. In addition we take
into account the upscaling problem (Lake and Srinivasan (2004)), in order to face the
limited resolution and the greater uncertainty of seismic data compared to well log
data and we integrate this step within the probabilistic inversion framework.
The workflow we propose (Figure 2.1) can be summarized as follows:
1. Rock physics model calibration: a rock physics model is established using well
log data to predict elastic attributes (velocities or impedances) from petrophysical
properties.
2. Linearized Bayesian seismic inversion: we estimate elastic properties from partially
stacked seismic angle gathers.
3. Conditional probabilities estimation: we calculate the conditional probabilities of
petrophysical variables and litho-fluid classes in a multiproperties, multiscale model.

The methodology includes the upscaling effect within an integrated probabilistic ap-
proach.
Prestack seismic data
Well data &

Rock Physics
Elastic attributes m (Ip, Is)
Petrophysics R (ĭ, C, sw)

Conditional
estimations
Litho fluid class
Figure 2.1: Flowchart of the probabilistic petrophysical properties estimation.
The rock physics model is a set of equations which transforms petrophysical vari-
ables in elastic attributes. The rock physics model type depends on the reservoir
rocks we deal with: the set of equations can be a simple regression on well data or
a more complex physical model (Mavko et al. (2009)). Once the rock physics model
has been calibrated on well logs, we can apply the model to situations not sampled by
log data and generate different scenarios by means of Monte Carlo simulations. This
approach is used to explore, for example, all possible ranges of porosity, saturation
and clay content and to simulate the corresponding acoustic and elastic responses.
The proposed methodology propagates the uncertainty from seismic data to petro-
physics combining three conditional probabilities. The first one is the probability of
elastic properties given seismic data obtained by a Bayesian approach to elastic inver-
sion (Buland and Omre (2003)). The second one is the probability of elastic attributes
at fine scale (high resolution) when the coarse scale values are known. The third one is
the probability of petrophysical properties conditioned by elastic attributes obtained
by integrating the rock physics model equations with Monte Carlo simulations and
generating different geological scenarios.
We applied the methodology to a clastic reservoir in the North Sea, where two
wells and four partial stacked seismic angle gathers are available. As a feasibility step
we tested the methodology on a calibration well using a synthetic seismic trace and we
compared two different approaches: GMMs and KDE. We then applied the method-
ology to the whole seismic volume and we obtained trace-by-trace the probability
density functions of the petrophysical variables and litho-fluid classes.
2.3 Probabilistic approach to petrophysical inver-

sion
The target of this section is to illustrate the probabilistic formulation of the joint
petrophysical inversion. We describe the derivation of the posterior probabilities of
petrophysical properties and litho-fluid classes, conditioned by seismic, using different
attributes (elastic, petrophysical, and categorical) and integrating data coming from
different sources (high resolution well data, coarse resolution data and seismic data).
The methodology is divided in three steps: 1) statistical rock physics modeling,
2) upscaling and 3) petrophysical inversion from seismic data.
In the following we will use m to indicate the acoustic and elastic properties,
typically impedances Ip and Is (m = [Ip Is ]T ), and R to represent the petrophysical
data, typically effective porosity, water saturation and clay content (R = [φ sw C]T ).
2.3.1 Statistical rock physics modeling

One of the important aspects of statistical rock physics is that it combines physical
models with statistics to account for situations not seen in the well data. If all the
variables are considered as random vectors, the rock physics model can be written as:
m = fRP M (R) + ε (2.1)

where fRP M represents the rock physics model and ε is the random error that describes
the degree of accuracy of the model.
For the prior distribution, which is the same at any depth, we assume a multivari-
ate Gaussian Mixture (GM), a linear combination of Gaussian distributions, with a
fixed number of components Nc :
Nc
X
P (R) = αk N(R; µkR , ΣkR ), (2.2)
k=1
where N indicates the multi-Gaussian distribution of the vector R with mean µkR
and covariance matrix ΣkR for all k = 1, ..., Nc and αk are the weights of the linear
combination (with N
P c
k=1 αk = 1).
This choice is motivated by two reasons. The first one is that this formulation
allows us to model each litho-fluid class detectable from a petrophysical point of view
as a single Gaussian component of the mixture. The second one is the analytical con-
venience of this approach since the analytical results valid for Gaussian distributions
can be extended also to Gaussian Mixtures. The number of components initially used
in our tests is three, because the litho-fluid classification we consider consists of shale,
oil sand and water sand.
The approach we propose for the estimation of the conditional probability P (R|m)
is semi-analytical: we generate a set of Ns samples from the prior distribution P (R),
we apply the rock physics model fRP M , we estimate the joint distribution assuming a
Gaussian Mixture distribution and we analytically derive the conditional distribution.
From the prior distribution, different scenarios can be generated by Monte Carlo
simulations: the petrophysical variables can be sampled from the prior distribution
and the elastic response can be computed by the rock physics model fRP M ; in the
pseudo logs generation, we model the vertical correlation of petrophysical properties
by means of a vertical variogram in order to obtain pseudo logs of elastic variables
with a realistic vertical correlation and subsequently perform upscaling on elastic
variables, in the probabilistic upscaling step.
We also made the additional assumption that the error ε in Eq. 2.1 is Gaussian
with zero mean and covariance Σε that can be estimated from well log data. With
this assumption we can state that
P (m|R) = N(m; µ(R), Σε ) (2.3)
where the mean µ(R) = fRP M (R) and the covariance matrix Σε can be assumed
independent of R and related only to ε .
This model allows us to account for the uncertainty associated to the rock physics
model predictions by means of Monte Carlo simulations and conditional probabili-
ties estimations. In fact we can generate a set of Ns samples {Ri }i=1,...,Ns from the
petrophysical prior P (R); then we compute the response of the rock physics model
µ(Ri ) = fRP M (Ri ), for all i = 1, ..., Ns and we generate Ns samples {mi }i=1,...,Ns from
the normal distributions N(m; µ(Ri ), Σ). " #
mi
The joint distribution P (m, R) can be estimated from the Ns samples .
Ri
i=1,...,Ns
If the joint distribution is a Gaussian Mixture
Nc
X
P (m, R) = πk N([m, R]T ; µk[m,R] , Σk[m,R]), (2.4)
k=1
then the conditional distribution P (R|m) is again a Gaussian Mixture. If the rock
physics model fRP M was linear, the joint distribution could be analytically derived
from the prior; but in general fRP M is not linear and the joint distribution P (m, R)
can be obtained from the Monte Carlo samples. The technique adopted to esti-
mate the parameters of the Gaussian components and the weights of the mixture is
Expectation-Maximization (EM) algorithm (Hastie and Tibshirani (1996)). We point
out that these weights can be interpreted as the indicator probability of the discrete
random variable that represents the litho-fluid class.
As a consequence, the conditional distribution P (R|m) is a Gaussian Mixture
Nc
X
P (R|m) = λk N(m; µkR|m , ΣkR|m ) (2.5)
k=1
and we can analytically compute its parameters; in particular the means and the
covariance matrices of the mixture components are given by
−1
µkR|m = µkR + ΣkR,m Σkm,m m − µkm

(2.6)
−1
ΣkR|m = ΣkR,R − ΣkR,m Σkm,m Σkm,R (2.7)
for each given m. The assumption that both petrophysical and elastic variables are
distributed as a Gaussian Mixture is compatible with the hypothesis of GM distribu-
tion for the prior model and it is reasonable if the rock physics model is not too far
from linear.
However, if these assumptions are not in agreement with well log data, a non
parametric approach for the conditional probability estimation P (R|m) should be
adopted. In this case, we propose to estimate the joint distribution P (m, R) by
applying Kernel density estimation on the Monte Carlo samples in a multidimensional
domain. Kernel density estimation is a non parametric technique that allows us to
estimate the probability distribution by fitting a base function, the kernel function,
at each data point including only those observations close to it.
The joint probability can be expressed as the sum of the contributions of the
same kernel function centered at each data point location (see Silverman (1986)); for
example in 2D, if m = [Ip ] and R = [φ]:
Ns
Ip − Ipi φ − φi

1 X
P (m, R) = P (Ip , φ) = K K . (2.8)
Ns hp hφ i=1 hp hφ

where K is the kernel function, Ipi , φi i=1,...,N are the data samples, and hp and hφ
s
are the scaling lengths (also called kernel widths). The kernel function K is a non-
negative symmetric function; in this work we used Epanechnikov kernel as in Doyen
(2007): (
3
4
(1 − x2 ) x ∈ [−1, 1]
K(x) = (2.9)
0 otherwise.
The scaling lengths, for each variable, control how far we incorporate observations
close to data points and they have to be assessed using training data.
In the current workflow, we estimate the joint distribution by KDE on a mul-
tidimensional grid, then the conditional distribution P (R|m) can be numerically
evaluated by definition:
P (m, R)
P (R|m) = R , (2.10)
P (m, R)dR
that corresponds to a normalization of the joint probability P (m, R) for each given
m.
Statistical formulation
We recall here the analytical results for conditional distributions of Gaussian Mix-
tures, extending the results valid in the Gaussian case. If the joint distribution is a
Gaussian Mixture of Nc components, we indicate the joint probability as
Nc
X
P (m, R) = πk N(y; µky , Σky ) (2.11)
k=1
with mean and covariance of each component given by

" # " #
k µkm k Σkm,m Σkm,R
µy = , Σy = ; (2.12)
µkR ΣkR,m ΣkR,R
" #
m
where y = .
R
Then the conditional distribution P (R|m) is again a Gaussian Mixture (see for
example Dovera and Della Rossa (2011))
Nc
X
P (R|m) = λk N(R; µkR|m , ΣkR|m ) (2.13)
k=1
where λk are the weights of the conditional distribution:
πk N(m; µkm , Σkm )

λk (m) = PNc , (2.14)
π N(m; µ l , Σl )
l=1 l m m
and the mean and the covariance of each component of the conditional distribution
can be analytically derived as follows:
−1
µkR|m = µkR + ΣkR,m Σkm,m m − µkm

(2.15)
−1
ΣkR|m = ΣkR,R − ΣkR,m Σkm,m Σkm,R , (2.16)
for each given m.

2.3.2 Probabilistic upscaling

The described methodology does not explicitly consider the difference in scale and
domain of the available data due to different sources of information. In fact the typical
domain of a rock physics model is depth with the resolution of well logs, whereas the
inverted seismic attributes are obtained in time domain with a lower resolution. The
objective of this section is to define a step in the methodology that accounts for
these differences and that is consistent with the general probabilistic approach we
are proposing. In general there are two main issues to take into account in the scale
reconciling problems: the computation of physically equivalent measures at different
scales and the correct propagation of the uncertainty from one scale to another.
The upscaling, both of petrophysical and elastic properties, is complicated by the
presence of spatial and vertical correlation in the heterogeneities distributions (see for
example Lake and Srinivasan (2004)), but in agreement with the choice of our petro-
physical inversion model we concentrate on the problem of coherently transforming
different measures from one resolution to another and of estimating the correspond-
ing changes in probability distribution. The approach we adopted to face the first
problem for elastic properties is Backus averaging (Backus (1962)), while we tackle
the second issue by estimating the conditional distribution of elastic parameters at
high resolution scale (fine scale) given the corresponding data at low resolution scale
(coarse scale).
The starting point for the probabilistic scale change is the rock physics model
with the associated uncertainty defined as in Eq. 2.1. If mf represents the fine scale
(log scale) vector of the elastic parameters and mc the corresponding coarse scale
(seismic) data, we indicate the change of scale with mc = g(mf ), where the function
g represents Backus averaging for velocities and linear average for density. Backus
upscaling is used to determine the effective transversely isotropic elastic constants of
layered media for long wavelength seismic waves.
In order to integrate the upscaling problem in our probabilistic framework we
propose the following scheme. If mf is conditionally distributed with probability
P (mf |R), the problem of the estimation of the distribution conditioned by mc can
n o
be solved by generating a set of Ns samples mfi according to P (mf |R) by
i=1,...,Ns
Monte Carlo simulation, andnapplying o the upscaling transformation g, so that we
f c
obtain a set of joint samples (mi , mi ) that can be used to estimate, with
i=1,...,Ns
Gaussian models for example, the conditional distribution P (mf |mc ).
The conditional distribution of the petrophysical parameters given coarse scale
elastic data P (R|mc ) can be obtained combining P (R|mf ) with P (mf |mc ) by means
of Chapman-Kolmogorov equation (Papoulis (1984)):
Z
c
P (R|m ) = P (R|mf )P (mf |mc )dmf , (2.17)
Rn
where n is the dimension of mf (n = 2 in the case of Ip and Is ). This expression is

the probabilistic model that includes the uncertainties due to rock physics (P (R|mf ))
and to upscaling (P (mf |mc )).
2.3.3 From seismic to petrophysics inversion

In order to obtain the posterior distribution of elastic parameters from seismic, we
use a linearized AVO inversion technique, in a Bayesian framework. The inversion
method adopted here assumes an isotropic and elastic medium and it combines the
convolutional model with Aki-Richards linearized approximation of Zoeppritz equa-
tions valid for vertical weak contrasts, as in Buland and Omre (2003).
If S refers to seismic data, the elastic model can be expressed as:
S = Gℓ + e, (2.18)
where G is the forward linearized operator including both convolution and weak
contrasts Aki-Richards approximation, ℓ is the vector of the logarithms of the whole
trace of the elastic parameters and e is a Gaussian error term with zero mean and
covariance Σe . We also assume that ℓ is distributed according to a multivariate
Gaussian prior ℓ ∼ N(ℓ; µℓ , Σℓ ).
Under these hypothesis it can be shown (Buland and Omre (2003)) that P (ℓ|S)
is again Gaussian:
P (ℓ|S) = N(ℓ; µℓ|S , Σℓ|S ) (2.19)
where
µℓ|S = µℓ + ΣTS,ℓ (ΣS )−1 (S − µS ) (2.20)

Σℓ|S = Σℓ − ΣTS,ℓ (ΣS )−1 ΣS,ℓ (2.21)
with µS = Gµℓ , ΣS = GΣℓ GT +Σe and ΣS,ℓ = GΣℓ is the cross-covariance between
the vector of parameters ℓ and seismic data S.
Once the conditional distribution P (ℓ|S) is known, also the lognormal distribution
at each vertical position z, P (mc |Sz ), can be derived (assuming a depth conversion
of seismic inversion results). The final step to compute the probability of the petro-
physical variables conditioned by seismic data P (R|Sz ) including the upscaling effect
(Eq. 2.17), can be written as:
Z
P (R|Sz ) = P (R|mc)P (mc |Sz )dmc . (2.22)
Rn
By means of Eq. 2.22 we finally obtain the posterior probabilities of the petrophysical
properties. We point out that even though Backus averaging is applied to upscale
elastic properties and estimate elastic properties at coarse scale, anisotropic effects
are not accounted for in the proposed seismic inversion method.
We can also introduce a further step to apply the same methodology in the discrete
domain, in litho-fluid classes classification studies for example. Formally we compute
the probability Z
P (πz |Sz ) = P (πz |R)P (R|Sz )dR (2.23)
Rn
where πz is the generic litho-fluid class at vertical position z, n is the dimension of R
(n = 3, if R = [φ sw C]T ) and P (πz |R) is the rock physics likelihood function.
In order to generate different realizations of litho-fluid classes conditioned by
seismic, including vertical correlation to model the vertical continuity of litho-fluid
classes, we can combine the posterior probability (Eq. 2.23) with a Markov chain
prior model (as in Larsen et al. (2006)):
Z Z
P (πz |Sz ) = P (πz |R)P (R|Sz )dR ∝ P (R|πz )P (πz )P (R|Sz )dR
Rn Rn
Y Z
∝ P (πz |πz−1 ) P (R|πz )P (R|Sz )dR (2.24)
z Rn
where z indicates depth and the probability P (πz |πz−1 ) can be obtained from the
downward Markov chain transition matrix of litho-fluid classes estimated on actual
well log classification (with the assumption that P (πz1 ) = P (πz1 |πz0 ) for notational
convenience).
2.4 Real case application
2.4.1 Methodology implementation

The methodology application is described for an oil saturated clastic reservoir but
it can be adapted to different saturation and lithology reservoir conditions, with the
choice of a suitable rock physics model.
First of all a rock physics model is calibrated at well location using velocity logs
and petrophysical curves obtained in formation evaluation analysis. The rock physics
model can be written in the following formulation
[Vp , Vs , ρ] = fRP M (φ, sw, C) + ε, (2.25)
where Vp and Vs are respectively P and S-waves velocities, ρ is the density, φ is the
effective porosity, sw is the water saturation, C is the clay content and ε is the error
that represents the difference between model predictions and real data. The function
fRP M can be an empirical relation or a theoretical set of equations such as granular
media models or effective media models (see Mavko et al. (2009)).
Secondly we estimate elastic attributes from seismic data: we use a reformula-
tion of the approximation of Zoeppritz equations by Aki-Richards (Aki and Richards
(1980)) in terms of impedances, and we jointly estimate P and S-impedances and
density following the Bayesian approach presented in Buland and Omre (2003). In
terms of impedances the reflection coefficient RP P as a function of the reflection angle
θ becomes:
2
!
1 ∆Ip I¯s ∆I s 1 1 I¯s 2 ∆ρ
RP P (θ) ∼ ¯ − 4 2 sin2 θ ¯ + − + 2 2 sin2 θ (2.26)
2
2 cos θ Ip I¯p Is 2 2 cos θ 2
I¯p ρ̄
where I¯p , I¯s , and ρ̄ are respectively the averages of impedances and density over the
reflecting interface; and ∆Ip , ∆Is , and ∆ρ are the corresponding contrasts. With
realistic noise levels, the inversion cannot retrieve reliable information about density
(as in Buland and Omre (2003)), for this reason in our real case application we do
not use density in the petrophysical inversion workflow.
Finally we calculate the conditional probabilities of petrophysical variables and
litho-fluid classes conditioned by seismic following the methodology described in the
theory section:
a). We assume a prior distribution of the petrophysical variables: in our case
P (φ, sw, C) is assumed as a trivariate GMMs to take into account the observed cor-
relations between variables in each litho-fluid class; in this case the prior is the same
at any vertical position.
b). We generate pseudo logs of petrophysical properties from the prior distribution
with a realistic vertical correlation, in two steps. We firstly create litho-fluid classes
profiles, for example using a first order Markov chain downward model (Larsen et al.
(2006)), and then we generate in each litho-fluid class petrophysical properties verti-
cally correlated using a variogram estimated on well data.
c). We apply the rock physics model fRP M to the petrophysical pseudo logs to obtain
the corresponding elastic attributes and we add a random error ε (Eq. 2.25); then we
compute fine scale impedances.
d). Using the random samples generated in step b) and c), we estimate the joint prob-
ability P (Ipf , Isf , φ, sw, C) and we derive the conditional probability of petrophysical
properties conditioned by impedances P (φ, sw, C|Ipf , Isf ) at fine scale.
e). We upscale the elastic properties applying sequential Backus averaging on a run-
ning window whose length is found by estimating the wavelength from the seismic
bandwidth and the average velocity. We then compute the conditional probabilities
at coarse scale:
Z
P (φ, sw, C|Ipc, Isc ) = P (φ, sw, C|Ipf , Isf )P (Ipf , Isf |Ipc , Isc )dIpf dIsf . (2.27)
R2
f). This last conditional probability is then combined by means of Chapman-Kolmogorov

equation (Papoulis (1984)) with the probability of elastic properties coming from lin-
earized Bayesian inversion P (Ipc , Isc |Sz ), to obtain the posterior probability of petro-
physical properties:
Z
P (φ, sw, C |Sz ) = P (φ, sw, C|Ipc, Isc )P (Ipc , Isc |Sz )dIpc dIsc . (2.28)
R2
at each vertical position z.

g). Finally we estimate the probability of litho-fluid classes conditioned by seismic:
Z
P (πz |Sz ) = P (πz |φ, sw, C)P (φ, sw, C |Sz )dφdswdC (2.29)
R3
and we integrate it with a Markov chain model by means of Eq. 2.24.

In the case of Gaussian Mixture assumption, the joint distribution P (Ipf , Isf , φ, sw, C)
is estimated using Expectation-Maximization algorithm (if the rock physics model
was linear, the joint distribution could be analytically derived from the prior). EM
is an iterative algorithm which allows us to find maximum likelihood estimates of
parameters in probabilistic models in the presence of missing data. EM is a two steps
method: Expectation step computes an expectation of the log-likelihood respect to
the current estimate of the distribution, Maximization step maximizes the expected
log-likelihood found in the previous step. The algorithm converges to the optimal
solution in a number of steps which depends on different factors, such as the distri-
bution shape and data dimensions (see Hastie and Tibshirani (1996)). It is used here
to estimate the parameters (means and covariance matrices) and the weights of the
Gaussian components of the mixture for the joint distribution in the case of multi-
modality of the data. Once the weights and the parameters of the joint distribution
are known, the conditional distribution is analytically derived using Eqs. 2.6 and 2.7.
In alternative we propose to use Kernel density estimation which allows us to
estimate on a multidimensional grid a probability density function. In this case we
apply Kernel density estimation to estimate the joint distribution P (Ipf , Isf , φ, sw, C),
extending Eq. 2.8 in 5D domain. In our implementation we use the same kernel
function, Epanechnikov kernel, for the five different variables and a specific scaling
length for each variable. The critical point of this approach is the calibration of
the scaling lengths: the higher they are, the farther from the data points are the
observations included in the distribution. The choice of the scaling lengths depends
on the number of data points and the spread of the distribution (Doyen (2007)). Once
the joint distribution is estimated, we compute the conditional distribution at fine
scale (Eq. 2.10) by normalizing the joint distribution at each given (Ipf , Isf ).
In both cases, parametric and non parametric, the following steps are similarly
performed: a Gaussian model is assumed for the upscaling step, while a lognormal
distribution is used for seismic inversion. These two probabilities are combined with
fine scale probabilities by means of Chapman-Kolmogorov equation (Eqs. 2.27 and
2.28).
2060
2080
2100
2120
Depth (m)
2140
2160
2180
2200
2220
2240
5000 9000 13000 3000 5000 7000 0 0.2 0.4 0 0.4 0.8 0 0.5 1 Litho−fluid classes
Ip (m/s g/cm3) Is (m/s g/cm3) Effective porosity Clay content Water saturation
Figure 2.2: Petrophysical curves derived from log interpretation of well A. From left
to right: P-impedance (well log in blue and rock physics model predictions in red),
S-impedance (well log in blue and rock physics model predictions in red), effective
porosity, clay content, water saturation and litho-fluid classification (oil sand in yel-
low, water sand in brown, shale in green)
2.4.2 Data application

The methodology has been applied to an oil saturated clastic reservoir in the North
Sea, using angle-stack seismic data and well log data coming from two wells of the
field: well A (relative coordinates: x=1530, y=450) used for model calibration and
well B (x=120, y=1000) used for methodology validation.
The input data of the rock physics model are the petrophysical curves obtained in
formation evaluation analysis: effective porosity, clay content and water saturation.
We focus our attention on a specific reservoir level (Figure 2.2), where we can identify
three litho-fluid classes: Oil Sand, Water Sand and Shale. The litho-facies have been
obtained by means of log-facies classification, on the basis of the petrophysical curves
and the available sedimentological information. The litho-fluid discrimination still
remains visible both in petro-elastic domain and in impedances domain (Figure 2.3).
The adopted rock physics model is the stiff sand model based on Hertz-Mindlin
contact theory. The model was previously calibrated on well A (Figure 2.4) and then
used on well B data; for the calibration, we performed a fluid substitution on velocities
of well A to obtain the corresponding velocities in wet conditions and we determined
the model parameters in order to obtain a good match with well data. The critical
porosity used is 0.4 and the coordination number is 7, while the effective pressure in
the reservoir is 70 MPa. For the solid phase we used a matrix model made by two
components: sand (mostly made of quartz) and wet clay (mostly made of illite), with
the parameters indicated in Table 2.1. The matrix parameters have been selected on
the basis of the available mineralogical information about the rock composition and
for the good match between model predictions and well log data (Figure 2.2). The
choice of considering clay as a mixture of mineral and clay bound water is coherent
with the choice of using effective porosity.
The stiff sand model was selected on the basis of the available geological infor-
mation and because it is appropriate to describe a well consolidated sand. In shale
the effective porosity is near to zero, so that the rock physics model reduces to the
computation of velocities and density of a matrix made of wet clay, by means of
Voigt-Reuss-Hill average and we obtain a good approximation of the velocities in
shale.
12000 shale
water sand
oil sand
11000
P−impedance (m/s g/cm3)

10000
9000
8000
7000
6000
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35

Effective porosity
7000
6500
6000
S−impedance (m/s g/cm3)
5500
5000
4500
4000
3500
3000
6000 7000 8000 9000 10000 11000 12000
P−impedance (m/s g/cm3)
Figure 2.3: Well log data distribution and litho-fluid classification of well A. (top)
P-impedance versus effective porosity color coded by litho-fluid class. (bottom) S-
impedance versus P-impedance color coded by litho-fluid class (oil sand in yellow,
water sand in brown, shale in green).
We describe here the implementation of the inversion methodology and its appli-
cation to the data (we recall that R = [φ sw C]T and m = [Ip Is ]T ). We assume
that the prior distribution P (R) is a Gaussian Mixture (Figure 2.5) which weights
are the actual proportions of litho-fluid classes. In particular we assume a Gaussian
Mixture distribution with truncations for φ and C, while a Gaussian Mixture Score
transformation (extension of the Normal Score transformation) is applied to water
saturation sw. The simulation and the inversion are conducted from the Gaussian
Mixture scores, and at the end of the simulation the results are back-transformed
to recover the actual saturation values. If we assume a large variability within each
ρ (g/cm3 ) K (GP a) µ (GP a)

wet clay 2.5 20 8
sand 2.7 33 36
Table 2.1: Rock physics model parameters: density ρ, bulk modulus K and shear
modulus µ of the matrix components.
litho-fluid class (large covariance matrices) we can also generate samples that are not
present in well data (Figure 2.5); the advantage of this assumption is that it allows
us to simulate the petroelastic properties of different scenarios.
Clay content
5500
C=0.1 0.7
5000
0.6
4500 C=0.3
0.5
Vp (m/s)
C=0.5
4000 0.4
C=0.7
0.3
3500
C=0.9
0.2
3000
0.1
2500
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Effective porosity
Figure 2.4: Rock physics model calibration on well A: velocity at wet conditions (ob-
tained performing a fluid substitution on well log velocities) versus effective porosity,
color coded by clay content. The curves are from the stiff-sand model for a mixture
of wet clay and sand, with clay content equal to (from bottom to top) 0.9, 0.7, 0.5,
0.3 and 0.1.
We then generate pseudo petrophysical curves in two steps. We firstly generate

litho-fluid classes profiles by means of a first order Markov chain downward model
using two transition matrices P1 and P2 honoring well A proportions and transition
−3
x 10
0.8 0.8
2.5
0.7 0.7
0.6 0.6 2
0.5 0.5
Clay content
Clay content
1.5
0.4 0.4
0.3 0.3 1
0.2 0.2
0.5
0.1 0.1
0 0
0.06 0.04 0.02 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 Probability
Probability Effective porosity
0
Probability
0.02
0.04
0.06
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Effective porosity
−3
x 10
1 1 3.5
0.9 0.9
3
0.8 0.8
0.7 0.7 2.5

Water saturation
Water saturation
0.6 0.6 2
0.5 0.5
1.5
0.4 0.4
0.3 0.3 1
0.2 0.2 0.5

0.1 0.1
0.06 0.04 0.02 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Probability
Probability Clay content
0
Probability
0.02
0.04
0.06
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Clay content
Figure 2.5: Prior distribution of petrophysical variables. (top) 2D marginal distribu-

tion of effective porosity and clay content with the associated 1D marginal distribu-
tions. (bottom) 2D marginal distribution of clay content and water saturation with
the associated 1D marginal distributions. The black crosses represent petrophysical
data of well A, the background color is the joint probability.
probabilities respectively above and under the oil-water contact (at 2182 m):
   
0.95 0.03 0.02 0.95 0.03 0.02
   
P1 =  0.20 0.80 0  P2 =  0.06 0.94 0 
   (2.30)
0.01 0.01 0.98 0 0.19 0.81
where rows correspond to shale, water sand and oil sand at generic depth z, while
columns refer to shale, water sand and oil sand at depth z−1 (the downward transition
from water sand to oil sand is impossible in both cases). If the well information is
not representative of reservoir conditions, litho-fluid classes proportions and as a
consequence the transition matrices can be modified. Then, for each litho-fluid class,
we can estimate a variogram to model the vertical correlation of the petrophysical
properties (the results for porosity are presented in Figure 2.6, as an example), and
we generate pseudo logs of petrophysical properties from the prior with a realistic
vertical correlation.
−3 −3 −3
x 10 Shale x 10 Water Sand x 10 Oil Sand
5 5 5
4.5 4.5 4.5
4 4 4
3.5 3.5 3.5

Variogram of porosity
3 3 3
2.5 2.5 2.5
2 2 2
1.5 1.5 1.5
1 1 1
0.5 0.5 0.5
0 0 0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Distance (m) Distance (m) Distance (m)
Figure 2.6: Experimental and fitted variograms of effective porosity for the three
litho-fluid classes. From left to right: variogram of effective porosity in shale, water
sand and oil sand (color lines are the experimental variograms and dotted lines are
the fitted variograms).
We then apply the rock physics model (fRP M ) to obtain the corresponding pseudo
logs of velocities and impedances. The error ε (see Eq. 2.1) added to the elastic vari-
ables, Ip and Is , computed with the rock physics model, is distributed as a bivariate
Gaussian distribution, and its parameters (the elements of covariance matrix) are
estimated from the difference between real data and rock physics model predictions
on well A logs (σp = 530 and σs = 280). Using the pseudo logs generated by means
of the rock physics model we perform the upscaling using sequential Backus averag-
ing in a running window of about 12.5 meters (the estimated wavelength is 125 m,
the operator length is obtained as wavelength/10 as in Avseth et al. (2005)) and we
estimate the conditional probabilities at coarse scale (Eq. 2.27).
Finally the linearized AVO inversion technique is used to jointly estimate the
posterior distribution of P and S-impedance and density. We estimated the four
wavelets independently for each available angle gather, while the trend for the prior
model has been obtained from well logs by filtering impedances logs to a high-cut
value of 4 Hz and by interpolating these logs along the interpreted horizons. The
probabilistic inversion approach is based on the convolutional model and Aki-Richard
linearized approximation of Zoeppritz equations in the limit of vertical weak contrasts.
The elastic parameters derived from seismic inversion are characterized by a log-
Gaussian random field.
The posterior probabilities of petrophysical properties and litho-fluid classes are
obtained by means of Eqs. 2.28 and 2.29 and the results are shown in the next section.
Stiff sand model
The stiff sand model is based on Hertz Mindlin grain contact theory (see Mavko
et al. (2009)). This model provides estimations for the bulk (KHM ) and shear moduli
(µHM ) of a dry rock assuming that the sand frame is a dense random pack of identical
spherical grains subject to an effective pressure P with a given porosity φ0 and an
average number of contacts per grain n (coordination number):
s
n2 (1 − φ0 )2 µ2mat P
KHM = 3 (2.31)
18π 2 (1 − ν)2
s
5 − 4ν 3 3n2 (1 − φ0 )2 µ2mat P
µHM = (2.32)
10 − 5ν 2π 2 (1 − ν)2
where ν is the grain Poisson’s ratio and µmat is the matrix shear modulus.
The matrix elastic moduli are obtained by Voigt-Reuss-Hill averages for a matrix
made of two components, wet clay (mixture of clay and clay bound water) and sand:
!
1 CKc + (1 − C)Ks 1−φ
Kmat = + C 1−C
(2.33)
2 1−φ Kc
+ Ks
!
1 Cµc + (1 − C)µs 1−φ
µmat = + C (2.34)
2 1−φ µc
+ 1−C
µs
where C is the volume of wet clay, φ is the effective porosity, Kc , µc , Ks , and µs are
respectively the bulk and shear moduli of wet clay and sand.
For effective porosity values between zero and the critical porosity φ0 , this model
connects the solid phase elastic moduli Kmat and µmat respectively with the elastic
moduli KHM and µHM of the dry rock at porosity φ0 , by interpolating these two
end members at the intermediate effective porosity values by means of the modified
Hashin-Strikman upper bound:
−1
φ/φ0 1 − φ/φ0 4
Kdry = 4 − 4 − µmat (2.35)
KHM + 3 µmat Kmat + 3 µmat 3
−1
φ/φ0 1 − φ/φ0 1
µdry = 1 − 1 − ξµmat , (2.36)
µHM + 6 ξµmat µmat + 6 ξµmat 6
9Kmat + 8µmat
ξ= .
Kmat + 2µmat
Gassmann’s equations are used for calculating the effect of fluid on velocities using
the matrix and the fluid properties (see Dvorkin et al. (2007) for the use of effective
porosity in Gassmann):
2
Kdry
1− Kmat
Ksat = Kdry + φ 1−φ Kdry
(2.37)
Kf l
+ Kmat
− 2
Kmat
µsat = µdry . (2.38)

From the rock saturated elastic moduli we finally obtain velocities as
s
Ksat + 34 µsat
Vp = (2.39)
ρ
µsat
r
Vs = . (2.40)
ρ
where ρ is the density of the saturated rock, estimated as a weighted linear average.
2.4.3 Results
We firstly applied the methodology using well log data coming from well A and a
synthetic seismic trace to verify the applicability and the validity of the method.
Through this feasibility step we compare two different statistical approaches and we
demonstrate the coherent propagation of the uncertainty through the three steps of
the methodology.
Following the approach presented in the previous sections we will show the results
of the petrophysical properties estimation in three different conditions: at fine scale,
at coarse scale and conditioned by seismic data. In the first step we take into account
only the uncertainty related to the rock physics model at fine scale, without consid-
ering the uncertainty associated to coarse scale and to seismic. In order to estimate
the conditional distribution P (R|mf ), the EM algorithm has been applied assuming
three mixture components (one component for each litho-fluid class) combined with
the analytical expression of Gaussian Mixtures. In Figure 2.7 (top) we display the
marginal conditional probabilities of effective porosity, clay content and water satu-
ration extracted from P (R|mf ) at fine scale. As the rock physics model is accurate,
the uncertainty propagated to petrophysics is quite small and the petrophysical prop-
erties estimation honors the actual curves of effective porosity, clay content and water
saturation derived from log interpretation.
In Figure 2.7 (bottom) we show the results of the probability estimation (Eq. 2.27)
conditioned by the upscaled impedances obtained by applying sequential Backus av-
eraging to rock physics model predictions: the probabilistic upscaling step allows us
to take into account the uncertainty associated to the scale change. The comparison
between Figure 2.7 (top) and Figure 2.7 (bottom) clarifies the impact of coarse res-
olution on uncertainty, which is, as expected, larger in the second case (P (R|mc )),
especially for water saturation.
Finally we combined the results of the statistical rock physics model with seismic
inversion performed with the Bayesian approach, by means of Eq. 2.28, in order
to obtain an estimation of the petrophysical properties conditioned by seismic data
P (R|Sz ). As a feasibility step we applied the methodology using synthetic seismic
data with a signal to noise ratio (SNR) equal to 5. The conditional distributions
Conditional probability distributions (fine scale) Probability

0.4
2060
0.35
2080
2100 0.3
2120
0.25
Depth (m)
2140
0.2
2160
0.15
2180
2200 0.1
2220
0.05
2240
0
0 0.2 0.4 0 0.2 0.4 0.6 0.8 0 0.5 1
Effective porosity Clay content Water saturation
Conditional probability distributions (coarse scale) Probability

0.25
2060
2080
0.2
2100
2120
0.15
Depth (m)
2140
2160
0.1
2180
2200
0.05
2220
2240
0
0 0.2 0.4 0 0.2 0.4 0.6 0.8 0 0.5 1
Figure 2.7: Petrophysical properties estimation at well A location: effective poros-

ity, clay content and water saturation probability distributions obtained with GMM.
(top) Petrophysical properties estimation conditioned by high resolution impedances,
extracted from P (R|mf ). (bottom) Petrophysical properties estimation conditioned
by upscaled data, extracted from P (R|mc). The background color is the conditional
probability. Black lines are the actual petrophysical curves, red dotted lines represent
P10, median and P90.
(Figure 2.8, top) show the multimodality of the petrophysical data and the increase of
uncertainty in particular in thin layers sequences. In the case of multimodal data, the
median is not a good estimator, but we can observe that the probability distributions
still capture the bimodality of petrophysical properties.
We compare now the previous results at seismic scale obtained with Gaussian
Mixture Models with the results obtained with Kernel density approach (Figure 2.8).
The two results are quite similar, in fact in both cases the petrophysical inversion
can capture the bimodality of each variable. The top of the reservoir is characterized
by a thick high porosity oil sand layer and it is well detected in both cases. In
this application Gaussian Mixtures are an appropriate solution and they provide a
good result because the litho-fluid classification of well data (which identifies the
components of the mixture) allows for a good discrimination of petrophysical and
elastic properties; also the approach based on KDE provides a good estimation of the
posterior probability as the Kernel density estimation recognizes the multimodality
of the data if the scaling lengths are correctly chosen. With respect to the GMM
approach, the non parametric approach is more computationally demanding and it
requires the tuning of the scaling lengths parameters.
Even though the linear correlation coefficients cannot be used as a full quantitative
measure of the inversion quality in the case of multimodal distributions, we tried
to evaluate the quality of the match between the inversion results and real data
computing the correlation coefficients between the estimated petrophysical properties
and the actual curves (Table 2.2). The analysis of the correlation coefficients confirms
what we observed from the probability densities, in particular how the scale change
affects the uncertainty.
The methodology has also been applied in a discrete domain, for litho-fluid clas-
sification based on seismic data: from the probability distributions of petrophysical
properties we predicted litho-fluid classes (Eq. 2.29) at well A location and we used
the resulting posterior probabilities to generate multiple realizations of litho-fluid
classes vertical sequences. The rock physics likelihood P (πz |R) has been estimated
using the petrophysical curves and the litho-fluid classification, assuming a Gaussian
distribution for each litho-fluid class.
Probability distributions conditioned by synthetic seismic (GMM) Probability

0.25
2060
2080
0.2
2100
2120
0.15
Depth (m)
2140
2160
0.1
2180
2200
0.05
2220
2240
0
0 0.2 0.4 0 0.2 0.4 0.6 0.8 0 0.5 1
Probability distributions conditioned by synthetic seismic (KDE) Probability

0.25
2060
2080
0.2
2100
2120
0.15
Depth (m)
2140
2160
0.1
2180
2200
0.05
2220
2240
0
0 0.2 0.4 0 0.4 0.8 0 0.5 1
Figure 2.8: Petrophysical properties estimation conditioned by synthetic seismic data

at well A location: effective porosity, clay content and water saturation probability
distributions extracted from P (R|Sz ) computed with GMM (top) and KDE (bot-
tom). The background color is the conditional probability. Black lines are the actual
petrophysical curves, red dotted lines represent P10, median and P90.
Correlation coefficient Correlation coefficient Correlation coefficient

of effective porosity of clay content of water saturation
Conditioned by
fine scale data 0.95 0.89 0.87
Conditioned by
coarse scale data 0.55 0.67 0.63
Conditioned by
synthetic seismic (GMM) 0.58 0.55 0.64
Conditioned by
synthetic seismic (KDE) 0.55 0.60 0.69
Table 2.2: Correlation coefficients between estimated petrophysical properties and

real data at well A location.
The results of the classification conditioned by seismic are shown in Figure 2.9
(top): we can observe a high probability value for the oil sand class reflecting the
thick high porosity sand layer at the top of the reservoir. It is important to observe
that, even though the maximum a posteriori (MAP) of litho-fluid classes probabilities
is not a good estimator in this case, the fluctuations of the probability curves have a
good match with the actual litho-fluid classes profile and they can be used as a prior
probability for multiple realizations. Integrating the probability of litho-fluid classes
conditioned by seismic with the probability obtained from the transition matrices
(Eq. 2.24), we can generate several realizations of litho-fluid classes profiles at well
location (Figure 2.9, bottom).
We used contingency analysis (Table 2.3) in order to evaluate the misclassification
errors, comparing the maximum a posteriori of the probability P (πz |Sz ) with the ac-
tual classification. In the contingency table we computed the absolute frequencies, the
reconstruction rate, the recognition rate and the estimation index. The reconstruc-
tion rate is obtained by normalizing the frequency table per row, while the recognition
rate is obtained by normalizing the frequency table per column. The reconstruction
rate represents the percentage of the samples belonging to a litho-fluid class (ac-
tual) which are classified in that class (predicted). The recognition rate represents
the percentage of the samples classified in a litho-fluid class (predicted) that actu-
ally belong to that class (actual). The information concerning under/overestimation
can be inferred from the estimation index which is defined as the difference between
the reconstruction rate and the recognition rate. A negative estimation index in the
2060
2080
2100
2120
Depth (m)
2140
2160
2180
2200
2220
2240
0 0.5 1 MAP of LFC probabilities Actual LFC

LFC probabilities
2060
2080
2100
2120
Depth (m)
2140
2160
2180
2200
2220
2240
Simulation 1 Simulation 2 Simulation 3 Simulation 4 Simulation 5 Actual LFC
Figure 2.9: Litho-fluid probabilistic classification conditioned by synthetic seismic

at well A location. (top) From left to right: probabilities of litho-fluid classes
P (πz |Sz ) based on petrophysical inversion, MAP of the probability and actual litho-
fluid classes. (bottom) Some realizations obtained with a Markov chain approach (oil
sand in yellow, water sand in brown, shale in green).
Predicted Predicted Predicted

shale water sand oil sand
f = 472 f = 104 f = 108
R = 69% R = 15.2% R = 15.8%
Shale
r = 69.2% r = 40% r = 19.4%
E = −0.2% E = −24.8% E = −3.6%
f = 114 f = 106 f = 101
R = 35.5% R = 33% R = 31.5%
Water sand
r = 16.7% r = 40.8% r = 18.1%
E = 18.8% E = −7.8% E = 13.4%
f = 96 f = 50 f = 349
R = 19.4% R = 10.1% R = 70.5%
Oil sand
r = 14.1% r = 19.2% r = 62.5%
E = 5.3% E = −9.1% E = 8%
Table 2.3: Contingency analysis of petrophysical properties estimation at well A

location (f = Absolute frequency; R = Reconstruction rate; r = Recognition rate; E
= Estimation index).
main diagonal indicates underestimation, while a positive estimation index indicates

overestimation; the off-diagonal terms describe in which class the samples are mis-
classified. In our case, oil sand could be reconstructed with probability 70.5% by
the inversion algorithm, and recognized with probability 62.5%, thus sand is over-
estimated (estimation index 8%). The actual oil sand samples not detected by the
inversion are mostly classified in shale (96 samples, reconstruction rate 19.4%). The
recognition rates of predicted oil sand tell us that some shale samples (108) and water
sand samples (101) are classified in oil sand, which is the reason of the overestimation
of oil sand. Similarly, the negative estimation index for water sand (−7.8%) in the
main diagonal of the contingency table shows an underestimation of water sand. In
some cases we cannot discriminate water sand from actual shale and oil sand from
actual water sand (relatively high estimation index of predicted water sand in actual
shale and of predicted oil sand in actual water sand). This result can be justified
by the rock physics template (Figure 2.3) where we can note the overlaps between
those classes. The misclassification between oil sand and shale is mainly due to the
upscaling effect on thin layers.
Finally we applied the methodology to the whole reservoir level using real seismic
data in order to obtain 3D volumes of petrophysical properties with the associated
uncertainty. First of all we performed a Bayesian inversion on a small 3D volume
including well A used for rock physics model calibration and well B used for method-
ology validation. The seismic volume (4 angle gathers available) contains about 10000
traces in a time window corresponding to a depth interval of approximately 250 m.
In Figure 2.10 we display two seismic sections (related to the partial angle stacks 20o
and 44o ), passing through the two wells. In Figure 2.11 we show the prior model
used for inversion and the inverted values with the associated uncertainties at well
locations, while the corresponding inverted impedances sections, Ip and Is , estimated
by Bayesian inversion, are displayed in Figure 2.12.
Well B Well A
2050
2100
Depth (m)
2150
2200
2250
300 600 900 1200 1500 1800 2100 2400

Distance (m)
2050
2100
Depth (m)
2150
2200
2250
300 600 900 1200 1500 1800 2100 2400

Distance (m)
Figure 2.10: 2D seismic sections passing through well A (on the right) and well B (on
the left): (top) angle stack 20o ; (bottom) angle stack 44o .
The final result of the study is the posterior probability of petrophysical properties
on the whole 3D volume. Figure 2.13 shows the probability distributions of effective
porosity at three different locations along the 2D section passing through the wells.
The comparison between the actual effective porosity curves and the estimated prob-
abilities gives evidence that at the top of the reservoir the estimation is more accurate
than in the lower part. In Figure 2.14 we display the maximum a posteriori of the
well A well A well B well B
2060 Prior model

P10 prior
P90 prior
2080 Well log
Upscaled log
Inverted log
2100 P10 inversion
P90 inversion
2120
2140
Depth (m)
2160
2180
2200
2220
2240
5000 9000 13000 3000 5000 7000 5000 9000 13000 3000 5000 7000
Ip (m/s g/cm3) Is (m/s g/cm3) Ip (m/s g/cm3) Is (m/s g/cm3)
Figure 2.11: Prior model and posterior distributions at well locations. From left
to right: P-impedance of well A, S-impedance of well A, P-impedance of well B,
S-impedance of well B. Blue curves are the actual logs, green curves represent the
upscaled data, black curves are the prior model, red curves represent the inverted
values. Dotted lines represent the P10 and P90.
posterior probabilities of effective porosity, clay content and water saturation. In the
upper part of the section we can clearly detect the overcap clay and the top of the
reservoir characterized by a high porosity sand filled by oil; in the lower part the thin
layers observed in the well logs are not detected and the uncertainty associated to the
inverted properties increases. This is mainly due to the quality of the seismic data,
which is higher at the top (SNR ≃ 3) and very poor at the bottom (SNR ≃ 1), as we
can observe in the seismic sections (Figure 2.10).
We also performed a litho-fluid classification based on seismic data; in Figure 2.15
we show the curves of the conditional probabilities of litho fluid classes at well A
Well B Well A Ip (m/s g/cm3)

2050 12000
11000
2100
10000
Depth (m)
2150
9000
2200 8000
7000
2250
6000
300 600 900 1200 1500 1800 2100 2400
Distance (m)
Is (m/s g/cm3)
2050 7000
6500
2100 6000
5500
Depth (m)
2150
5000
4500
2200
4000
2250 3500
3000
300 600 900 1200 1500 1800 2100 2400
Distance (m)
Figure 2.12: 2D sections of inverted impedances: (top) inverted P-impedance; (bot-

tom) inverted S-impedance.
location conditioned by the corresponding seismic trace and some realizations ob-
tained integrating the posterior probability of litho-fluid classes with the Markov
chain model: the quality of Markov chain realizations is acceptable at the top of the
reservoir and quite poor at the bottom where the SNR of seismic data is very low.
The results on the 3D volume, for example for effective porosity, are shown in
Figure 2.16 by extracting from the estimated effective porosity volume a crossline
passing for well B and an inline for well A. Finally, in Figure 2.17 we propose a 3D
visualization of the hydrocarbon sands probability: the oil sand probability cube has
been thresholded to reveal the areas where the probability of oil sand litho-fluid class
occurrence is greater than 0.7.
Figure 2.13: Probability distributions of effective porosity at wells locations (well A

on the right, well B on the left) and at an intermediate location between the two
wells. Black lines are the actual effective porosity curves.
Well B Well A Effective porosity

2050
0.2
2100
0.15
Depth (m)
2150
0.1
2200
2250 0.05
300 600 900 1200 1500 1800 2100 2400

Distance (m)
Clay content
2050 0.7
2100 0.6
Depth (m)
0.5
2150
0.4
2200
0.3
2250 0.2
0.1
300 600 900 1200 1500 1800 2100 2400
Distance (m)
Water saturation
2050
0.9
2100 0.8
Depth (m)
0.7
2150
0.6
2200 0.5
0.4
2250
0.3
300 600 900 1200 1500 1800 2100 2400

Distance (m)
Figure 2.14: Estimation of effective porosity (top), clay content (middle), and wa-
ter saturation (bottom) in the 2D section obtained from the mode of the posterior
distributions.
2060
2080
2100
2120
Depth (m)
2140
2160
2180
2200
2220
2240
0 0.5 1 Simulation 1 Simulation 2 Simulation 3
LFC probabilities
2060
2080
2100
2120
Depth (m)
2140
2160
2180
2200
2220
2240
Simulation 4 Simulation 5 Simulation 6 Actual LFC
Figure 2.15: Litho-fluid probabilistic classification conditioned by real seismic at well

A location. From top left to bottom right: probabilities of litho-fluid classes P (πz |Sz )
based on petrophysical inversion, six realizations obtained with a Markov chain ap-
proach, and actual litho-fluid classification (oil sand in yellow, water sand in brown,
shale in green).
Figure 2.16: Estimation of effective porosity along two lines extracted from the 3D
volume.
Figure 2.17: Isoprobability surface of 70% probability of oil sand litho-fluid class. The
background slices represent two 2D sections of probability of oil sand occurrence.
2.5 Discussion
The feasibility test based on synthetic seismic shows the propagation of the differ-
ent sources of uncertainty through the different steps and the applicability of the
methodology with both the proposed statistical approaches. The real case applica-
tion, integrating well data and real seismic, pointed out that the results are quite
satisfactory as long as the quality of seismic is acceptable. In particular the use of
Gaussian Mixture seems to be a valid approach for the classification of petrophysical
and categorical parameters, which can be applied to real cases with reduced compu-
tational time.
The application of the rock physics model is not computationally demanding
whereas the estimation of the conditional probability in a Bayesian framework can be
quite hard to obtain because it requires the estimation of the joint distribution in a
space of high dimensions. Gaussian Mixture Models are a suitable solution because
of their analytical convenience, especially when the distributions of petrophysical and
elastic attributes describe different litho-fluid classes features. The non parametric
alternative, Kernel density estimation, is more computationally demanding because
it requires the numerical evaluation of a joint probability on a multidimensional do-
main. A more efficient method to tackle the multidimensional extension of KDE is
based on Fast Fourier Transform (FFT). In fact KDE can also be seen as a convolu-
tion, so that we can reduce the computational time by realizing the convolution by
means of FFT (Buland et al. (2008)). However, one of the most critical point for
the estimation of probabilities by means of Kernel density is the choice of the scaling
lengths parameters, which have to be determined through different trials.
The main simplification we adopted in our approach is the overlooking of the
spatial correlation of petrophysical variables for the estimation of the conditional dis-
tributions, in order to reduce the dimension of the probability space. However we
take into account the vertical correlation in seismic inversion by including a vertical
correlation in the prior covariance matrix of the vector of elastic parameters and in the
prior covariance matrix of the error on seismic amplitudes for each angle gather (Bu-
land and Omre (2003)); whereas the spatial correlation is not explicitly accounted for,
because we adopt a trace-by-trace inversion approach to jointly estimate impedances

and densities from seismic. We remark that the lateral continuity of our results is
mainly related to the imaging, in fact part of the lateral correlation of seismic data is
imposed by the migration operator which is a spatial filter whose correlation length
is associated with the Fresnel zone.
In order to perform the final step from continuous petrophysical variables to litho-
fluid classes modeling, Gaussian Mixture Models seem to be an appropriate approach
as they can express the multimodal features of the petrophysical variables in the dif-
ferent litho-fluid classes. The integration of more advanced geostatistical techniques
could be a significant improvement in order to use the probabilistic information related
to litho-fluid classes to generate multiple realizations for reservoir characterization.
2.6 Conclusions
The presented methodology aims to propagate the uncertainty from seismic to petro-
physical properties, including the effect of scale change, the seismic noise error and the
degree of approximation of the physical models. Statistical rock physics, combined
with the probabilitic approach adopted for seismic inversion, is proposed in order to
quantify the uncertainty. The main results of the methodology are the probability
distributions of the estimated petrophysical parameters, that can be used to assess
the reliability of reservoir properties estimation. In order to obtain the posterior dis-
tribution of the petrophysical properties, we point out that one of the key point of
our methodology is the use of Gaussian Mixture Models and the identification of the
weights of the mixture as the indicator probability of the litho-fluid classes.
Even though the considered uncertainty factors do not cover all the possible
sources, the 1D feasibility test shows that the main effects due to scale changes
and seismic noise are taken into account and that these two factors can explain an
important part of the uncertainty.
In the application case, the method works better in the upper layers, where the
signal to noise ratio is high, rather than in the lower layers, where the signal to noise
is low. In conclusion, where the signal to noise is acceptable, the probabilistic petro-
physical evaluation on the real case shows the applicability of the methodology and
that the reliability of the seismic data is coherently propagated to the petrophysical
properties prediction.
The proposed methodology can be applied to all reservoirs where elastic charac-
terization of petrophysical properties is possible and where the physical link can be
described by a suitable rock physics model.
Chapter 3
Sequential Gaussian mixture

simulation
3.1 Abstract
We present here a method for generating realizations of the posterior probability den-
sity function of a Gaussian Mixture linear inverse problem in the combined discrete-
continuous case. This task is achieved by extending the sequential simulations method
to the mixed discrete-continuous problem. The sequential approach allows us to gen-
erate a Gaussian Mixture random field that honors the covariance functions of the
continuous property and the available observed data. The traditional inverse theory
results, well known for the Gaussian case, are first summarized for Gaussian Mix-
ture models: in particular the analytical expression for means, covariance matrices,
and weights of the conditional probability density function are derived. However,
the computation of the weights of the conditional distribution requires the evaluation
of the probability density function values of a multivariate Gaussian distribution, at
each conditioning point. As an alternative solution of the Bayesian inverse Gaus-
sian Mixture problem, we then introduce the sequential approach to inverse problems
and extend it to the Gaussian Mixture case. The main novelty compared to the ap-
proach presented in Chapter 2 is that in this method we sample from the posterior
50
CHAPTER 3. SEQUENTIAL GAUSSIAN MIXTURE SIMULATION 51
pdfs rather than just taking a statistical estimator, which allows us to obtain mul-
tiple realizations all honoring the initial conditioning measurements. The Sequential
Gaussian Mixture Simulation (SGMixSim) approach is presented as a particular case
of the linear inverse Gaussian Mixture problem, where the linear operator is the iden-
tity. Similar to the Gaussian case, in Sequential Gaussian Mixture Simulation the
means and the covariance matrices of the conditional distribution at a given point
correspond to the kriging estimate, component by component, of the mixture. Fur-
thermore, Sequential Gaussian Mixture Simulation can be conditioned by secondary
information to account for non-stationarity. Examples of applications with synthetic
and real data, are presented in the reservoir modeling domain where realizations of
facies distribution and reservoir properties, such as porosity or net-to-gross, are ob-
tained using Sequential Gaussian Mixture Simulation approach. In these examples,
reservoir properties are assumed to be distributed as a Gaussian Mixture model. In
particular, reservoir properties are Gaussian within each facies, and the weights of
the mixture are identified with the point-wise probability of the facies.
3.2 Introduction
Inverse problems are common in many different domains such as physics, engineering,
and earth sciences. In general, solving an inverse problem consists of estimating the
model parameters given a set of observed data. The operator that links the model
and the data can be linear or non-linear.
In the linear case, estimation techniques generally provide smoothed solutions.
Kriging, for example, provides the best estimate of the model in the least-squares
sense. Simple kriging is in fact identical to a linear Gaussian inverse problem where
the linear operator is the identity, with the estimation of posterior mean and covari-
ance matrices with direct observations of the model space. Monte Carlo methods can
be applied as well to solve inverse problems Mosegaard and Tarantola (1995) in a
Bayesian framework to sample from the posterior; but standard sampling method-
ologies can be inefficient in practical applications. Sequential simulations have been
introduced in geostatistics to generate high resolution models and provide a number
of realizations of the posterior probability function honoring both prior information

and the observed values. Deutsch and Journel (1992) and Goovaerts (1997) give de-
tailed descriptions of kriging and sequential simulation methods. Hansen et al. (2006)
proposes a methodology that applies sequential simulations to linear Gaussian inverse
problems to incorporate the prior information on the model and honor the observed
data.
We propose here to extend the approach of Hansen et al. (2006) to the Gaus-
sian Mixture case. Gaussian Mixture models are convex combinations of Gaussian
components that can be used to describe the multi-modal behavior of the model and
the data. Sung (2004), for instance, introduces Gaussian Mixture distributions in
multivariate nonlinear regression modeling; while Hastie and Tibshirani (1996) pro-
poses a mixture discriminant analysis as an extension of linear discriminant analysis
by using Gaussian Mixtures and Expectation-Maximization algorithm Hastie et al.
(2002). Gaussian Mixture models are common in statistics (see, for example, Has-
selblad (1966) and Dempster et al. (1977)) and they have been used in different
domains: digital signal processing Reynolds (2000) and Gilardi et al. (2002), engi-
neering Alspach and Sorenson (1972), geophysics Grana and Della Rossa (2010), and
reservoir history matching Dovera and Della Rossa (2011).
In this chapter we first present the extension of the traditional results valid in
the Gaussian case to the Gaussian Mixture case; we then propose the sequential
approach to linear inverse problems under the assumption of Gaussian Mixture dis-
tribution; and we finally show some examples of applications in reservoir modeling.
If the linear operator is the identity, then the methodology provides an extension
of the traditional Sequential Gaussian Simulation (SGSim, see Deutsch and Journel
(1992), and Goovaerts (1997)) to a new methodology that we call Sequential Gaus-
sian Mixture Simulation (SGMixSim). The applications we propose refer to mixed
discrete-continuous problems of reservoir modeling and they provide, as main result,
sets of models of reservoir facies and porosity. The key point of the application is that
we identify the weights of the Gaussian Mixture describing the continuous random
variable (porosity) with the probability of the reservoir facies (discrete variable).
3.3 Theory: Linearized Gaussian Mixture Inver-

sion
In this section we provide the main propositions of linear inverse problems with Gaus-
sian Mixtures (GMs). We first recap the well-known analytical result for posterior
distributions of linear inverse problems with Gaussian prior; then we extend the result
to the Gaussian Mixtures case.
In the Gaussian case, the solution of the linear inverse problem is well-known
Tarantola (2005). If m is a random vector Gaussian distributed, m ∼ N(µm , Cm ),
with mean µm and covariance Cm ; and G is a linear operator that transforms the
model m into the observable data d
d = Gm + ε, (3.1)
where ε is a random vector that represents an error with Gaussian distribution

N(0, Cε ) independent of the model m; then the posterior conditional distribution of
m|d is Gaussian with mean and covariance given by
µm|d = µm + Cm GT (GCm GT + Cε )−1 (d − Gµm ) (3.2)

Cm|d = Cm − Cm GT (GCm GT + Cε )−1 GCm . (3.3)
This result is based on two well known properties of the Gaussian distributions:
(A) the linear transform of a Gaussian distribution is again Gaussian; (B ) if the
joint distribution (m, d) is Gaussian, then the conditional distribution m|d is again
Gaussian.
These two properties can be extended to the Gaussian Mixtures case. We as-
sume that x is a random vector distributed according to a Gaussian Mixture with
P Nc k k
Nc components, f (x) = k=1 πk N(x; µx , Cx ), where πk are the weights and the
distributions N(x; µkx , Cxk ) represent the Gaussian components with means µkx and
covariances Cxk evaluated in x. By applying property (A) to the Gaussian compo-
nents of the mixture, we can conclude that, if L is a linear operator, then y = Lx
is distributed according to a Gaussian Mixture. Moreover, the pdf of y is given by

f (y) = N k k T
P c
k=1 πk N(y; Lµx , LCx L ).
As a matter of fact, by using the definition of characteristic function, we can write
T T Lx T t)T x
Φy (t) = E(et y ) = E(et ) = E(e(L ) = Φx (LT t) (3.4)
where t ∈ ℜN , y ∈ ℜN , x ∈ ℜM and L : ℜM → ℜN . As x is a Gaussian Mixture,

then the characteristic function Φx (s) is a linear combination of the characteristic
functions of the Gaussian components; then
Nc
X
Φy (t) = πk Φxk (LT t) (3.5)
k=1
where Φxk is the characteristic function of the Gaussian component k. Property (A)
applied to each Gaussian component allows to conclude that also y is distributed
according to a Gaussian Mixture and the pdf is given by
Nc
X
f (y) = πk N(y, Lµkx , LCkx LT ) . (3.6)
k=1
Similarly we can extend property (B ) to conditional Gaussian Mixture distribu-

tions. The well-known result of the conditional multivariate Gaussian distribution
has already been extended to multivariate Gaussian Mixture models (see, for exam-
ple, Alspach and Sorenson (1972)). In particular, if (x1 , x2 ) is a random vector whose
joint distribution is a Gaussian Mixture
Nc
X
f (x1 , x2 ) = πk fk (x1 , x2 ), (3.7)
k=1
where fk are the Gaussian densities, then the conditional distribution of x2 |x1 is
again a Gaussian Mixture
Nc
X
f (x2 |x1 ) = λk fk (x2 |x1 ), (3.8)
k=1
and its parameters (weights, means, and covariance matrices) can be analytically
derived. The coefficients λk are given by
πk fk (x1 )
λk = PNc , (3.9)
ℓ=1 π ℓ fℓ (x1 )
where fk (x1 ) = N(x1 ; µkx1 , Ckx1 ); and the means and the covariance matrices are
−1 k
µkx2 |x1 = µkx2 + Ckx2 ,x1 Ckx1 x1 − µkx1

(3.10)
−1 T
Ckx2 |x1 = Ckx2 − Ckx2 ,x1 Ckx1 Ckx2 ,x1 ,

(3.11)
where Ckx2 ,x1 is the cross-covariance matrix.

In fact, the general conditional distribution can be written by definition as: f (x2 |x1 ) =
f (x1 ,x2 )
f (x1 )
. We then explicitly write the numerator and the denominator by evaluating
the single Gaussian components of the mixture. If x = (x1 , x2 ) is distributed accord-
ing to a Gaussian Mixture distribution, then the single components of the joint pdf
are Gaussian with means and covariances given by
" # " #
k µkx1 k Ckx1 ,x1 Ckx1 ,x2
µx = , Cx = (3.12)
µkx2 Ckx2 ,x1 Ckx2 ,x2
where x1 ∈ ℜN1 and x2 ∈ ℜN2 . Now the property (B) of Gaussian distributions
applied to each single component of the mixture allows us to write the conditional
distributions for each component as Gaussian fk (x2 |x1 ) with means and covariances
given by Eqs. 3.3 and 3.3. If the joint density of each component is written as the
product fk (x1 , x2 ) = fk (x2 |x1 )fk (x1 ), then the Gaussian Mixture joint pdf of (x1 , x2 )
is
Nc
X
f (x1 , x2 ) = πk fk (x2 |x1 )fk (x1 ) . (3.13)
k=1
Similarly, the denominator of the full conditional pdf can be written by using the
definition of marginal density as
Z Z Nc
!
X
f (x1 ) = f (x1 , x2 )dx2 = πk fk (x1 , x2 ) dx2 =
ℜN2 ℜN2 k=1
Nc
X Z Nc
X
= πk fk (x1 , x2 )dx2 = πk fk (x1 ) (3.14)
k=1 ℜN2 k=1
where fk (x1 ) is the marginal density of the k th Gaussian component. If we explicitly

write the numerator and the denominator of the conditional distribution f (x2 |x1 ),
we obtain P Nc
k=1 πk fk (x2 |x1 )fk (x1 )
f (x2 |x1 ) = PNc (3.15)
ℓ=1 πℓ fℓ (x1 )
If we set the new coefficients λk as in Eq. 3.3, then the full form of conditional pdf
can be written as in Eq. 3.3 and the so obtained pdf is again a Gaussian Mixture.
By combining these propositions, the main result of linear inverse problems with
Gaussian Mixture can be derived.
Theorem 3.3.1. Let m be a random vector distributed according to a Gaussian Mix-

ture m ∼ N k k k
P c
k=1 πk N(µm , Cm ), with Nc components and with means µm , covariances
Ckm , and weights πk , for k = 1, . . . , Nc . Let G : ℜM → ℜN be a linear operator, and
ε a Gaussian random vector independent of m with 0 mean and covariance Cε , such
that d = Gm + ε, with d ∈ ℜN , m ∈ ℜM , ε ∈ ℜN , then the posterior conditional
distribution m|d is a Gaussian Mixture.
Moreover, the posterior means and covariances of the components are given by
µkm|d = µkm + Ckm GT (GCkm GT + Cε )−1 (d − Gµkm ) (3.16)

Ckm|d = Ckm − Ckm GT (GCkm GT + Cε )−1 GCkm , (3.17)
where µkm and Ckm , are respectively the prior mean and covariance of the k th Gaussian
component of m. The posterior coefficients λk of the mixture are given by
πk fk (d)
λ k = P Nc , (3.18)
ℓ=1 πℓ fℓ (d)
where the Gaussian densities fk (d) have means µkd = Gµkm and covariances Ckd =
GCkm GT + Cε .
3.4 Theory: Sequential approach

Based on the results presented in the previous section, we introduce here the sequen-
tial approach to linearized inversion in the Gaussian Mixture case. We first recap the
main result for the Gaussian case Hansen et al. (2006).
The solution of the linear inverse problem with the sequential approach requires
some additional notation. Let mi represent the ith element of the random vector m,
and let ms represent a known subvector of m. This notation will generally be used
to describe the neighborhood of mi in the context of sequential simulations. Finally

we assume that the measured data d are known having been obtained as a linear
transformation of m according to some linear operator G.
Theorem 3.4.1. Let m be a random vector, Gaussian distributed, m ∼ N(µm , Cm )

with mean µm and covariance Cm . Let G be a linear operator between the model
m and the random data vector d such that d = Gm + ε, with ε a random error
vector independent of m with 0 mean and covariance Cε . Let ms be the subvector
with direct observations of the model m, and mi the ith element of m. Then the
conditional distribution of mi |(ms , d) is again Gaussian.
Moreover, if the subvector ms is extracted from the full random vector m with the
linear operator A such that ms = Am, where the ith element is mi = Ai m, with Ai
again linear, then the mean and variance of the posterior conditional distribution are:
" #
h i −1 ms − Aµm
µmi |(ms ,d) = µmi + Ai Cm AT Ai Cm GT C(ms ,d) (3.19)
d − Gµm
" #
T
2 2
h i −1 AC m A i
σm i |(ms ,d)
= σm − Ai Cm AT Ai Cm GT C(ms ,d) , (3.20)
GCm ATi
i
2
where µmi = Ai µm , σm i
= Ai Cm ATi , and
" #
ACm AT ACm GT
C(ms ,d) = . (3.21)
GCm AT GCm GT + Cε
To clarify the statement we give the explicit form of the operators Ai and A. In
particular, Ai is written as
h i
Ai = 0 0 ... 1 ... 0 , (3.22)
with the one in the ith column. If the subvector ms has size n, ms = {mi1 , mi2 , . . . , min },
and m has size M; then the operator A is given by
 
0 0 ... 1 ... 0
 
 0 ... 1 0 ... 0 
A= . . , (3.23)
 
 .. .. .. .. .. ..
 . . . . 

0 1 ... 0 0 0
where A has dimensions n×M and the ones are in the i1 , i2 , . . . , in columns. Theorem
3.4.1 can be proved using the properties (A) and (B ) described in Section 3.3 (see
Hansen et al. (2006)). Then, by using Theorem 3.3, we extend the result to the
Gaussian Mixture case.
Theorem 3.4.2. Let m be a random vector distributed according to a Gaussian Mix-

ture, m ∼ N k k k
P c
k=1 πk N(µm , Cm ), with Nc components and with means µm , covariances
Ckm , and weights πk , for k = 1, . . . , Nc . Let G a linear operator such that d = Gm+ε,
with ε a random error vector independent of m with 0 mean and covariance Cε . Let
ms be the subvector with direct observations of the model m, and mi the ith element
of m. Then the conditional distribution of mi |(ms , d) is again a Gaussian Mixture.
Moreover, the means and variances of the components of the posterior conditional
distribution are:
" #
h i −1 ms − Aµkm
µkmi |(ms ,d) = µkmi + Ai Ckm AT Ai Ckm GT Ck(ms ,d) (3.24)
d − Gµkm
" #
2 (k) 2 (k)
h i −1 ACkm ATi
σmi |(ms ,d) = σm − Ai Ckm AT Ai Ckm GT Ck(ms ,d) ,(3.25)
GCkm ATi
i
2 (k)
where µkmi = Ai µkm , σmi = Ai Ckm ATi , and
" #
ACkm AT ACkm GT
Ck(ms ,d) = . (3.26)
GCkm AT GCkm GT + Cε
The posterior coefficients of the mixture are given by
πk fk (ms , d)
λk = PNc , (3.27)
ℓ=1 πℓ fℓ (ms , d)
where the Gaussian components fk (ms , d) have means

" #
k
Aµ m
µk(ms ,d) = (3.28)
Gµkm
and covariances Ck(ms ,d) .

In the case where the linear operator is the identity, the associated inverse problem
reduces to the estimation of a Gaussian Mixture model with direct observations of the
model space at given locations. In other words, if the linear operator is the identity,
the theorem provides an extension of the traditional Sequential Gaussian Simulation
(SGSim) to the Gaussian Mixture case. We call this methodology Sequential Gaussian
Mixture Simulation (SGMixSim), and we show some applications in the next section.
3.4.1 Proof
First of all we observe that the joint vector (m, ε) is distributed according to a Gaus-
sian Mixture. If we consider the following linear operator
 
Ak 0
 
B=  A 0 
 (3.29)
G 1N
then, we observe that  

mk " #
 
 md  = B · m
(3.30)
  ε
s
and the joint vector (mk , md , s) is distributed according to a Gaussian Mixture. The
statement follows by assuming x1 = (md , s) and x2 = mk , and deriving the condi-
tional distribution x2 |x1 .
3.5 Application
We describe here some examples of applications with synthetic and real data, in the
context of reservoir modeling. First, we present the results of the estimation of a
Gaussian Mixture model with direct observations of the model space as a special case
of Theorem 3.4 (SGMixSim). In our example, the continuous property is the porosity
of a reservoir, and the discrete variable represents the corresponding reservoir facies,
namely shale and sand. This means that we identify the weights of the mixture
components with the facies probabilities. The input parameters are then the prior
distribution of porosity and a variogram model for each component of the mixture.
The prior is a Gaussian Mixture model with two components and its parameters are
the weights, the means, and the covariance matrices of the Gaussian components. We
assume facies prior probabilities equal to 0.4 and 0.6 respectively, and for simplicity we
assume the same variogram model (spherical and isotropic) with the same parameters
for both. We then simulate a 2D map of facies and porosity according to the proposed
methodology (Figure 3.1). The simulation grid is 70x70 and the variogram range of
porosity is 4 grid blocks in both directions. The simulation can be performed with
or without conditioning hard data; in the example of Figure 3.1, we introduced four
porosity values at four locations that are used to condition the simulations, and we
generated a set of 100 conditional realizations (Figure 3.1). When hard data are
assigned, the weights of the mixture components are determined by evaluating the
prior Gaussian components at the hard data location and discrete property values
are determined by selecting the most likely component.
As we previously mentioned, the methodology is similar to Hansen et al. (2006),
but the use of Gaussian Mixture models allows us to describe the multi-modality of the
data and to simulate at the same time both the continuous and the discrete variable.
SGMixSim requires a spatial model of the continuous variable, but not a spatial model
of the underlying discrete variable: the spatial distribution of the discrete variable
only depends on the conditional weights of the mixture (Eq. 3.27). However, if the
mixture components have very different probabilities and very different variances (i.e.
when there are relatively low probable components with relatively high variances),
the simulations may not accurately reproduce the global statistics. If we assume,
for instance, two components with prior probabilities equal to 0.2 and 0.8, and we
assume at the same time that the variance of the first component is much bigger than
the variance of the second one, then the prior proportions may not be honored. This
problem is intrinsic to the sequential simulation approach, but it is emphasized in case
of multi-modal data. For large datasets or for reasons of stationarity, we often use
a moving searching neighborhood to take into account only the points closest to the
location being simulated Goovaerts (1997). If we use a global searching neighborhood
(i.e. the whole grid) the computational time, for large datasets, could significantly
increase. In the localized sequential algorithm, the neighborhood is selected according

to a fixed geometry (for example, ellipsoids centered on the location to be estimated)
and the conditioning data are extracted by the linear operator (Theorem 3.4) within
the neighborhood. When no hard data are present in the searching neighborhood and
the sample value is drawn from the prior distribution, the algorithm could generate
isolated points within the simulation grid. For example, a point drawn from the first
component could be surrounded by data, subsequently simulated, belonging to the
second component, or viceversa. This problem is particularly relevant in the case of
multi-modal data especially in the initial steps of the sequential simulation (in other
words when only few values have been previously simulated) and when the searching
neighborhood is small.
To avoid isolated points in the simulated grid, a post-processing step has been
included (Figure 3.1). The simulation path is first revisited, and the local conditional
probabilities are re-evaluated at all the grid cells where the sample value was drawn
from the prior distribution. Then we draw again the component from the weights of
the re-evaluated conditional probability. Finally, we introduce a kriging correction of
the continuous property values that had low probabilities in the neighborhood.
Next, we show two applications of linearized sequential inversion with Gaussian
Mixture models obtained by applying Theorem 3.4. The first example is a rock
physics inverse problem dealing with the inversion of acoustic impedance in terms of
porosity. The methodology application is illustrated by using a 2D grid representing a
synthetic system of reservoir channels (Figure 3.2). In this example we made the same
assumptions about the prior distribution as in the previous example. As in traditional
sequential simulation approaches, the spatial continuity of the inverted data depends
on the range of the variogram and the size of the searching neighborhood; however,
Figure 3.2 clearly shows the multi-modality of the inverted data. Gaussian Mixture
models can describe not only the multi-modality of the data, but they can better
honor the data correlation within each facies.
The second example is the acoustic inversion of seismic amplitudes in terms of
acoustic impedance. In this case, in addition to the usual input parameters (prior
distribution and variogram models), we have to specify a low frequency model of
Figure 3.1: Conditional realizations of porosity and reservoir facies obtained by SG-
MixSim. The prior distribution of porosity and the hard data values are shown on
top. The second and third rows show three realizations of porosity and facies (grey is
shale, yellow is sand). The fourth row shows the posterior distribution of facies and
the ensemble average of 100 realizations of facies and porosity. The last row shows
the comparison of SGMixSim results with and without post-processing.
impedance, since seismic amplitudes only provide relative information about elastic
contrasts and the absolute value of impedance must be computed by combining the
estimated relative changes with the low frequency model (often called prior model in
seismic modeling). Once again, the discrete variable is identified with the reservoir
facies classification. In this case shales are characterized by high impedance values,
and sand by low impedances. The results are shown in Figure 3.3. We observe that
even though we used a very smoothed low frequency model, the inverted impedance
log has a good match with the actual data (Figure 3.3), and the prediction of the
discrete variable is satisfactory compared to the actual facies classification performed
at the well. In particular, if we perform 50 realizations and we compute the maxi-
mum a posteriori of the ensemble of inverted facies profiles, we perfectly match the
actual classification (Figure 3.3). However, the quality of the results depends on the
separability of the Gaussian components in the continuous property domain.
Figure 3.2: Linearized sequential inversion with Gaussian Mixture models for the
estimation of porosity map from acoustic impedance values. On top we show the true
porosity map and the acoustic impedance map; on the bottom we show the inverted
porosity and the estimated facies map.
Finally we applied the Gaussian Mixture linearized sequential inversion to a layer

map extracted from a 3D geophysical model of a clastic reservoir located in the
North Sea (Figure 3.4). The application has been performed on a map of P-wave
velocity corresponding to the top horizon of the reservoir. The parameters of the
variogram models have been assumed from existing reservoir studies in the same area.
In Figure 3.4 we show the map of the conditioning velocity and the corresponding
histogram, two realizations of porosity and facies, and the histogram of the posterior
distribution of porosity derived from the second realization. The two realizations have
been performed using different prior proportions: 30% of sand in the first realization
and 40% in the second one. Both realizations honor the expected proportions, the
multi-modality of the data, and the correlations with the conditioning data within
each facies.
1000
Actual data
Prior model
1020
Realization 1
Inverted data
1040
1060
1080
Time (ms)
1100
1120
1140
1160
1180
4000 8000 12000 −200 0 200 Facies Maximum a posteriori Actual facies
P−impedance Seismic amplitudes (Realization 1) of facies classification
3
(m/s g/cm )
Figure 3.3: Sequential Gaussian Mixture inversion of seismic data (ensemble of 50

realizations). From left to right: acoustic impedance logs and seismograms (actual
model in red, realization 1 in blue, inverted realizations in grey, dashed line represents
low frequency model), inverted facies profile corresponding to realization 1, maximum
a posteriori of 50 inverted facies profiles and actual facies classification (sand in yellow,
shale in grey).
Figure 3.4: Application of linearized sequential inversion with Gaussian Mixture mod-
els to a reservoir layer. The conditioning data is P-wave velocity (top left). Two
realizations of porosity and facies are shown: realization 1 corresponds to a prior
proportion of 30% of sand, realization 2 corresponds to 40% of sand. The histograms
of the conditioning data and the posterior distribution of porosity (realization 2) are
shown for comparison.
3.6 Conclusion
In this chapter, we proposed a methodology to simultaneously simulate both continu-
ous and discrete properties by using Gaussian Mixture models. The method is based
on the sequential approach to Gaussian Mixture linear inverse problem, and it can
be seen as an extension of sequential simulations to multi-modal data. Thanks to

the sequential approach used for the inversion, the method is generally quite efficient
from the computational point of view to solve multi-modal linear inverse problems
and it is applied here to reservoir modeling and seismic reservoir characterization.
We presented four different applications: conditional simulations of porosity and fa-
cies, porosity-impedance inversion, acoustic inversion of seismic data, and inversion
of seismic velocities in terms of porosity. The proposed examples show that we can
generate actual samples from the posterior distribution, consistent with the prior in-
formation and the assigned data observations. Using the sequential approach, we can
generate a large number of samples from the posterior distribution, which in fact are
all solutions to the Gaussian Mixture linear problem.
Chapter 4
Stochastic inversion for facies

modeling
4.1 Abstract
The main objective of this work is to present a new stochastic methodology for seis-
mic reservoir characterization that combines advanced geostatistical methods with
traditional geophysical models, in order to provide fine-scaled reservoir models of fa-
cies and reservoir properties, such as porosity, and net-to-gross. The methodology
we propose is a stochastic inversion where we simultaneously obtain earth models of
facies, rock properties, and elastic attributes. It is based on an iterative process where
we generate a set of models of reservoir properties by using sequential simulations,
calculate the corresponding elastic attributes through rock physics relations, compute
synthetic seismograms and, finally, compare these synthetic results with the real seis-
mic amplitudes. The optimization is performed through a stochastic technique, the
probability perturbation method, that perturbs the probability distribution used to
generate the initial realization and allows obtaining a facies model consistent with
all available data through a relatively small number of iterations. The probability
perturbation approach is based on a probabilistic method called Tau model, which
provides an analytical representation to combine single probabilistic information into
a joint conditional probability. The advantages of probability perturbation method
67
CHAPTER 4. STOCHASTIC INVERSION FOR FACIES MODELING 68
are that it transforms a 3D multiparameter optimization problem into a set of 1D

optimization problems and it allows us to include several probabilistic information
through the Tau model. The method has been tested on a synthetic case where we
generated a set of pseudo-logs and the corresponding synthetic seismograms. We
then applied the method to a real well profile, and finally extended it to a 2D seismic
section. The application to the real reservoir study includes data from three wells
and partially stacked near and far seismic sections, and provided as a main result the
set of optimized models of facies, and of the relevant petrophysical properties, to be
used as initial static reservoir models for fluid flow reservoir simulations.
4.2 Introduction
One of the aims of reservoir modeling is to describe the spatial variability of reservoir
properties: facies, and the corresponding petrophysical properties, such as porosity,
permeability, net-to-gross, and fluid saturation. The estimation of reservoir properties
from seismic data is a complex underdetermined non-linear inverse problem. Several
techniques, both deterministic and probabilistic, have been developed to solve the
problem and estimate the optimal reservoir model Bosch et al. (2010) to be used as
initial model in fluid flow simulations. We can classify all the existing methodolo-
gies in two categories: 1) multi-step inversion methods and 2) stochastic inversion
approaches.
In multi-step inversion methods, the problem of estimating reservoir properties
from seismic data is split into two or more sub-problems: generally elastic properties
are first derived from partial stacked seismic data by elastic inversion; then facies
are pointwise classified from the resulting volumes of elastic attributes by statistical
techniques, such as, for example, discriminant analysis, neural networks, or Bayesian
classification (see Avseth et al. (2001), and Mukerji et al. (2001a)). If a Bayesian
elastic inversion (Buland and Omre (2003)) is performed, we obtain in the first step
a set of volumes of probability of elastic properties which can be used with a suitable
likelihood function to classify seismic facies through the Bayesian approach (Doyen
(2007)). In more recent approaches, reservoir properties such as porosity and clay
content are estimated from inverted seismic velocities (Grana et al. (2009)) and fa-
cies distribution can be subsequently derived from the reservoir properties volume.
Similarly, in Grana and Della Rossa (2010), a three-step probabilistic approach based
on Gaussian mixture models is introduced to estimate the probability of seismic
elastic attributes, reservoir properties, and litho-fluid classes or facies. The tradi-
tional Bayesian framework (Tarantola (2005)) has been also adopted for problems of
litho-fluid prediction from seismic data, as presented in Buland et al. (2008). The
probabilistic approach allows us to correctly propagate the uncertainty associated
with input data and physical model approximations to the posterior probability of
reservoir properties. For non-Gaussian posterior distributions, different estimators
can be used to obtain the most likely model, such as mean, median or maximum a
posteriori. However if the inversion does not include any sampling method of the
posterior distribution, the resolution of the estimated properties is the same as that
of the input conditioning data (seismic amplitudes) and the final volumes of facies
and reservoir properties are representative of a coarser scale than the characteristic
scale of reservoir dynamic models. As a consequence, these methodologies require
the integration with geostatistical methods to include seismic inversion results into
reservoir models (e.g., Mukerji et al. (2001a)). The most common strategy (Doyen
(2007)) is to perform sequential simulations to generate high resolution facies models,
by conditioning the simulation with the ”coarse-scaled” volume of facies estimated
from seismic. Facies models can be generated by either two-point (sequential indi-
cator simulation, see e.g. Deutsch and Journel (1992)) or multi-point geostatistics
(single normal equation simulation, see e.g. Remy (2009)). Both methods allow one
to include secondary information derived from seismic data to condition the simu-
lations. The corresponding models of continuous reservoir properties are generated
by sequential Gaussian simulation, conditioned by the facies model. Other methods
have been recently proposed, mainly in reservoir history matching, including geome-
chanical models to condition reservoir simulations (Wilschut et al. (2011)).
On the other hand, stochastic inversion approaches are generally based on the
iterative application of a forward model and the inversion step is performed using
deterministic or stochastic optimization techniques: in particular models of subsur-

face properties (facies and rock properties) are generated; then suitable rock physics
transforms are applied to generate the corresponding volumes of the elastic proper-
ties; finally synthetic seismic volumes are computed and compared to real seismic
data to evaluate the mismatch. The initial models are usually generated using pre-
viously mentioned geostatistical techniques (sequential indicator simulation or multi-
point geostatistics, and sequential Gaussian simulation) to create fine-scaled reservoir
models (González et al. (2008)). The final model is found by applying a suitable op-
timization method. The main limitation of stochastic inversion techniques is that
the optimization step, in real applications, can be computationally expensive. The
optimization cannot be applied independently point-by-point, because the objective
function depends on seismic data which represent coarse scale information reflecting
contrast between sub-surfaces. Moreover, as the solution of the inverse problem could
have local mimima, the final model could depend on the initial model. Different ini-
tial models could lead to different optimized models with the same seismic response
especially when layers thinner than the seismic resolution are included in the reservoir
model. Different optimization methods can be used. In González et al. (2008), the
optimization is deterministic and it is performed trace-by-trace, and the optimized
profile at the current trace is then used to condition the following simulations. Other
methods have been introduced: for example, Bosch et al. (2009) propose an iterative
optimization based on Newton’s method to simultaneously update the multi-property
model. Another family of stochastic inversion approaches is based on Markov chain
Monte Carlo methods (Eidsvik et al. (2004a), Larsen et al. (2006), Gunning and
Glinsky (2007), Rimstad and Omre (2010), Ulvmoen and Omre (2010), Hansen et al.
(2012)). Several geostatistical methods have been proposed to generate ensembles
of reservoir property realizations: two-point geostatistics (for example, sequential
indicator and sequential Gaussian simulations, or pluri-Gaussian methods, Doyen
(2007)) and multi-point geostatistics methods (González et al. (2008)) are the most
common. Some of these methods have been combined with optimization techniques
(such as simulated annealing, genetic algorithms, gradual deformation, probability
perturbation method and ensemble Kalman filter) to obtain optimal models of reser-
voir properties. These methods have been applied to elastic inversion or simultaneous
inversion of elastic and reservoir properties and facies (see Sams et al. (1999), Merletti
and Torres-Verdin (2006), and Sams et al. (2011)).
We propose here a new approach that aims at estimating fine-scaled reservoir
models in a stochastic inversion by combining geostatistical methods, such as sequen-
tial simulations (Deutsch and Journel (1992)) and a stochastic optimization tech-
nique called probability perturbation method (PPM, Caers and Hoffman (2006)),
with classical geophysical methods, such as seismic convolution and rock physics
models (Mavko et al. (2009). Our methodology is mainly aimed to determine the
optimal facies model for the reservoir. In our approach, we use sequential indicator
simulation (SISim, Journel and Gomez-Hernandez (1989)) to generate facies models,
and the probability perturbation method to perturb the probability used in SISim.
At each optimization step, a new facies model is generated; reservoir properties, in
particular porosity and clay content (or net-to-gross), are then simulated by sequen-
tial Gaussian simulation conditioned by the facies distribution; elastic properties are
subsequently calculated by applying a rock physics model and converted in the corre-
sponding time domain; and finally the synthetic seismic response is computed with a
traditional convolutional model (Figure 4.1). The optimization objective function is
the 2-norm of the difference between the synthetic seismic and the real seismic data.
A similar approach has been presented in González et al. (2008), with the target being
the direct inversion of facies with the integration of rock physics models and multi-
point geostatistics. However, in their method, at each iteration of the optimization,
the perturbation of the facies model is performed directly on the realization, whereas
in our approach we perturb the underlying probability distribution used to generate
the model. We introduce the probability perturbation method to obtain the optimal
model in a reasonable number of iterations. In our methodology we also account
for non-stationarity by introducing an additional probability distribution that can be
derived from different sources (seismic or AVO attributes, for example).
We first apply the stochastic inversion to a synthetic case with the objective of re-
constructing the actual facies classification, in order to test the validity of the method.
Figure 4.1: Workflow of stochastic inversion.
As the proposed methodology includes different methods and models and several pa-
rameters have to be calibrated or assumed from prior knowledge or information from
nearby fields, we propose a sensitivity analysis that investigates their effect on the
corresponding estimations. The method is then applied to a well log profile and to
a 2D seismic section of a real seismic reservoir characterization study in the North
Sea (offshore Norway). Several studies have been published on a number of nearby
fields in the North Sea (Avseth et al. (2001), Mukerji et al. (2001a), and Avseth et al.
(2005)). In this example we integrated into the methodology a further probability
derived from seismic data by means of a traditional Bayesian approach, to speed up
the convergence and account for non-stationarity.
4.3 Methodology
The inversion methodology we propose attempts to directly integrate the petroelastic
model and facies classification into the seismic inversion workflow. The flowchart of
the method is shown in Figure 4.1. In the following subsections we will describe each
step of the method and the techniques that are used. The methodology is presented
for a clastic reservoir, but it can be adapted to different lithology reservoir conditions,
with the choice of a reliable rock physics model and a suitable facies classification.
4.3.1 Geostatistical methods

The application of geostatistics to reservoir modeling aims to integrate data from vari-
ous sources (well, seismic and production data) into a consistent model to describe the
rock properties of the reservoir and their spatial continuity. Sequential simulations
are geostatistical methods that can be used to generate realizations of a probabil-
ity density function of either discrete or continuous properties. These methods are
based on various stochastic algorithms and they are applied in reservoir modeling
to generate different realizations of reservoir properties. This procedure produces
high-resolution simulations of the property we are interested in, by sequentially vis-
iting the grid cells of a 1D, 2D or 3D space, along a random path. In each cell, the
simulated value is drawn from the local conditional distribution, which depends on
the prior distribution and on the previously simulated values in the neighborhood of
the given cell. This procedure is repeated for all the cells of the grid. The methods
nowadays available can be divided into two big categories: two-point geostatistics and
multi-point geostatistics. Two-point geostatistics algorithms are in general faster as
they only account for the correlation between two spatial locations at a time; the spa-
tial continuity of the property distribution being ensured by variogram models. On
the other hand, multi-point geostatistics takes into account the correlation between
multiple spatial points but as it is very complex to analytically treat the associated
conditional probability, the multi-point statistics are inferred from a training image
generated for example by unconditional Boolean modeling. In our approach we use
two-point geostatistics algorithms, but if a suitable training image is available, with
size at least larger than the reservoir size, then multi-point geostatistics could be
easily integrated.
The two most common algorithms in two-point geostatistics are: sequential in-
dicator simulation, SISim, and sequential Gaussian simulation, SGSim (see Deutsch
and Journel (1992) or Goovaerts (1997)). Sequential indicator simulation deals with
discrete random variables, for example facies in reservoir modeling, while sequential
Gaussian simulation deals with continuous random variables, for example porosity.
In our approach, facies are first simulated by SISim possibly with a secondary
conditioning data derived from seismic (seismic facies probability, for example); then
porosity is simulated by using SGSim. In particular, the simulation of porosity for
each facies is performed independently of the simulations for other facies; each sim-
ulation is performed over the whole 3D grid, then the simulations are re-assembled
into the final simulated porosity realization according to the facies classification. To
grid-cells not belonging to the reservoir layer, a constant value of porosity (equal or
close to 0) is assigned. Finally other reservoir properties (for example, net-to-gross,
irreducible water saturation, and even permeability if necessary) are subsequently sim-
ulated by sequential Gaussian co-simulation (CoSGSim) with porosity distribution or
previously simulated properties, as secondary information. In CoSGSim, a continuous
variable is simulated by accounting for the correlation with another variable: this is
the case for example of porosity and net-to-gross or porosity and permeability.
At the end of this step, we obtain a reservoir model of facies, porosity φ, and
net-to-gross ntg. In our approach, net-to-gross is converted into clay content (vclay )
by assuming vclay = 1 − ntg. Fluid saturation distributions (water, oil and gas satu-
rations, namely sw, so, sg) could be simulated as well, but in order to obtain realistic
litho-fluid models we prefer to impose in the model the oil- (or gas-) water contact and
the gas-oil contact (if present) and assume a constant distribution of the fluid within
the so-obtained fluid layers. To increase the realism of the model, we could however
simulate the irreducible water saturation through sequential Gaussian simulation or
co-simulation or deterministically distribute it by assuming empirical relations with
other properties such as porosity or permeability. One of the favorable features of
sequential simulation is the ability to incorporate different types of conditioning data.
However, if A is the unknown property and B and C are the conditioning data (for
example hard data B and soft data C), then in many cases the analytical expression
of the posterior probability P (A|B, C) can be difficult to obtain. Journel (2002) has
proposed an efficient method to integrate secondary data (soft data C) in the proba-
bility model P (A|B) to get the posterior probability P (A|B, C). To combine P (A|B)
and P (A|C) into P (A|B, C), Journel (2002) proposes the following expression:
1
P (A|B, C) = (4.1)
1+x
where τ1
x b c τ2
= (4.2)
a a a
and
1 − P (A) 1 − P (A|B) 1 − P (A|C)
a= , a= , c= (4.3)
P (A) P (A|B) P (A|C)
P (A) is the prior distribution of the unknown property. The ratios a, b, and c can
be interpreted as the distance to an event occurring. For example the ratio a is the
distance to the event A occurring, prior to knowing the information associated with
B and C: if P (A) = 1, then a = 0 and A is certain to occur. The parameters
τ1 and τ2 account for the redundancy for each set of conditioning data B and C
(Krishnan (2008)). Setting τ1 = τ2 = 1 is equivalent to assuming a form of conditional
independence between P (B|A) and P (C|A) expressed in terms of permanence of ratio
(Eq. 4.2). In other words we assume that the incremental contribution of data event C
to knowledge of A is the same after or before knowing B; this assumption is however
less restrictive than assuming independence between data B and C.
The parameter τ2 can be modified to tune the contribution of the conditioning
data C: if τ2 > 1 then the influence of C is increased (in our context, this could be the
case where C is crosswell seismic where the resolution is higher than common surface
seismic); if 0 < τ2 < 1 then the influence of C is decreased (this could be the case
where the quality of the seismic is not optimal and the low resolution of seismic could
obscure facies transitions). In our work we assume τ1 = τ2 = 1, but a preliminary
sensitivity analysis at the well location is necessary to investigate the effect of these
parameters. Several possible definitions are proposed for the information content
measure related to Tau-parameters (Liu (2003)), but the determination of the optimal
parameter values is still object of research.
Tau-model formulation allows avoiding the computation of the probabilities P (B|C)
and P (C|A, B) or P (B|A, C) that appear in the exact decomposition of P (A|B, C)
and that are generally more difficult to calculate. In our application the unknown
property A is the facies classification within the reservoir, B are the hard data and C
is the seismic information.
4.3.2 Geophysical forward model

In order to compute the seismic response of the earth models generated by sequential
simulations, we first calculate the elastic properties, such as velocities or impedances,
within the reservoir model and subsequently compute the corresponding seismic sig-
nature. We observe that in many real applications, the geocellular grid (corner-point
grid) used to model petrophysical and dynamic properties in the reservoir does not
coincide with the seismic grid. In particular geocellular grid cells are usually larger
than the bin size of the seismic survey, which calls for a downscaling of the grid
(Castro et al. (2009)). Furthermore, the velocity models must be converted from
depth domain to time domain, in order to perform seismic convolution and obtain
synthetic seismic volumes. Depth-to-time conversion necessarily requires an accu-
rate background velocity model, which is consistent with the seismic processing steps
performed on the real seismic dataset.
Elastic properties are usually computed through a rock physics model. This model
is a set of equations that transforms petrophysical variables, typically porosity, miner-
alogy (clay and sand content, for example, in clastic reservoirs), and fluid saturations,
into elastic properties, such as P-wave and S-wave velocities and density (or, as in
many practical applications, P- and S-impedance). The rock physics model type de-
pends on the reservoir rocks we are dealing with: the set of equations can be a simple
regression on well data or a more complex physical model (Mavko et al. (2009)). Gen-
erally the model is first calibrated on well logs, where both petrophysical and elastic
properties are available: in fact most of the models traditionally used contain one or
more parameters (for example critical porosity and coordination number for granular
media models, Mavko et al. (2009)) that should be determined from core analysis or
estimated by comparing well logs and rock physics model predictions. Once the rock
physics model has been calibrated on well logs, the model is applied, point-by-point,
to the volumes of reservoir properties generated by sequential simulations. In the case
of a clastic reservoir, these properties are generally porosity φ, clay content vclay , and
fluid saturations sw, so, and sg. Clay content is usually computed from net-to-gross
as vclay = 1 − ntg; this assumption does not account for mineralogical texture related
to laminated and dispersed shale that can influence the elastic properties of the rock.
However if more accurate petrophysical relations linking net-to-gross to lithological
properties are available, they can be included in the forward model. The rock physics
forward model can be described as follows: first we estimate the elastic properties
of the solid phase, i.e. bulk and shear modulus of the matrix, Kmat and µmat , and
density ρmat , by using solid phase mixing laws (Voigt-Reuss-Hill average or Hashin-
Shtrikmann bounds); then we compute the elastic properties of the fluid phase, i.e.
bulk modulus Kf l and density ρf l , by using fluid mixing laws (Reuss average or Brie’s
law); dry rock properties are then computed from solid phase properties by using one
of the available theories in literature (for example, granular media or inclusion mod-
els) to obtain dry rock bulk and shear moduli, Kdry and µdry ; finally the saturated
rock properties, Kdat and µsat , are calculated by Gassmann’s equations (Mavko et al.
(2009)). Density of the saturated rock is computed as a linear combination of matrix
density ρmat and fluid density ρf l weighted by their respective volume fractions
ρ = φρf l + (1 − φ)ρmat (4.4)
and P-wave and S-wave velocities are calculated by definition as function of saturated
elastic properties, Kdat and µsat , and density ρ:
s
Ksat + 4/3µsat
VP = (4.5)
ρ
µsat
r
VS = (4.6)
ρ
The result of this first step of the forward modeling is a set of volumes of elastic
attributes: typically, P-wave and S-wave velocity, VP and VS , and density ρ, or P-
and S-impedances IP and IS . The rock physics model can be facies dependent, such
as in the field study we propose in the Application section. Data are subsequently
re-sampled in the seismic grid and converted from depth to time. In our workflow,
we downscaled the data using a subsampling technique, but more advanced down-
scaling methods could be used (Castro et al. (2009)). Depth-to-time conversion can
be performed by applying a velocity model obtained by collocated cokriging using
the stacking velocity volume used for the processing of seismic amplitudes and sonic
well-logs filtered at low frequency.
The second phase of the forward modeling step is the computation of the synthetic
seismic signature. If partial stacked seismic data (dobs (θ)) are available, then we
compute the corresponding angle stacks, otherwise if only post stack data are available
we compute only the zero angle seismic traces. We describe here the partial stack case,
as the post stack can be seen as a particular case of this application. Synthetic seismic
traces (dsynth (θ)) are computed here by seismic convolution: the forward modeling is
based on a convolutional model and Zoeppritz equations (Aki and Richards (1980)).
Specifically, at each trace, the synthetic seismogram is computed by convolving the
wavelets (estimated from the real dataset) with the reflection coefficients series:
dsynth (t, θ) = wsynth (t, θ) ∗ RP P (t, θ) (4.7)
where t it the travel-time, wsynth (t, θ) is the vector of the angle-dependent wavelets
and RP P (t, θ) the vector of reflection coefficients. Seismic reflection coefficients
RP P (t, θ) depend on the angle and the material properties of the subsurface: an
isotropic, elastic medium is completely described by P-wave and S-wave velocity and
density. For angles smaller than the critical angle of the seismic dataset, we can al-
ternatively use a linearized weak contrast approximation of Zoeppritz equations (Aki
and Richards (1980)).
We point out that in order to perform the convolution, we estimate the wavelets
independently for each available angle gather. Both forward models, seismic con-
volution and rock physics model, lead to underdetermined inverse problems. In our
approach, in order to solve these problems we adopt a stochastic optimization method

in the inversion.
4.3.3 Stochastic optimization algorithm

The forward model results are included in an optimization loop to find the optimal
model of facies (F). We perform a seismic-driven stochastic optimization where the
objective function is the 2-norm of the difference between synthetic seismic data
(dsynth ) and the real seismic amplitudes (dobs ).
The stochastic optimization algorithm we used in our methodology is based on
the probability perturbation method (Caers and Hoffman (2006)) and Tau model
(Journel (2002) and Krishnan (2008)). In our approach the target proportions and
the variogram models are assumed to be assigned, but they could be made stochastic
and optimized simultaneously. However preliminary sensitivity tests in 1D showed
that the convergence of the algorithm could be more than 10 times slower.
We describe the methodology for a generic reservoir with NF facies. In particular
the categorical variable F can assume NF possible values fk (for k = 1, , NF ). The
facies value at a given location is coded using a set of indicator variables i(u, fk )
(
1 if fk occurs at u
i(u, fk ) = (4.8)
0 otherwise
where u = (x, y, z) denotes a generic spatial location, corresponding to a grid cell in

the reservoir grid.
We first select a random seed and determine a random path of simulation; then we
generate an initial realization of facies by using SISim and according to the selected
variogram. We then simulate porosity by using SGSim, cosimulate other rock prop-
erties by using co-SGSim and apply the forward model to compute elastic properties
and synthetic seismic data. This realization honors the hard data (for example the fa-
cies profiles at the well locations) but it does not necessarily match the seismic data.
The initial realization is then perturbed: in the probability perturbation method
rather than perturbing the initial realization directly, we propose a perturbation of
the probability model used to generate the realization. We denote the underlying
probability model of SISim as P (Fk |G) where G indicates the hard data (well data
or previously simulated values). We remind that, in SISim, the probability of facies
given hard data G is obtained at each location by indicator kriging.
At the following step of the optimization we propose a new probability, P (Fk |dobs ),
as a linear combination of the indicator model i0 (u, fk ) associated to the initial real-
ization and the prior probability P (Fk ) of the facies:
P (Fk |dobs ) = (1 − r)i0 (u, fk ) + rP (Fk ) (4.9)
where r is the scalar deformation parameter to be optimized between 0 and 1

(Caers and Hoffman (2006)). In our approach, we used a uniform discretization of
the interval [0, 1] with spacing 0.1, which means that we evaluated the probability in
Eq. 4.9 for 11 values of the parameter r. For each facies, the probability in Eq. 4.9 is
a function of spatial location u, but for simplicity of notation we omitted the spatial
dependency.
At each spatial location u, we now have to combine P (Fk |dobs obtained from
Eq. 4.9, with the prior probability P (Fk ) and the probability P (Fk |G) obtained from
indicator kriging conditioned to hard data, to obtain the probability P (Fk |G, dobs ).
This is done by using the Tau-model (Journel (2002)):
τ1
1 b c τ2
P (Fk |G, dobs ) = , x=a (4.10)
1+x a a
where τ1 and τ2 are the Tau-model parameters, and a, b, and c, are obtained by
Eq. 4.3 with A = Fk , B = G, and C = dobs . In the case τ1 = τ2 = 1 (i.e, in case of
conditional independence) Eq. 4.10 simplifies as follows:
a
P (Fk |G, dobs ) = , (4.11)
a + bc
We sample from the distribution P (Fk |G, dobs ) to generate a new facies model
ir (u, fk ) and we repeat the above described reservoir modeling, by simulating rock
properties, computing elastic attributes and synthetic seismic data dsynth (θ, r). For
each facies, the probability of Eq. 4.10 depends on the scalar parameter r, in other
words Eq. 4.3 provides a set of distributions, and the forward model result is a set of
models that depends on the deformation parameter r. For each model we calculate
the objective function
Nθ
X
O(r) = ωθk ||dobs (θk ) − dsynth (θk , r)||2 (4.12)
k=1
where Nθ is the number of angle stacks and the ωθk are the weights assigned to
the different angle stacks based on the quality of the seismic dataset. For example
we could choose the weights directly proportional to the signal-to-noise ratio of the
angle stacks. We finally perform a 1D optimization on the deformation parameter
r. However the search space provided by the set of distributions in Eq. 4.9 is too
limited because it is obtained as linear deformation of two realizations. Thus we
introduce another optimization loop where we change the random seed and the opti-
mal realization ir (u, fk ) obtained at the previous step replaces the initial realization
i0 (u, fk ).
The optimization step is indeed performed within two nested loops. In the outer
loop, we change the random seed until a good match between the synthetic seismic
traces of the trial model and the real seismic traces is achieved. At each step we
perform a 1D optimization (inner loop) on the deformation parameter r of the proba-
bility perturbation method to obtain the parameter that minimizes the error between
the synthetic and real seismic data. If the error of the new model is less than the
error of the previous model, we accept the new model and we set i0 (u, fk ) = ir (u, fk )
otherwise we change the random seed, and we repeat the previously described steps.
We iterate this procedure until the error is less than a fixed tolerance value T , which
can be selected depending on the quality of seismic data, for example in terms of
signal-to-noise ratio.
The basic structure of the algorithm (Figure 4.2) can be described as follows:
1. Select a random seed, generate an initial realization of facies (namely i0 (u, fk ))

using SISim; simulate rock properties and apply the geophysical forward model;
2. Perform a seismic-driven stochastic optimization using the probability pertur-

bation method:
(a) in the outer loop change the random seed and iterate to obtain a good
match, i.e O(r) < T ;
(b) in the inner loop perform a 1D optimization to obtain the optimal defor-
mation parameter r.
In this inner loop, we propose a new probability P (Fk |dobs ), obtained as a

linear combination of the realization i0 (u, fk ) and the prior probability P (Fk )
of the facies (Eq. 4.9), we compute the conditional probability P (Fk |G, dobs )
by using Tau-model (Eq. 4.10), we generate a new facies model ir (u, fk ), apply
the forward model and evaluate the objective function of the objective function
of Equation 4.12. If O(r) < T , then we stop the algorithm, otherwise we set
ir (u, fk ) = i0 (u, fk ) and repeat the procedure (steps 2a and 2b), with a different
random seed.
Figure 4.2: Flowchart of PPM algorithm.

Two-point geostatistics
Sequential Gaussian simulation (SGSim) is a geostatistical method that allows us to

generate several realizations of a continuous property which honor: 1) the hard data
(if any) used to condition the simulation, 2) the target distribution (if the distribution
of the property is not already a standard Gaussian pdf) and 3) the variogram model
which describes the spatial continuity of the property. In SGSim we sequentially visit
the grid cells along a random path. At each cell, we simulate a value by sampling
from a Gaussian distribution with mean equal to the kriging estimate at that location
and variance equal to the kriging variance. Kriging estimate and variance at the
given location are computed by solving the kriging system (Goovaerts (1997)). The
procedure is repeated for all the cells in the grid. As in many practical applications,
when the target distribution is not Gaussian, we usually apply a preliminary normal
score transformation, perform the simulation and back-transform the results at the
end. Sequential Gaussian simulation provides higher detailed map of the simulated
property compared to the corresponding smoothed kriging map. Multiple simulations
are generated by changing the random seed.
Sequential indicator simulation can be seen as a generalization of SGSim. It is
based on the concept of indicator variable, i.e. a binary variable, which is the indicator
of occurrence of an event. SISim is generally applied to simulate discrete properties.
The probability of a certain cell assuming a certain value of a discrete property,
given the set of neighboring values, is calculated by indicator kriging. In indicator
kriging we estimate the probability of a certain categorical event at a given location
as a weighted linear combination of the indicator data falling within the searching
neighborhood. As in traditional kriging the weights are obtained by solving the linear
system of kriging equations which accounts for the indicator spatial covariance model
(Goovaerts (1997)). The methodology relies on the result that the expected value of
a binary indicator is the probability of the corresponding categorical event occurring;
kriging, as a least square error estimation method, allows calculating this probability.
At each generic location xn of the random path we compute the indicator kriging
probability pIK
fk of the generic categorical event fk :
n−1
X
pIK
fk = πfk + wi (i(xi , fk ) − πfk ) (4.13)
i=1
where πfk is the prior proportion of the categorical event fk , i(xi , fk )i=1,,n−1 are the
indicators of the previously simulated values at the locations xi ; and wi are the
kriging weights. By computing this distribution at each location and sampling from
it, SISim allows us to perform simulations of discrete variables such as facies or litho
fluid-classes.
4.3.4 Secondary information

To speed up the convergence we can include a further probability term in the Tau
model, for example, the probability of facies obtained by inverted seismic attributes,
P ∗ (Fk |S), where P ∗ represents the pointwise probability of facies conditioned by a
set of seismic attributes S at seismic scale, i.e. at low resolution. The probability of
facies at seismic scale can be obtained by using different methods and it can be con-
ditioned by different data, for example seismic impedances, P ∗ (Fk |IP , IS ), or seismic
amplitudes, P ∗ (Fk |dobs ), as in Grana and Della Rossa (2010), or AVO properties R0
and Gr , P ∗ (Fk |R0, Gr ) (see Mukerji et al. (2001a)). This step allows us to account
for low resolution secondary information, i.e. the probability of facies conditioned by
seismic, which improves the convergence speed and accounts for non-stationarity of
the data. The star symbol in the following will indicate the probability of facies at
seismic scale used as a low resolution trend to condition stochastic inversion simula-
tions. This probability can be integrated into the workflow in different ways: in our
method we used the Tau-model, by modifying Eq. 4.10 as follows:
τ1 τ3
1 b c τ2 d
P (Fk |G, S, dobs ) = , x=a (4.14)
1+x a a a
where
1 − P ∗ (Fk |, S)
d= (4.15)
P ∗ (Fk |, S)
However we point out that other methods could be adopted, such as, collocated
cokriging or Bayesian updating in sequential indicator simulations (Doyen (2007)).
Different sources of information can be used to obtain low resolution estimation of

facies distribution, i.e. to derive the probability of facies at seismic scale, P ∗ (Fk |S).
We adopt here a probabilistic approach to seismic facies classification consisting of
three main steps: seismic inversion to recover elastic attributes from seismic ampli-
tudes; estimation of petrophysical properties from elastic attributes; and facies classi-
fication to classify seismic facies from petrophysical properties (Grana and Della Rossa
(2010)). A full Bayesian approach has been adopted, based on the integration of the
probabilities obtained from Bayesian seismic inversion and statistical rock physics
model. First of all, a Bayesian seismic inversion is performed (Buland and Omre
(2003)) to obtain the probabilities of impedances from seismic amplitudes (step a).
Then a probabilistic characterization of petrophysical properties is applied (Grana
and Della Rossa (2010)) to estimate the probability of porosity, and clay content
(step b) by integrating the statistical rock physics model with the probabilities of
impedances obtained from Bayesian elastic inversion. Finally a probabilistic facies
characterization is performed (Grana and Della Rossa (2010)): the estimation of
facies probabilities P ∗ (Fk |S) conditioned by seismic attributes (step c) is obtained
combining petrophysical properties probabilities (step b), log-facies classification, and
seismic information from Bayesian elastic inversion (step a). The final results of this
probabilistic multi-step approach are the probability volumes P ∗ (Fk |S) of seismic
litho-facies, that are used in stochastic inversion as additional information in the Tau
model to condition the geostatistical simulations and account for non-stationarity.
In general P ∗ (Fk |S) can be obtained from any set of seismic attributes, such as
inverted impedances, AVO attributes, full waveform inversion properties; however
the weight of this information, i.e. the exponent τ3 in the Tau model (Eq. 4.14),
should be tuned after a sensitivity analysis, as all these properties are derived from
the same seismic dataset which appears in the optimization objective function. In
our application we tested the following set of parameters 0.5, 1, 2.5. By assuming
a high exponent τ3 , we increase the convergence speed but we tend to disregard the
prior information related to the spatial continuity model described by the variogram,
in particular we generally obtain a model with a resolution closer to the seismic one.
This result can be explained by the fact that we are accounting seismic information in
two different terms, c and d, with a relatively high weight in Eq. 4.14. According to
the theory, the parameters of the Tau-model measure the additional contribution of
the probabilistic information they are associated with. Since the parameter t3 is as-
sociated to the seismic-derived probability, and this information is already accounted
by the probability term P ∗ (Fk |dobs ), we tend to exclude values greater than or equal
to 1.
4.4 Application
The application of stochastic inversion is first presented in a one-dimensional synthetic
case to show the different steps of the method and verify its reliability. Then we
applied the methodology on a real reservoir study in the North Sea where a complete
dataset, including a set of well log data coming from three wells and partial stacked
seismic data, is available.
4.4.1 Synthetic case

For the synthetic test, we built a set of pseudo-logs to mimic a realistic depositional
turbiditic system along a well profile (namely well A). This synthetic case models a
clastic reservoir filled by oil: the top of the reservoir is supposed to be positioned at
around 2000m depth at the well location and the oil-water contact is fixed at 2100m.
We focus on a depth interval of 200m, by assuming a thick overcap of clay on top
of the reservoir. We created a synthetic facies sequence and a set of pseudo-logs
mimicking the behavior of rock and elastic properties associated to the facies profile;
and we finally generated synthetic seismic traces at three different angle stacks: 12o ,
24o and 36o. Three facies are defined: sand, silty-sand, and shale. We first modeled
the synthetic stratigraphic sequence in the well by a first-order Markov chain and
then distributed the corresponding rock properties.
Markov chains are a statistical tool that has been used in geophysics to simulate
facies sequences to capture the main features of the depositional process (Krumbein
and Dacey (1969)). Markov chains are based on a set of conditional probabilities that
describe the dependency of the facies value at a given location with the facies values
at the locations above (upward chain) or the locations below (downward chain). The
chain is said to be first-order, if the transition from one facies to another depends only
on the immediately preceding facies. The conditional probability of the transitions
are the elements of the so-called transition matrix P, where the generic element Pij
represents the probability of a transition from the facies i located above the interface
to the facies j located below. In our example, we estimated the input parameters, i.e.
prior proportions and transition probabilities, from a real well dataset. The transition
matrix is estimated by counting the number of transitions in the facies classification
at the well:  
0.9 0.05 0.05
 
P=
 0 0.93 0.07 
 (4.16)
0.05 0 0.95
Rows correspond to shale, silty-sand, and sand at the generic depth location z,
and columns correspond to shale, silty-sand, and sand at the generic depth location
z − 1. In other words, in our facies profile, we never have a shale on top of a silty-sand
or a silty-sand on top of a sand. The terms on the diagonal of the transition matrix
are related to the thickness of the layers: in fact the higher are the numbers on the
diagonal, the higher is the probability to observe no transition (i.e. high probability
that a facies has a transition to itself), and as a consequence the thicker will be the
layer. We define the first sample of the well profile as shale in order to have a shale
layer above the top of the reservoir. At the next step the facies value is sampled from
the conditional probability P (Fi |Fi−1 ), and we iterate the sampling till the bottom
of the interval (Figure 4.3).
Through this method we generated a facies profile which is assumed to be the
true model of this synthetic example. The facies proportions in this well profile are
0.28, 0.34, and 0.38 respectively for sand, silty-sand and shale. We then generated
pseudo-logs of rock properties, namely porosity and clay content (Figure 4.3). The
pseudo-logs must be vertically correlated within each facies and at the same time
they must be correlated between each other, as porosity in general almost linearly
depends on the clay content of the rock. In our dataset, we created the pseudo-logs
Figure 4.3: Synthetic well log dataset (well A), from left to right: effective porosity,
volume of clay, water saturation, P-wave and S-wave velocity, density and facies profile
(green represents shale, brown represents silty-sand, and yellow represents sand).
of porosity and clay content by sampling from three bivariate Gaussian distributions,
one for each facies. We assumed a correlation of 0.8 for every facies and we included
a vertical correlation by multiplying (Kronecker product) the covariance matrices of
each distribution by a spatial covariance matrix obtained from a 1D (vertical) expo-
nential variogram with correlation range 7.5 m (for every facies). The simulated logs
are finally re-assembled into the final pseudo-logs according to the facies classifica-
tion profile. Then we computed the corresponding pseudo-logs of density and elastic
properties, P-wave and S-wave velocity, by means of the soft sand model (Mavko
et al. (2009)). Rock physics models are generally good approximations of the elastic
behavior of rocks, but these relations cannot account for the heterogeneity and the
natural variability of the rocks in the subsurface. We then added a random error,
vertically correlated (the correlation range is 1 m), to mimic a more realistic behavior
similar to measured well log data. Traditional rock physics crossplots and the esti-
mated probability distributions are shown in Figure 4.4. Finally we computed the
synthetic seismograms corresponding to three angle stacks: 12o , 24o and 36o by using
three Ricker wavelets with three different center frequencies 30, 25, and 20 Hz PP
reflection coefficients have been computed with Aki-Richards approximation.
In the inversion methodology, we assume that the rock physics model is known (but
Figure 4.4: (Top) Rock physics crossplots of well dataset: clay content versus effec-
tive porosity (top left), P-wave velocity versus effective porosity (top right), S-wave
velocity versus P-wave velocity (mid left), and VP /VS ratio versus P-impedance (mid
right), color coded by facies classification (green represents shale, brown represents
silty-sand, and yellow represents sand). (Bottom) Joint probability of petrophys-
ical properties distribution: conditional probability contours color coded by facies
(bottom left), and joint probability surface (bottom right).
the error is unknown) and the three wavelets corresponding to the three angle stacks
are known as well. We finally assume that the facies proportions of the true model are
known and correspond to the real proportions of the well log profile. The variograms
of facies and rock properties used in sequential simulations and the cross-correlation
between rock properties have been estimated from the pseudo-logs (Figure 4.5). For
all the three facies we assumed a Gaussian model with the following correlation ranges:
3, 10, and 4, respectively for shale, silty-sand, and sand.
Figure 4.5: Variograms of porosity estimated at the well location, from top to bottom:
variogram of porosity in shale, silty-sand, and sand.
We show the results of the stochastic inversion methodology applied to the syn-
thetic well A, assuming perfect signal-to-noise ratio, in Figures 4.6 and 4.7. The
facies profile classified by our approach has a good match with the actual classifi-
cation; good results are obtained after only 25 iterations, which correspond to 275
evaluations of the forward model (since each iteration requires 11 forward model eval-
uations to locate the optimal parameter r). The more we proceed with the iterations,
the lower is the acceptance rate: in fact even if we perform 100 iterations the improve-
ment compared to the result obtained after 25 steps is quite small. In other words
the stochastic optimization used in this approach quickly reaches a sufficiently small
neighborhood of the minimum, then the convergence to the exact minimum becomes
slower. We point out that since seismic data are generally noisy, we do not want to
match perfectly the data, but only match the data within a certain tolerance (T ).
The convergence can be sped up, by introducing secondary information describing
the probability of facies at seismic frequency.
Figure 4.6: Stochastic inversion results at well location, from left to right: actual
facies classification, initial realization, and partial results of the optimization loop
after 3, 10 and 25 iterations classification (green represents shale, brown represents
silty-sand, and yellow represents sand). The last result (right plot) is the optimized
model according to the fixed tolerance.
We then applied the methodology several times by using different random seeds:
the result of the optimization after a fixed number of iterations will be statistically
similar but every time different in the details: in Figure 4.8, we show the variability
Figure 4.7: Synthetic seismograms (red) corresponding to the optimized model of

Figure 4.6 compared to input seismic traces (black). From left to right: near, mid
and far stack, corresponding to the incident angles of 12o, 24o , and 36o .
between the solutions we obtained by plotting 10 of 25 different runs; this variabil-

ity mainly depends on the tolerance T we fixed to satisfy the convergence criterion
(O(r) < T ).
Figure 4.8: Set of 10 different realizations obtained by stochastic inversion. We show

10 optimized models (obtained from 10 different runs) compared to the actual clas-
sification (green represents shale, brown represents silty-sand, and yellow represents
sand).
Finally we used this synthetic example to perform a sensitivity analysis on different

parameters (for each case, one parameter is changed at the time):
1. Signal to noise ratio of seismic data: SNR=5 (good quality seismic) or SNR=2.5;
2. Rock physics model: known (we use the same rock physics model used for the
model generation) or unknown;
T shale T silty-sand T sand

C shale 0.81 0.06 0.13
C silty-sand 0.11 0.74 0.15
C sand 0.07 0.05 0.88
Table 4.1: Confusion matrix of the reference case (T stands for true facies, C stands
for classified facies).
3. Wavelets: known (we use the same wavelets used for the synthetic seismic
generation) or estimated from real seismic data;
4. Variograms: real variogram or a different one with wrong correlation ranges;
5. Prior proportions of facies: real proportions or different ones;
6. Number of angle stacks: post stack seismic or partial stack seismic (2 or 3

angles);
7. Fluid effect: oil water contact known or ignored.
The results of the different cases have been quantitatively compared by computing
for each case the corresponding confusion matrix associated to the facies classification.
The confusion matrix is a tool used in supervised learning to visualize the quality of
the classification: in our application, each column of the matrix represents the per-
centage of samples in a predicted facies, whereas each row represents the percentage
of samples in the actual facies. The confusion matrix of the reference case shown in
Figure 4.6 (right plot) is summarized in Table 4.1. For all the three facies we obtain
a satisfactory reconstruction rate.
The results of the sensitivity tests are collected in Table 4.2, where we report
the main diagonal of the confusion matrix and the percentage of correctly classified
samples (normalized sum of the trace of the confusion matrix). Even though these
statistics are not exhaustive to evaluate the quality of the inversion, this sensitivity
analysis confirms that the number of angle stacks and the quality of the seismic data
are the major sources of uncertainty in seismic reservoir characterization studies. The
rock physics model is essential in this methodology; however the degree of accuracy of
Diagonal confusion matrix Percentage of identified samples

True model 1.00, 1.00, 1.00 100 %
Reference case 0.81, 0.74, 0.88 81.0 %
Low signal-to-noise (SNR=2.5) 0.76, 0.72, 0.73 73.6 %
Rock physics model (stiff sand) 0.92, 0.69, 0.78 79.6 %
Wavelets 0.79, 0.83, 0.77 79.6 %
Variogram (long range) 0.71, 0.48, 0.79 66.0 %
Variogram (short range) 0.76, 0.59, 0.69 68.0 %
Biased prior proportions 0.93, 0.52, 0.66 70.3 %
Number of angle stacks (Nθ = 1) 0.55, 0.33, 0.68 52.0 %
Number of angle stacks (Nθ = 2) 0.84, 0.49, 0.75 69.3 %
Fluid effect 0.80, 0.76, 0.77 77.6 %
Table 4.2: One-way sensitivity analysis of the synthetic inversion test: the first column
show the different cases, in the second column we report the main diagonal of the
confusion matrices of the different cases, in the third column we show the average of
the elements of the main diagonal (sum of the trace of the matrix normalized by the
number of facies).
the model is usually quite high as the model can be calibrated at the well locations by
using well data. Finally, in our tests, we observed that even if we underestimate the
correlation range of the variogram, we still obtain good results in the optimization;
conversely if we overestimate the range, then the optimization model cannot reproduce
the correct thickness of the sediments, but it tends to create thicker layers. However
we point out that these results are not completely general and they depend on the
parameters we chose. For example, if we use the wrong rock physics model but the
predictions are close enough to the observed data values (for example, a multi-linear
regression), or if we ignore the fluid effect but the velocity in hydrocarbon sand is
close to the velocity in brine sand, then the difference between the inversion results
with the correct parameters and the ones with wrong assumptions could be small.
Similarly the results of the sensitivity analysis on variogram parameters and prior
proportions could be worse, if we introduce a more significant bias in the parameters.
4.4.2 Real case

As a final step, we applied the methodology to a real seismic reservoir characterization
study, in the North Sea (offshore Norway). It is a deepwater clastic reservoir made
of sand and shale and filled by oil. Four seismic-scale sedimentary litho-facies can
be identified: soft sand and stiff sand (both filled by oil in the upper part of the
reservoir), silty-sand with clay dispersed in it, and shale. The reservoir is located
at approximately 2150m; the oil-water contact was measured at 2190m at the well
locations and it is supposed to be known in the inversion.
The data available are partial stacked seismic data (near stack corresponding to
8o , and far stack to 26o), a set of horizons for the reservoir level we are interested in,
and 5 well log datasets. Well 2 has been used as calibration well as it contains the
main acquired logs: P-wave velocity, S-wave velocity, gamma ray, density, neutron
porosity and resistivity. Data of well 2 and well 3 are used to condition the simulations
along the 2D line used to test the methodology in 2D. Finally well 5, that is located
outside the seismic survey, is used as an additional test.
First, the method has been tested at well 2 location to predict the facies distribu-
tion from seismic data. The litho-facies classification has been performed using sedi-
mentological information, core analysis, and a clustering technique applied to petro-
physical curves: effective porosity, volume of clay and volume of sand (Figure 4.9). A
rock physics model has been calibrated at the well location: the more suitable model
for this scenario is a constant-cement sand model (Avseth et al. (2005)). This model
is a combination of the contact-cement model and friable sand model (Mavko et al.
(2009)), where we assumed critical porosity equal 0.4 and coordination number 9
(Figure 4.10); however we point out that in some wells (wells 2 and 5) we can identify
a relatively thick layer of soft sand at the top of the reservoir, therefore in this facies
a soft sand model has been applied to explain the low velocity values measured at the
well locations in the corresponding intervals. From well data we can also infer the
marginal distribution of petrophysical properties conditioned by the facies classifica-
tion (Figure 4.11): the estimated pdfs are used in the forward model to generate the
simulated realizations at each iteration of the stochastic inversion. We assumed for
simplicity, Gaussian distributions, but SGSim does not require that the prior distri-
bution is Gaussian and other distributions could be used. The descriptive statistics
of the different properties are shown in Table 4.3.
The results of the 1D application with a perfect synthetic seismic trace (SNR=8)
Figure 4.9: Real case application: well log dataset from well 2 (calibration well).
From left to right: P-wave and S-wave velocity, effective porosity, clay content, water
saturation, and actual facies classification (shale in green, silty-sand in brown, stiff
sand in light brown, soft sand in yellow).
are shown in Figure 4.12, where we reconstruct the litho-facies classification at well
2 location. In this test we assumed that the wavelets were known, in other words the
wavelets used in the inversion were the same as those used in the forward modeling
performed to generate the synthetic seismic trace. The actual facies profile is severely
non-stationary and the estimation of variogram parameters was not reliable; we then
assumed the same exponential model for all the facies with correlation range equal
to 2.5m. We notice that even if we start from an initial model with a short vertical
correlation range, we obtain a good result in the inversion: after 50 iterations (right
plot) the error between the input seismic and the synthetic seismic generated from
the optimized model is lower than the fixed tolerance T ′ , and the match between the
optimized facies profile and the actual classification is satisfactory. The tolerance T ′
is fixed such that the ratio between the variance of the signal and the variance of
the residuals approximates the SNR: in this example with perfect synthetic seismic
we stop the convergence when the ratio approximates 10. The upscaled classification
Figure 4.10: Rock physics model: (left) P-wave velocity versus effective porosity;
(right) S-wave velocity versus effective porosity, color coded by clay content. Black
lines represent constant-cement sand model for different clay contents (from top to
bottom: 0%, 25%, 50%, 75%, and 100%).
(see Stright et al. (2009)) over a support of 1m is provided for comparison. As

already pointed out the tolerance value used to stop the inversion process depends
on a number of factors, in particular the quality of seismic data (signal-to-noise and
resolution): the more reliable is the seismic dataset, the smaller can be the tolerance.
In most of the cases, we do not want to perfectly match the seismic data, but we prefer
to obtain a model (or a set of models) that match the data within a fixed tolerance.
In Figure 4.13 we show two different sets of inverted models: on top we show 25
Shale Silty-sand Stiff sand Soft sand

Porosity 17% 24% 28% 30%
Volume of clay 47% 35% 28% 14%
VP (m/s) 2591 2872 2820 2581
VS (m/s) 1112 1292 1313 1198
Density (g/cm3 ) 2.26 2.18 2.13 2.12
Table 4.3: Mean values of petro-elastic properties in the different facies. Values of
porosity and clay content have been estimated from well log data, elastic properties
values have been computed by rock physics model.
Figure 4.11: Marginal probability density functions of effective porosity and clay
content conditioned by facies classification. The pdfs of petrophysical properties are
used to distribute rock properties within the reservoir model at each iteration of
stochastic inversion (shale in green, silty-sand in brown, stiff sand in light brown, soft
sand in yellow).
inverted models obtained by imposing a small tolerance T ′ (on average 51 iterations

are needed to reach the required accuracy); on bottom we show 25 inverted models
obtained with a larger tolerance equal to 1.2T ′ (33 iterations required on average).
By decreasing the tolerance value, we improve the match between the input seismic
and the synthetic seismic of the generated models, but we increase the computational
time to reach the convergence condition. The different variability within the two sets
of realizations is shown by the e-types of the two ensembles (Figure 4.13). The e-type
is the ensemble average of the set of models and it is a continuous variable.
We then performed the same inversion exercise with the real collocated seismic
Figure 4.12: Inversion results at well 2 location with synthetic seismic data, from left
to right: actual facies classification, upscaled facies profile, seismic facies probability,
initial model, optimized model after 50 iterations (shale in green, silty-sand in brown,
stiff sand in light brown, soft sand in yellow).
trace (Figure 4.14). The quality of the inversion is worse than the previous case but
it is still satisfactory as pointed out by the confusion matrix reported in Table 4.4,
where we obtain high recognition indexes for all the facies. In both applications,
with synthetic and real seismic, we used as secondary information a low resolution
probability estimated at the well location. In this application we assumed that the
parameters of the Tau-model are: τ1 = τ2 = 1 and τ3 = 0.5. The use of this probability
is necessary to account for the non-stationarity of the actual facies classification. In
Figure 4.15 we show the convergence of the methodology with and without secondary
information and we show the boxplots of the normalized error of 25 optimization
runs, each of them consisting of 50 iterations. We notice that the convergence is
much slower if no secondary information is used, and the average error is generally
higher.
T shale T silty-sand T stiff sand T soft sand

C shale 0.95 0.05 0 0.00
C silty-sand 0.09 0.81 0.10 0
C stiff sand 0.07 0.08 0.72 0.13
C soft sand 0 0.10 0 0.90
Table 4.4: Confusion matrix of the inversion results obtained by stochastic inversion
at well 2 location (T stands for true facies, C stands for classified facies).
We then applied the methodology to two other wells (wells 3 and 5), by using the
same parameters for variograms, prior distributions of properties and forward models
calibrated at well 2. Well 3 does not present soft sand in the classification, whereas the
scenario in well 5 is similar to the calibration well. In both cases the inversion results
are good (Figure 4.16), even though we notice that we cannot reproduce the thin
layers below a certain thickness. We could actually obtain a set of models with the
same vertical correlation observed at the well by using a shorter correlation range and
a larger tolerance T and/or a smaller value of the Tau-model parameter τ3 (Eq. 4.14)
related to the low resolution information. In fact if we assign a high weight to seismic
data (low tolerance T and/or high parameter τ3 ) we tend to match the seismic data
with higher accuracy and we cannot recover thin layers under the seismic resolution.
If a larger tolerance is fixed or a lower weight is assigned to seismic data, we could
recreate thin layers according to the input variograms; however, in this application
we retain that some of the layers visible in the actual profile are due to clustering
artifacts in the log-facies classification process.
As conclusion of this study we perform the stochastic inversion in terms of facies of
a 2D seismic section passing through wells 2 and 3. The near and far stacks are shown
in Figure 4.17 and the top horizon (in time domain) of the reservoir is superimposed.
The seismically derived probability of facies, P ∗ (Fk |S) has been computed following
the approach proposed in Grana and Della Rossa (2010). The maximum a posteriori
of the so-estimated seismic facies probability converted in depth within the geocellular
grid, is shown, as a reference, in Figure 4.18 (bottom right). The velocity model used
for the time-to-depth conversion has been obtained by applying a kriging method to
the 2D section by using filtered sonic logs (at a frequency of 4 Hz).
The initial model has been generated using SISim: the 2D variograms of the facies
have been estimated using information from previous studies on this field and nearby
fields of the same area, and the prior information assumed by averaging the facies
proportions at the well locations. The 2D variograms describe the spatial continuity
model of the facies. For each facies we assumed an exponential model with the
following parameters: the lateral correlation range is 1000m for soft sand, 2000m for
stiff sand, 2500m for silty-sand and 800m for reservoir shale; the azimuth is 0o for all
the facies. The optimization is performed for all traces simultaneously.
The main result of this study is the optimized facies model (Figure 4.18, top): this
result honors the prior information and the spatial continuity of the data; furthermore,
the optimized realization honors the seismic data. Here we show the results after 10
iterations, corresponding to 110 2D simulations, in terms of facies, porosity, P-wave
velocity, and the corresponding synthetic seismic (Figure 4.18). For simplicity, the
porosity of the non-reservoir shale has been set constant equal to 0.05. As expected
the areas with higher variability are the sequences of silty-sand and stiff sand in the
lower part of the reservoir.
As this application points out, the main advantage of stochastic techniques is that
the estimated facies model has a higher resolution than the model obtained from the
maximum a posteriori of the probability of facies directly inferred from seismic, and it
can be directly used as initial model in fluid flow simulation, without the integration
of additional geostatistical methods.
Figure 4.13: Multiple realizations with different tolerance conditions: (top) 25 sim-
ulations obtained with a small tolerance; (bottom) 25 simulations obtained with a
larger tolerance (shale in green, silty-sand in brown, stiff sand in light brown, soft
sand in yellow). For each set of simulations we plot the ensemble average (e-type)
and we compare the results with the upscaled facies classification at well 2 location.
Figure 4.14: Inversion results at well 2 location with real seismic data, from left to
right: actual facies classification, upscaled facies profile, seismic facies probability,
initial model, optimized model after 50 iterations (shale in green, silty-sand in brown,
stiff sand in light brown, soft sand in yellow).
Figure 4.15: (Top) Convergence plot of stochastic inversion results with and with-
out low frequency information (blue and red symbols respectively) as a function of
iteration number. (Bottom) Boxplots of 25 runs consisting of 50 iterations with and
without low frequency information (left and right respectively).
Figure 4.16: Inversion results at well 3 (left plots) and well 5 (right plots) locations.
We compare the actual classification with the optimized model obtained by stochastic
inversion (shale in green, silty-sand in brown, stiff sand in light brown, soft sand in
yellow).
Figure 4.17: 2D seismic section passing through wells 2 and 3: (top) near angle stack
corresponding to 8o , (bottom) far angle stack corresponding to 26o . The black line
represents the top horizon, in time domain, corresponding to the interpreted top of
the reservoir.
Figure 4.18: Inversion results along the 2D section shown in Figure 4.17. On the
left we show the optimized model of reservoir facies (top left), porosity (mid left) and
P-wave velocity (bottom left), obtained by stochastic inversion. On the right we show
the corresponding synthetic seismic sections, near (top right) and far (top left) and
the maximum a posteriori (MAP) of seismic facies probability (converted in depth
and mapped in the geocellular grid) used as secondary information in the inversion
and obtained by multi-step inversion (shale in green, silty-sand in brown, stiff sand
in light brown, soft sand in yellow).
4.5 Discussion
In this chapter we summarized the main results obtained by applying a new method-
ology for inversion of seismic data in terms of facies. Most of the geostatistical
methodologies published so far, require large computational times to be performed
for real reservoir studies. The introduction of the probability perturbation method
allows us to reduce the computational time compared to other stochastic optimization
methods. The probability perturbation method incorporates the Tau model, which
is a probabilistic model to account for information coming from different sources. By
using the Tau model, we can integrate well and seismic data and we could potentially
extend the methodology to include other information such as EM data or production
data.
In our approach we used two-point geostatistics methods, to obtain models of
facies and petrophysical properties; however the presented stochastic inversion could
be applied also with multi-point based priors, if a suitable training image is available.
Multi-point geostatistics could be necessary in litho-fluid inversion, to avoid non-
physical scenarios such as brine sand on top of oil sand. In our study we assumed
that the oil water contact was known.
We point out that in both cases, it is important to include secondary information
in the methodology, specifically in the Tau model: this information represents the
probability of facies at seismic scale, or in a broad sense, the low resolution prob-
ability of facies. This additional step has two goals: the first one is to speed up
the convergence as shown in our application, the other one is to account for non-
stationarity. Two-point and multi-point geostatistics are based on the assumption of
stationarity, but in most of the cases this assumption is not completely satisfied. In
our real case application for example, silty-sand and stiff sand approximately have
a stationary behavior in the reservoir layer, but this is not true for soft sand, which
appears only at the top of the reservoir. The scenario is more complex if we consider
the shale layers at the top and at the bottom of the reservoir. The layers bordering
the reservoir can be neglected in reservoir modeling if a reliable set of horizons is
provided; however in seismic reservoir characterization the elastic contrasts at the
top and at the bottom of the reservoir have a key role in seismic interpretation and
inversion. The introduction of the probability of seismic facies, even though with a
low resolution, allows us to account for the non-stationarity observable in the vertical
stratigraphy. Moreover, we observed in the proposed real field application, that the
integration of the secondary probability information reduces the number of iterations
by a factor of 5 (assuming τ3 = 0.5). We point out that secondary information could
be integrated in different ways: in addition to the Tau model used in our research,
we could include this additional probability data in sequential simulations by using
cokriging or Bayesian updating; and it could be derived from different sources of
information, such as full waveform inversion, AVO attributes, or EM methods.
Non-stationarity is also visible in petrophysical and elastic properties and it can
be recognized by the presence of trends as a function of depth. To overcome this
issue, we introduced in the real case application a deterministic trend of porosity
and elastic properties to integrate/correct the rock physics model which does not
explicitly account for depth-dependency. The trend observable in porosity could be
due to different mechanisms and also to changes in lithology; for example we suspect
that the layer of shale at the bottom of the reservoir has a different mineralogy than
the overcap clay, even though the amount of clay recognized by means of gamma ray
log is approximately the same. In general the depth trend is not linear (see Rimstad
and Omre (2010)) and several models are available, but in a relatively small depth
interval it can be approximated by a linear regression.
We observe that in the forward modeling we used a convolutional model, which is
a Born single scattering approximation of the first order. However, other techniques
could be used such as the reflectivity methods (Kennett algorithm) or 2-D Born fil-
tering. Similarly different rock physics models could be introduced to link elastic and
petrophysical properties. A limitation of this methodology is the presence of tuneable
parameters, such as the Tau model exponents, variogram parameters, and the opti-
mization tolerance. The determination of the Tau-model weights is still a subject of
research and a methodology to quantitatively express the data-dependency or the de-
pendency of the probabilistic information, through the Tau-model exponents, is still
missing. Especially when redundant information is incorporated in the Tau-model,
it is recommended to perform a preliminary sensitivity analysis on the parameters of

the Tau model, possibly at the well location where a facies classification is generally
provided by formation evaluation studies. In our study we used τ1 = τ2 = 1 and
τ3 = 0.5 after some synthetic tests performed at the well location to investigate the
effect of the Tau model parameters, but the final choice is partially a subjective de-
cision. The optimization tolerance T significantly influences the variability of the set
of models of multiple realizations: the smaller is the tolerance the smaller is the vari-
ability within the set of models obtained from different runs. This choice is related to
how well we want to match the data. This parameter depends on several conditions
related to the seismic dataset of the reservoir study and it should be assessed by trial
and error. A similar conclusion can be drawn for variogram parameters: in this case
an assessment of the parameters is harder, especially for the correlation ranges in the
lateral directions, due to the lack of multiple wells distributed spatially. We have
seen that, in the 1D application, the key is to select the appropriate parameters to
reconcile the expected vertical variogram and the match with real seismic data; but
this assessment cannot be done in a 2D or 3D application unless a very large number
of wells are available. The assumptions related to the choice of correlation ranges and
anisotropic parameters of the variogram model are crucial; in real reservoir studies
this choice can only rely on prior geological information of the field or nearby fields.
We observed that solutions are better, with finer scale features, when the variogram
range is underestimated. As pointed out by an anonymous reviewer, this behavior
may be quite general. By giving a short range we specify a prior information outside
the seismic bandwidth and the inversion process can adapt better the solution to the
specified prior (i.e. it has more freedom to arrange the thin layers in a way that
gives a good seismic response). On the contrary if we specified a long range, the
prior may be less complementary to the seismic frequency content. In the case of a
wrong prior (too large thickness linked to an overestimation of the ranges) it may be
then more difficult for the inversion algorithm to adapt the (thick) layers to give the
assigned seismic response. Seismic data could be used to estimate these properties,
but the lateral continuity of seismic is generally affected by the migration operator
applied to data, which could lead to an overestimation of the lateral ranges of the
variograms. One possible future development is to use a multi-parameter approach in

the probability perturbation method (MP-PPM), where we optimize more than one
parameter simultaneously. One particular property that could be made stochastic
and perturbed in this manner is the facies proportion, especially when few wells are
available in the field.
4.6 Conclusion
We presented a new methodology for facies and reservoir properties modeling which
combines traditional geophysical models, such as rock physics and seismic forward
modeling, with geostatistical methods. The proposed approach is a stochastic op-
timization based on probability perturbation method and Tau model. The use of
sequential simulations allows us to generate fine-scaled models, while the probabil-
ity perturbation method guarantees that the optimized model matches the real data
within a fixed tolerance. The main advantage of this technique is that it provides
high resolution models of facies and the associated properties in a moderate amount
of computational time. The method is fast, especially if secondary information is pro-
vided. This secondary information can be obtained from seismic attributes through
different techniques: if the secondary information is not taken into account, the con-
vergence could be quite slow, especially in complex sequences of thin layers, and
the assumption of stationarity of geostatistical methods cannot be satisfactorily over-
come. Different probabilistic information from different sources and at different scales
can be integrated into the methodology thanks to the use of the Tau model. The ap-
plication to the real well data shows that the methodology can be applied to complex
reservoirs with good results.
Chapter 5
Statistical methods for log

evaluation
5.1 Abstract
Formation evaluation analysis, rock physics models and log-facies classification, are
powerful tools to link the physical properties measured at wells with petrophysical,
elastic and seismic properties. However this link can be affected by several sources of
uncertainty. In this chapter, we propose a complete statistical workflow for obtaining
petrophysical properties at the well location and the corresponding log-facies classi-
fication. This methodology is based on traditional formation evaluation models and
cluster analysis techniques but it introduces a full Monte Carlo approach in order
to account for uncertainty evaluation. The workflow includes rock physics models in
log-facies classification to preserve the link between petrophysical properties, elastic
properties and facies. The use of rock physics model predictions guarantees obtaining
a consistent set of well log data that can be used to calibrate the usual physical mod-
els used in seismic reservoir characterization and to condition reservoir models. The
final output is the set of petrophysical curves with the associated uncertainty, the
profile of the facies probabilities and the entropy, or degree of confusion, related to
the most probable facies profile. The full statistical approach allows us to propagate
the uncertainty from data measured at well location to the estimated petrophysical
113
CHAPTER 5. STATISTICAL METHODS FOR LOG EVALUATION 114
curves and facies profiles. We apply the proposed methodology to two different well
log studies to show its applicability, the advantages of the new integrated approach
and the value of uncertainty analysis.
5.2 Introduction
In reservoir modeling and seismic reservoir characterization most of the physical mod-
els used are calibrated at well location. However well log data do not provide direct
measurements of the reservoir properties we account for in reservoir models: the de-
sired reservoir properties are generally obtained from the measured well logs through
formation evaluation analysis. This process contains several sources of uncertainty.
Similarly log facies classification can be severely affected by the uncertainty of petro-
physical curves performed in formation evaluation. Furthermore, in some cases, the
classification could be not linked to seismically derived attributes if rock physics mod-
eling results are not included in the classification methodology.
The first goal of this work is to present a new methodology for log-facies classi-
fication based on both petrophysical and acoustic/elastic properties in order to link
log-facies to seismic inverted attributes. The second added value of the workflow is
the introduction of Monte Carlo simulations to generate several realizations of petro-
physical and elastic curves to obtain different log-facies profiles which are used to
infer facies probabilities and facies uncertainty.
The first step of the methodology is formation evaluation analysis. The petrophys-
ical evaluation of subsurface formations requires the combined efforts of log measure-
ments and core data together with a quantitative log interpretation model. The main
results obtained from the petrophysical interpretation of well logs are the volumes of
certain formation components (solid matrix and fluids) at each data level which com-
bine the measurements provided by several tools such as well log resistivity, acoustic,
density, neutron, nuclear magnetic resonance, fluid sampling, coring, and imaging (for
details we refer the reader to Darling (2005), and Ellis et al. (2007)). The standard
approach consists of optimizing simultaneously equations described by one or more
interpretation models. Formation evaluation is done by solving the so-called inverse
problem, in which well log measurements and response parameters are used together
to compute volumetric results. An improved approach has been recently introduced
by Heidari et al. (2010). However, there is always some uncertainty in the inferred
parameters obtained by the inversion process since errors in any of the measurements
lead to errors in the final values (see Fylling (2002), Verga et al. (2002), Kennedy
et al. (2010), and Viberti (2010)). Moreover, other sources of uncertainties in the
determination of the volume fractions come from the theoretical models used in the
interpretation and the input parameters characterizing them. Hence a statistical
quantitative log interpretation is performed in order to obtain reliable results and to
assess the uncertainty for the characterization.
Rock physics modeling represents the second step. Rock physics is mainly used to
establish the link between petrophysical parameters and acoustic/elastic properties.
In our approach, the uncertainty from formation evaluation is propagated through
the well-calibrated rock physics model in order to provide a full probabilistic petro-
elastic model. This chapter only covers the application of the basic models used in
rock physics and the statistical approach. For those who are interested in in-depth
knowledge of rock physics, we refer to Bourbie et al. (1988), Nur and Wang (1989),
Wang and Nur (1992), Wang and Nur (2000), Avseth et al. (2005) and Mavko et al.
(2009). Statistical rock physics has been introduced by Mukerji et al. (2001a) and it
is essentially based on Monte Carlo simulations. An extensive explanation is included
in Doyen (2007).
Finally a log-facies classification is performed at each depth location of the well.
Reservoir facies are usually defined from sedimentological information and core anal-
ysis. This classification must then be consistent with the one performed on well
log data in order to link petrophysical properties with reservoir facies. Tradition-
ally, log-facies classification is based on multivariate techniques (in particular cluster
analysis, see Kaufman and Rousseeuw (1990)) introduced to automatically identify
common features within well log data and computed curves (generally petrophysical
properties). In this work, we also account for elastic properties in order to guarantee
a clear discriminability in the petro-elastic domain, through rock physics modeling.
This issue is of key importance for the following step of seismic facies classification
performed on seismic data to obtain reliable reservoir models. In this chapter we do

not focus on the classification of the seismic data but on the initial well log analysis
and log-facies classification.
The probabilistic petro-elastic model yields a statistical log-facies classification.
Thus the probability of occurrence of different facies at the well is obtained. The
concept of entropy (Shannon (1948)) is then exploited in order to quantify the uncer-
tainty related to the discrete random variable describing the facies distribution along
the well.
At each step of the workflow, namely formation evaluation analysis, rock physics
modeling, and log-facies classification, we account for the sources of uncertainty as-
sociated with measured data, numerical models and natural heterogeneity in order to
provide for each estimated curve, the corresponding confidence interval at each depth
location along the well.
The crucial added value is the assessment and propagation of the uncertainty at
each step in order to finally provide a reliable classification fully based on the amount
of information characterizing the entire workflow. In general, formation evaluation,
rock physics models and log-facies analysis are commonly implemented neglecting the
uncertainty involved in these processes and this fact can lead to an erroneous final
interpretation. This is exactly the pitfall the probabilistic petro-elastic classification
tries to avoid by quantifying the reliability of each step and of the final result. More-
over, the discriminated facies are automatically characterized from the petrophysical
and acoustic/elastic point of view, which is a key factor in data integration in seismic
reservoir characterization. Eidsvik et al. (2004b) presented a probabilistic log-facies
formulation using hidden Markov models to get the posterior probability distributions
of the log facies accounting for uncertainties in the log data and rock physics model,
but did not account for the uncertainties in the formation evaluation analysis.
Differently from previously published works in formation evaluation analysis and
log-facies classification, the uncertainty propagation problem is faced by using a full
Monte Carlo approach for each step of the method. Monte Carlo methods allow
us to numerically estimate the posterior probability density functions (PDFs) of the
variables we are interested in: in formation evaluation we estimate the posterior
PDFs by propagating, point by point, the uncertainty from acquired well logs; in
rock physics modeling we estimate the posterior PDFs by propagating the uncertainty
from the previously obtained petrophysical curves; finally both PDFs are combined in
log-facies classification where we sample from these distributions and we combine the
results to statistically classify the facies, point by point, and estimate the associated
uncertainty. The so obtained uncertainty represents the uncertainty associated to
the degree of approximation of the physical models used in the method and the
uncertainty related to error measurements. However random errors are also taken
into account to represent the natural variability and heterogeneity of the rocks.
The whole methodology is first illustrated by applying all the steps to the same
well log dataset of an offshore West Africa field. We then perform the same study on
another real case, located in the North Sea, where a complex sedimentological facies
classification has been identified.
5.3 Methodology
The methodology can be divided into three parts:
1. Quantitative Log Interpretation (QLI);
2. Rock Physics Model (RPM) computation;
3. Log-Facies Classification (LFC).
The flowchart of the methodology is shown in Figure 5.1. In this specific case it
refers to a clastic environment, but it can be applied to different scenarios if suitable
log interpretation and rock physics models are available.
In the following we analyze in detail these three steps and the uncertainty propa-
gation through them.
5.3.1 Quantitative Log Interpretation

The petrophysical characterization of subsurface formations is an inverse problem and
generally requires an integrated approach. Well log measurements and core data are
Figure 5.1: Flowchart of the probabilistic petrophysical properties estimation.
combined together with a Quantitative Log Interpretation (QLI) model. QLI, also
known as formation evaluation, provides the volume fractions of interest (solid matrix
and fluids) at each depth level by solving the corresponding inverse problem, in which
well log measurements and response parameters are reconciled with the theoretical
interpretation models. In particular, the inverse problem is based on an opportune
cost function that expresses the distance between the observed measurements and
the predictions of the model chosen to describe the system. The final aim is to
minimize the cost function and determine the solution (i.e. the volume fractions of
the formation components) that optimally mimics the observables.
Theoretical response-equations in QLI come from a simplified description of the
physics involved in the well log instruments and are considered together with several
constraints. These are typically used in order to impose geological and/or petrophys-
ical prior knowledge to avoid physically meaningless results and define the domain
space of the problem. In details, QLI can be divided into three different phases:
• system parameterization, defining the set of parameters (unknown volumes)

which characterizes the formation;
• direct modeling which involves the laws (linear and/or non-linear) able to gen-
erate the synthetic values for the observables, once the parameters describing
the model are fixed;
• inverse modeling, playing its role in volumetric quantification.
System parametrization and direct modeling are quite straightforward; even so, it
is worth mentioning a few points of discussion that need to be taken into account to
properly define the domain space.
Mathematically, at a given depth x , the vector d(x) represents the observational
log data (e.g. gamma ray, neutron, density and so on). QLI model is determined by
the components of a vector p(x) containing the depth-dependent unknown volumetric
fractions, such as porosity φ, volume of clay vclay , water saturation sw and so on. For
simplification purposes, from now on the dependency upon the spatial position x is
disregarded and it is assumed that all the recorded log measurements are sampled
with a constant depth step. Formally we can write:
p = [φ, vclay , sw , ...] (5.1)
Once a well is drilled into the formation, to maintain the pressure balance, some
mud fluid is pumped into the borehole. Often, drilling is done at over-balanced
condition and mud has a pressure slightly higher than the formation pore pressure.
This pressure gradient induces mud filtrate to seep into the porous rock system.
Hence, well log measurements are affected by, at least, the existing step invasion
profile due to the circulating fluid. The mud invasion divides the formation in two
separated parts: the so-called flushed zone (usually labeled XO) and the unflushed
one, or deep zone (DE). This splits the set of unknowns into two subsets, and the
vector p can be written as:
p = (pXO , pDE ). (5.2)
Shallow reading log measurements are assumed to respond only to volumes of

formation components in the first zone (pXO ). Similarly, deep reading logs only see
the unflushed zone pDE ). Some tools, which have a medium depth of investigation,
are assumed to be influenced by both zones, and their response equations contain
terms for all formation components, regardless of the zone.
On the other hand, physically-based assumptions constrain and reduce the domain
space of the unknowns. For instance, the standard normalized-to-unity investigated
total volume fractions (both in shallow and deep zone) need to be achieved. More-
over, it is common to assume lateral continuity: all the solid formation components
extend infinitely from the borehole at zero degrees dip and porosity in the flushed
and undisturbed zone is the same, regardless of the type of fluids filling the pore
space. So, the sum of the fluid volumes in the flushed zone is equal to the sum of
fluid volumes in the unflushed zone. In a more compact form, all of the constraints
can be collected in the function . Synthetic log data are obtained by applying the
direct modeling operator for a given fixed values of the model parameters u(p):
Synthetic log data s are obtained by applying the direct modeling operator fQLI (p)
for a given fixed values of the model parameters p :
s = fQLI (p). (5.3)
The forward model is a set of equations, linear or non-linear, which describes the
link between volumetric curves (p) and petrophysical properties measurable at the
well (d).
Inverse modeling tries to determine the components of the parameter vector p,
using the observational data d. The optimization problem involves a cost function
to be minimized in order to reduce the discrepancy e between the observed data and
the synthetic data generated by the direct modeling:
e = s − d = fQLI (p) − d. (5.4)
The definition of the cost function introduces the concept of a weighted 2-norm for
the vector e. As a matter of fact, since the physical observables and measurements
come from several well log tools and involve absolute values which are in general
very different (and have different units), we have to normalize the 2-norm in order
to treat all the quantities at the same level. This normalization mainly depends on
typical tool accuracy. Moreover an additional term, a multiplier, is incorporated in

order to weight the different tools with respect to their influence in the final answer
and it is based on log analyst experience. The understanding of how a tool can
affect the results and the knowledge of tool physics is required to select the multiplier
values. Finally, normalizations and weight multiplier factors can be organized into a
weighting matrix W. So, the cost function reads

1 T
j(p) = e We . (5.5)
2 u(p)
which is a scalar and it can be represented as a function of the unknowns collected

in the vector p. The aim of optimization is to determine the optimal solution p∗ that
minimizes the above cost function, under the set of constraints u(p) representing the
solution space for the problem. The so-obtained vector p∗ represents the output and
solution of the standard QLI.
We now set the scenario regarding the uncertainty analysis and evaluation for the
QLI methodology. As a matter of fact, there is always some degree of uncertainty in
the inferred volumes p∗ obtained by the inversion process since each piece of infor-
mation introduced in system parametrization and direct modeling can be affected by
several errors. Quantification and propagation of these uncertainties throughout the
inversion problem is here considered and assessed together with appropriate methods.
The main sources of uncertainty in QLI come from well log measurements, het-
erogeneity of rock system, possible thin layered intervals, invasion effects, lithological
and/or textural assumptions, environmental correction issues, simplified interpreta-
tion models and input parameters. A complete analysis of the above uncertainties
would involve a full knowledge of the individual contributions and such detailed pre-
liminary evaluation is rarely, if ever, available. Thus, the practical quantification prob-
lem simplifies the classification of sources into three macro-groups: methodological,
systematic and random errors (see Theys (1991), Theys (1994), and Theys (1997)).
Methodological errors are mainly due to an incorrect choice of interpretation mod-
els and related parameters. In general, model selection is one of the most critical steps
in a reliable QLI and, for sure, it is mandatory to keep in mind all of the simplify-
ing hypotheses of the real system it intends to describe. Since an inaccurate model
selection could introduce significant errors to the interpretation, at this particular

stage core measurement calibration and validation are fundamental. Petrophysical
experience plays a role as well. Potential errors associated to an incorrect choice of
the theoretical models are difficult to quantify and their magnitude may vary signif-
icantly. This is to state that methodological errors, once assessed, should be mainly
used as a sort of guidelines on the applicability limits and constraints underlying each
model.
Systematic errors are defined as reproducible inaccuracies of the measurement
mainly due to imprecision of the instrumental system, processing of the data, envi-
ronmental conditions and so on. Systematic errors cannot be removed by repeated
runs, thus they need to be recognized and corrected before any calculation.
Random errors are those that need to be accounted for and propagated through-
out the QLI process. In fact, random errors are mostly associated to the physics
of the well log measurement system and cannot be corrected because they cannot
be reproduced. Statistical variations in count rates or signal noise are examples of
random errors. Also the uncertainty associated to the parameters used to correct
the measurements for any environmental effect has an important impact in the final
petrophysical characterization and must be accounted for.
Given the broad spectrum of random uncertainty, a statistical treatment of the
problem is suitable. Despite this apparent need, most formation evaluations in prac-
tice are based on a single deterministic description that can reproduce with a certain
quality and confidence the well observations. The single deterministic model and its
description make the adopted model rather unsuitable for uncertainty assessment.
On the other hand, the statistical approach combines the deterministic QLI with
a Monte Carlo Simulation (MCS) in order to build a probabilistic framework to study
the natural variability of the results (Figure 5.2). In order to avoid confusion, in what
follows deterministic QLI means with no MCS implemented.
We start from a number of uncertain inputs representing log measurements. Un-
certainty in each input variable is represented by a probability density function (PDF)
Figure 5.2: Schematic representation of petro-elastic uncertainty estimation and log-

facies classification through Monte Carlo simulations of petrophysical and elastic
properties.
whose range of values correspond to the nominal uncertainty provided with the mea-
surement. In particular, Gaussian distributions are used to represent the random vari-
able d̃ ∼ N(d, Σd ) and describe the uncertainty of acquired data: the measured value
d represents the mean, the standard deviation is equal to the instrumental/processing
error (collected in the covariance matrix Σd ), and N is a compact notation for the
normal distribution.
This set of distributions, estimated at each depth location along the well profile,
provides the bandwidth of tolerance for the response variables of the forward model.
Model uncertainties can be added to the forward modeling fQLI by means of a
random noise εQLI , distributed with zero mean and a prior covariance matrix, so that
the forward model becomes:
s = fQLI (p) + εQLI . (5.6)

In most of the cases the forward model can be represented as:
s = [gr, nt, dn, rs] = fQLI (φ, vmin , sw ) + εQLI . (5.7)
where φ is porosity, vmin is the vector of the volumes of mineral fractions, sw

is water saturation, gr is Gamma ray, nt is neutron porosity, dn is the acquired
density and rs is resistivity, and it allows computing the volumetric fractions (porosity,
volume fractions of mineral, and saturations) from the acquired logs. In our approach
we include also velocity data to guarantee the consistency with rock physics modeling
step that follows.
The simulation is performed by randomly sampling from the PDFs of d̃ at each
data level the values for each input log required for the interpretation model (see
also Viberti (2010)). Suitable vertical correlated and conditioned Monte Carlo sam-
pling can be implemented in a straightforward way in order to perform a geological
driven sampling and minimize the probability of occurrence of physically meaningless
scenarios.
Once a set of values has been sampled, the QLI cost function is optimized and
the results are stored. When a statistically representative number of realizations have
been drawn the results can be sorted and histograms created to approximate the local
PDFs of the uncertain output p̃
p̃ = F (d̃) (5.8)
where F = (fQLI + εQLI )−1 .
Thus, the single vector p∗ obtained in the deterministic QLI is now enhanced to
account for the uncertainty of p̃.
Once this goal is achieved, all inferences can be obtained from the posterior PDFs
by computing statistics relative to individual parameters. Since the final result is
presented as probability density functions, this tool provides a unified framework for
volume estimates and for the uncertainty associated with them.
Formation evaluation analysis equations
We recap here the basic equations for formation evaluation analysis (Darling (2005),
and Ellis et al. (2007)). Starting from well log measurements, formation evaluation
computes volumetric curves of the corresponding formation components. Assuming

a unitary investigated total volume then the following relation holds:
N
X
vh = 1 (5.9)
h=1
where 0 < vh < 1 is the volumetric fraction of the single hth component and N is the
total number of minerals and fluids defining the modelled formation.
Fundamental in dealing with response equations is the knowledge of how clay is
handled and the concept of wet versus dry clay. Here, we take the clay as wet so
the volume fraction, vclay , is intended to consider this phase composed of dry clay
and associated bound water. As already mentioned, well log measurements could be
affected by the existing step invasion profile due to the circulating mud fluid. The mud
invasion divides the formation in two separated parts: 1) the flushed zones (XO) and
2) the unflushed deep zone (DE). Well logs tools have different depths of investigation
and can be influenced only by one zone (shallow reading log tools and deep reading
logs) or by both zones; then their response can contain terms for only one or for all
formation components. Mathematically this is taken into account by a special factor,
0 < Ψi < 1, called the invasion factor, which controls how much influence for each
log i (such as neutron porosity or density) comes from the XO zone. The remaining
influence (1 − Ψi ), comes from the DE zone. In well log interpretation processes it is
common to assume a lateral homogeneity and as a consequence to assume that the
sum of the fluid volumes in the flushed zone is equal to the sum of fluid volumes in
the unflushed zone:
Nf luid
X
(vXOk − vDE k ) = 0 (5.10)
k=1
where the sum only considers the fluid involved.

Other constraints involve the feature of the mud used. For example, when oil
based mud is used the invading fluid is hydrocarbon. Therefore, we must impose a
constraint that limits the volume of water in the flushed zone to be less than or equal
to the volume of water in the undisturbed zone on the basis of the normal response
equations. The reverse is true for water based mud.
Finally it is also possible to limit the sum of the fluid volumes to be less than a
particular critical porosity. This critical porosity is the value which marks the limit
between a saturated rock and a suspension and it will be relevant also for the rock
physics model.
In general, a response equation is a mathematical description of how a given
measurement mi (where i is the generic index to identify a well log type) varies with
respect to each formation components. The simplest linear response equations are of
the form
NX
solid Nf luid Nf luid
X X
mi (v) = m̃j vj + (1 − Ψi ) m̃XOk vXOk + Ψi m̃DE k vDE k (5.11)
j=1 k=1 k=1
where m̃j , m̃XOk , m̃DE k are the response parameters for the corresponding formation
components. This form is typical for neutron, density, and sonic logs. As gamma
ray log only considers solid components, then it is not affected by invasion and the
response equation does not involve the invasion factor.
On the other hand, the resistivity/conductivity model used is non-linear and in
our formation we consider two different log measurements: one related to the flushed
zone (shallow reading log) and the other in the undisturbed zone (deep reading).
The response equations are the same for both situations, but the parameters involved
depends on the investigated zone.
For conductivity (C), we consider the theoretical form of the Indonesian equation
(Poupon and Leveaux (1971)) since it is widely used in shaly sand formations:
√ n2
√

p (1− 12 vclay ) Cw 12 (m+ mφ2 ) vw
C= Cclay vclay + √ φ (5.12)
a φ
where Cclay and Cw are the specific conductivities of clay and water respectively, vw
is the volume of water and the parameters m, m2 , n and a come from empirical
observations.
5.3.2 Rock physics modeling

A Rock Physics Model (RPM) is represented by a set of equations that transform
petrophysical variables to acoustic/elastic variables. RPM can be a simple regression
on well data or a more complex physical model characterized by a number of parame-

ters that need to be estimated, such as elastic moduli of matrix and fluid components,
critical porosity, aspect ratio, and/or coordination number (Mavko et al. (2009)).
Generally, the standard workflow starts from the model calibration from well log
data, or the estimation of the physical parameters involved, in order to obtain a
good match between RPM predictions and well log measurements. The model type
depends on the geological environment and the parameter estimation is given by the
solution of an inverse problem. In the traditional rock physics workflow, the generic
RPM can be written in the following concise form:
r = fRP M (p). (5.13)
where the vector p represents the petrophysical input obtained from QLI and
fRP M comprises the set of equations defining the model. The output vector r is
the vector containing acoustic and elastic properties, typically P-wave velocity (VP ),
S-wave velocity (VS ) and density (ρ):
r = (VP , VS , ρ) (5.14)
In literature, there are many different RPMs and they can be classified in three
big classes (see Avseth et al. (2005), and Mavko et al. (2009)):
• empirical models: in this case RPM usually is a simple regression using elastic
moduli or directly velocities;
• granular media models: these are based on Hertz-Mindlin contact theory which
assumes that the rock is represented by a random pack of spherical grains;
• inclusions models: these models describe the rocks as a sequence of inclusions

(typically ellipsoidal) until the desired pore fraction is achieved.
The various models differ for the calculation of the dry rock properties. In fact,
the usual scheme of a RPM starts from the computation of fluid and solid phase
(matrix) properties. Then, dry rock properties, i.e. the properties of the solid phase
with its own porosity, are calculated with equations depending on the RPM chosen.
Finally, the effect of fluids using Gassmann equation is included.
In general, once the RPM has been calibrated to well log data, to evaluate the
fluid effect, different saturation scenarios could be considered: a ”brine” scenario
where the hydrocarbon column in the reservoir is substituted by water and a ”full
hydrocarbon” scenario where hydrocarbon saturation is constantly increased up to
one minus irreducible water saturation.
As described, our aim is to assess the uncertainty in RPM by propagating it
from QLI probabilistic output. From the previous QLI analysis we have a number of
realizations of the petrophysical variables representing basic volumetric properties of
the formation: uncertainty is represented by the obtained PDFs of p̃. Thus, fRP M
links the uncertain QLI inputs to multiple RPM output variables. In the end, the
output PDFs of elastic properties r̃ are generated through a MCS (Figure 5.2, top),
extending the deterministic RPM to the probabilistic case:
r̃ = G(p̃) (5.15)
where G = (fRP M + εRP M )−1 .

This means that the random errors of the first step (QLI) are propagated through
RPM. In order to account for the uncertainty associated with the RPM approxima-
tion, a random noise εRP M can be added to the fRP M + direct modeling operator.
MCS still provides a simple way to propagate all uncertainties in the RPMs since the
latter are in general non-linear (except for linear empirical models).
As we previously mentioned, we will restrict our description to a clastic reser-
voir, without losing generality. For high-porosity clastic reservoirs, common rock
physics models are the so-called granular media models (see Mavko et al. (2009)).
These models are based on Hertz-Mindlin contact theory (see Dvorkin et al. (1994),
Dvorkin and Nur (1996), and Gal et al. (1998)) which describes a rock as a random
pack of spherical grains. In the reservoir layer, in addition to input petrophysical
curves, i.e. effective porosity, volumes of mineral components, and saturations, the
applied rock physics model requires the following input data: fluid and matrix prop-
erties, reservoir pressure and temperature. Fluid parameters and reservoir condition
information are generally available from PVT analysis and well tests measurements,
whereas solid phase properties are inferred from core mineralogical analysis. In reser-
voir characterized by low porosity rocks, inclusion models (e.g. see Avseth et al.
(2005), and Mavko et al. (2009)) are more appropriate.
Soft sand model
Granular media models are based on Hertz Mindlin contact theory (Mavko et al.
(2009)) and, they provide estimates for the bulk (KHM ) and shear moduli (µHM ) of
the dry rock assuming that the frame is a dense random pack of identical spherical
grains subject to an effective pressure P with a given critical porosity φ0 and an
average number of contacts per grain n. Critical porosity is the value which marks
the porosity limit between a saturated rock and a suspension. We recap here the set
of equations of soft sand model used in the application case. First Hertz Mindlin
equations are given by:
31
P [n(1 − φ0 )µmat ]2

KHM = (5.16)
18[π(1 − ν)]2
31
3P [n(1 − φ0 )µmat ]2

5 − 4ν
µHM = (5.17)
5(2 − ν) 2[π(1 − ν)]2
where ν is the grain Poisson’s ratio, namely
3Kmat − 2µmat
ν= (5.18)
2(3Kmat + µmat )
and Kmat and µmat are the solid phase (zero porosity) elastic moduli.
For porosity values ranging between zero and the critical porosity, the soft sand
rock physics model connects the grain elastic moduli Kmat and µmat with the elastic
moduli KHM and µHM of the dry rock at critical porosity. This is done by inter-
polating these two end members in the intermediate porosity values by means of a
heuristic modified Hashin-Shtrikman lower bound:
−1
φ/φ0 1 − φ/φ0
Kdry = + − 4/3µHM (5.19)
KHM + 4/3µHM Kmat + 4/3µHM
−1
φ/φ0 1 − φ/φ0
µdry = + − 1/6ξµHM (5.20)
µHM + 1/6ξµHM µmat 1/6ξµHM
9KHM + 8µHM
ξ= (5.21)
KHM + 2µHM
On the other hand, the stiff-sand model connects the two end-members by with
the modified Hashin-Shtrikman upper bound (Mavko et al. (2009)). Gassmann’s
equation is then used for calculating the effect of fluid:
2
Kdry
1 − Kmat
Ksat = Kdry + φ K
(5.22)
Kf l
+ K1−φ
mat
− K 2dry
mat
µsat = µdry (5.23)
where Kf l the fluid bulk modulus, and Ksat and µsat are the rock saturated moduli.
Fluid properties can be obtained by means of Batzle-Wang formulas (Batzle and
Wang (1962)), a set of empirical equations which allows calculating bulk moduli and
densities of fluid components starting from fluid analysis parameters (such as gas
gravity, oil gravity, gas-oil-ratio, temperature and pressure). Finally, velocities are
obtained as follows s
Ksat + 4/3µsat
VP = (5.24)
ρ
µsat
r
VS = (5.25)
ρ
where ρ is the fluid saturated bulk density.
5.3.3 Log-facies classification

The goal of this section is to present a new methodology for Log-Facies Classifica-
tion (LFC) and the related uncertainty evaluation. The classification of facies at well
location is a key important step in reservoir modeling: the static reservoir model es-
sentially consists of stochastic simulations which are driven by log-facies classification
at the well locations.
We propose here a classification based on both petrophysical and acoustic/elastic
properties in order to link log-facies to seismic inverted attributes which are often
used as soft conditioning data in reservoir simulation. However, different sources

of uncertainty affect high resolution log-facies classification. We have classified the
associated uncertainty in two main groups: the uncertainty related to petrophysical
curves obtained in QLI and the uncertainty associated to elastic properties (veloci-
ties or impedances) recovered by RPM. We finally couple the so obtained results by
Monte Carlo simulations to generate several realizations of log-facies profiles which are
used to infer facies uncertainty from the probabilistic analysis previously performed
(Figure 5.2, bottom).
Here, we first present the standard approach in order to set the basis of the statis-
tical methodology. LFC usually combines a multivariate statistics technique (cluster
analysis) and interpretation of prior sedimentological information. Cluster analysis is
a well-known technique that helps to group objects according to their mutual simi-
larity. Petrophysical and elastic logs can be statistically processed to find clusters, by
using, for example,an unsupervised hierarchical clustering algorithm. This algorithm
groups the log data based on their statistical similarity into hierarchically ordered set
of clusters. During the characterization phase, an initial number of clusters is selected.
The identified clusters are compared to quantitative and qualitative information de-
rived from cores (routine and special core analyses, lithological, sedimentological and
petrographical descriptions). The integration of information from different sources
(logs, core measurements, lithology, sedimentology, etc) is a key step in the charac-
terization of the different clusters. In particular, a suitable training set t is selected
from QLI and RPM outputs (p∗ and r respectively) and LFC is run including the
data all along the well, providing a depth-dependent, discrete, class vector, collecting
the results of the classification. In detail, the log-facies vector gives the classified
litho-classes at a given depth value. Eventually, some of the clusters are grouped to
identify and classify log-facies for use in the 3D geological model. In the conventional
workflow, LFC does not consider the uncertainty associated to input curves (both
petrophysical and acoustic/elastic ones).
In order to evaluate the uncertainty related to LFC, we use the Monte Carlo
approach. From the previous steps of the methodology (QLI and RPM), we have
obtained PDFs of petrophysical and acoustic/elastic curves (p̃ and r̃) and these define
the probabilistic training set t̃. In particular, from this set of PDFs collected in t̃, N
realizations can be sampled and then used to perform N LFC profiles through MCSs,
and estimate the posterior probability on facies classification, in addition to the most
probable facies profile: in other words, cluster analysis is applied to the different QLI
and RPM realizations to obtain the related scenarios of log-facies. Then, from the so-
generated profiles we can count, at each depth location, the frequency of occurrence
of each log-facies, and infer a depth-by-depth posterior probability distribution.
The results are collected in a vector c whose entries are the probability of each
classified log-facies, so that if have a set of M possible discriminated facies, their
probabilities of occurrence are c = (c1 , ..., cM ). Sedimentological information can be
further integrated by multiplying the probability vector by the confusion matrix ob-
tained from the sedimentological classification (e.g. the matrix accounting for the
degree of confusion among the different facies with those provided by sedimentolo-
gists). We finally observe that, from vector c, for instance, the most likely facies
scenario can be estimated.
We can take a further step in order to quantify the uncertainties associated to
the probabilistic LFC result. In fact, the most probable scenario, as the output of a
probabilistic workflow, should be coupled with a confidence value showing the amount
of information brought by the discrete classification. For this reason the concept of
entropy (Shannon (1948)) can be exploited. This function is a statistical parameter
able to provide some numerical insights on the intrinsic variability of the discrete
variable studied with respect to all the outcomes of the classification.
The scalar information entropy (see Shannon (1948), or Mavko and Mukerji
(1998), and Mukerji et al. (2001b)) associated to the probability vector c is defined
as follows:
M
X
T
h = −c logM (c) = − ci logM (ci ), (5.26)
i=1
where the logarithm is computed in base M (number of facies) since we have classified
M different facies; this restricts and normalizes the entropy to the range [0, 1].
As a simple example, assume that M = 3 and at a particular point of the depth
profile, the three facies are equally probable: c = ( 31 , 13 , 31 ). Hence, the entropy is
h = 1, namely the maximum entropy value and this means that uncertainty is at its
maximum. On the other hand, if c = (1, 0, 0) we get h = 0, giving a minimum entropy
scenario and no uncertainty. In-between situations are characterized by entropy values
that quantify the degree of disorder brought by the classification: 0 < h < 1. Entropy
measure of uncertainty goes beyond measures such as variance and covariance, and
can be used for categorical variables (e.g. facies).
The information provided by entropy analysis on LFC, comprising uncertainty
from QLI and RPM, will be discussed in two real examples.
5.4 Application: first example

The methodology has been applied to a real dataset from an exploration well (here
called A), offshore West Africa. Well A is vertical and drilled in a deep offshore clastic
reservoir made by soft sandstones of a turbidite channel complex. The prospect is
a typical combined structural-stratigraphic trap. It is composed by a quite large
turbidite sandstone channel which drapes over the north flank of a large salt diapir.
Sparse shale beds are present and analysis of core samples revealed that the clay is
a mixture of illite and kaolinite (respectively 70% and 30%). The reservoir fluid is
oil. A comprehensive set of wireline logs (standard and high resolution) has been
acquired to provide a reliable formation evaluation: gamma ray (gr), resistivity (rs),
sonic (sn), density (dn), neutron (nt) and nuclear magnetic resonance (nmr) logs.
Globally, the quality of the acquired data is very good (Figure 5.3). Following our
notation the depth-dependent data vector is:
d = (gr, rs, sn, dn, nt). (5.27)
Deterministic QLI via a straightforward fQLI resulted in the generation of miner-

alogical volumes (vsand , vsilt , and vclay ), effective porosity φ, and water saturation sw
:
p = (vsand , vsilt , vclay , φ, sw ). (5.28)
In well A, invasion effects were negligible so that pXO = pDE . In Figure 5.4, the
results of our local minimization, p∗ , in other words the final set of petrophysical
Figure 5.3: Well log dataset interval of well A, from left to right: P-wave and S-wave
velocity; density; neutron porosity; gamma ray.
curves performed in Quantitative Log interpretation, are shown. In Figure 5.5, we

give an example of comparison between our result and a standard commercial software
output, which shows a satisfactory match between the two estimated set of data.
A rock physics model is then used to establish the link between petrophysical
parameters p∗ and acoustic properties r. For this reservoir study, we used the soft
sand (uncemented) model, based on Hertz-Mindlin contact theory (Dvorkin and Nur
(1996)).
Petro-elastic property uncertainty assessment and facies classification has been
performed in the lower reservoir whose top is located at approximately 2510m. In
the reservoir layer the applied RPM requires the following input data from QLI:
effective porosity, volumes of mineral components (sand, silt and clay) and fluid sat-
urations. The parameters include effective pressure of 35 MPa, a critical porosity of
0.4 while the coordination number is 9. In this particular case effective pressure has
been assumed constant within the reservoir but in general pressure effects on veloc-
ities should be taken into account. Several models have been developed to account
for pressure effect on velocities (e.g. Eberhart-Phillips et al. (1989)) or on dry rock
Figure 5.4: Petrophysical curves performed in Quantitative Log Interpretation, from

left to right: porosity (total porosity in blue, effective porosity in red); volumetric
fraction curves (clay in green, quartz in yellow, and silt in brown); water saturation;
and finally the cumulative volumetric display (shale in orange, silt in brown, quartz
in yellow, water in blue, and oil in green; red dashed line represents 1 minus porosity
and it separates solid and fluid phase).
elastic moduli (MacBeth (2004)); however we point out that these methods must be
calibrated (typically using lab measurements at different pressure regimes) to deter-
mine the empirical parameters. The rock physics model is calibrated at well location,
by comparing the velocity and density estimated from the elastic parameters with the
velocity and density from the recorded logs in the borehole. The calibration is per-
formed in wet condition (meaning that a preliminary fluid substitution is performed
on well logs) in order to avoid the effect of fluid in rock parameters calibration. Fo-
cusing on the depth interval including the lower reservoir layer (2480m - 2600m), the
results can be seen in Figure 5.6 . The calibration is performed with a trial and error
method; optimization techniques could be used as well but these methods do not
guarantee that the optimized parameters still preserve their physical meaning. Rock
Figure 5.5: Petrophysical outputs comparison (red curves represent our results; blue
curves represent standard commercial software outputs).
physics model calibration, in many practical applications, can be quite complex as the
physical-mathematical model cannot account for heterogeneity and natural variability
of the rock. In our application the model overpredicts sonic log velocities at 2520m
depth and does not match the high peak of P-wave velocity at 2550m depth. The
mismatch at 2520m is due to lack of accuracy of the model, while the velocity peak
at 2550m could be a measurement error or represent a different lithology. The lack of
accuracy of the rock physics model in some small depth intervals should be partially
compensated by the variability introduced in the following Monte Carlo simulations.
Traditional log-facies characterization is achieved through cluster analysis, carried
out using key-curves (defining the training set t) selected to be effective porosity, clay
content and VP /VS ratio:
VP
t= φ, vclay , . (5.29)
VS
In particular, VP /VS ratio refers to brine condition to avoid any fluid effect due to
hydrocarbon presence. In this case, a preliminary analysis showed that VP /VS ratio
Figure 5.6: Calibration of the rock physics model, from left to right P-wave and S-
wave velocity, and VP /VS ratio (black curves represent the actual sonic log, blue dashed
curves represent the predicted rock physics model). The rock physics model has been
calibrated in wet condition and then applied to the well scenario by Gassmann fluid
substitution.
and clay content have a good correlation; however the training set could include only
VP , or VS (or both of them), if these logs are identified as good lithological indicators
for the reservoir (see second example).
The training set t was statistically processed by means of a hierarchical agglom-
erative clustering algorithm with Ward’s minimum variance linkage method (Ward
(1963)). Hierarchical algorithms find successive clusters using previously established
clusters (i.e. the resultant classification has an increasing number of nested classes).
Thus, given a data set consisting of N0 objects, agglomerative clustering methods
generate outputs where objects are gradually partitioned. Ward’s minimum variance
linkage method makes use of squared Euclidean distances to define the dissimilar-
ity among clusters. The most common and practical way to visualize the results is
through a plot called dendrogram: this allows reconstructing the merging history of
the objects of the studied data set from the beginning (each object forms a cluster
of its own) to the end of the clustering process (all objects are in the same cluster).
The dendrogram for our case application is plotted in Figure 5.7.
Figure 5.7: Dendogram associated to log-facies classification. A dendrogram consists

of many U-shaped lines connecting objects in a hierarchical tree. The stem of each U
represents the distance between the two objects being connected. Red clusters refer
to the connecting histories of the three recognized facies: low concentration turbidite
(LCT), mid concentration turbidite (MCT), high concentration turbidite (HCT).
In traditional log-facies analysis, the number of the classes (facies) that can be
identified from sedimentological information is often higher than the number of classes
of interest in reservoir models: the reason is that the quality of the data at the
well location, and the indirect measured data far away from the well (seismic and
electromagnetic data) do not allow us to distinguish sedimentological features with

the same accuracy. Based on the dendrogram that shows a very stable clustering
(Figure 5.7) and on core analysis calibration, a 3-facies classification, based on sand
concentration, is determined (Figure 5.8) and it consists of:
• low concentration turbidite (LCT) facies in green;
• mid concentration turbidite (MCT) facies in brown;
• high concentration turbidite (HCT) facies in yellow.
Figure 5.8: Log-facies classification performed at well location: LCT in green, MCT in
brown, and HCT in yellow. Log-facies classification is derived by using petrophysical
curves (porosity and clay content) and velocity data (VP /VS ratio). On the right
we show two crossplots in petrophysical (top right) and petroelastic (bottom right)
domain, color coded by facies classification.
The proposed facies classification is based on the percentage of quartz and clay
in the facies as derived from interpreted log curves. This classification is simplified
compared to sedimentological models but it is strongly consistent with the facies clas-
sification of the static reservoir model where porosity and net-to-gross are distributed
Log Standard Deviation

Neutron 7%
Density 0.015 g/cm3
Gamma Ray 5%
Resistivity 10%
Sonic 7%
Table 5.1: Standard deviations associated to log measurements.
following distributions estimated from well logs in each facies. Actually facies LCT
could be subdivided into 2 sub-classes, namely low and very low concentration tur-
bidite (which includes non-reservoir shale), as thin interbeddings of hemipelagic shales
have been observed in the well and as two clusters can be identified in the dendogram.
However these two classes are not distinguishable from the petro-elastic point of view
within the reservoir.
MC simulations are then exploited to provide the uncertainty propagation and
evaluation. The starting point are the normal PDFs associated to the set of wireline
logs (d̃) used in QLI with with standard deviations which are typical of the log
measurements and of the processing of raw data: Table 5.1 collects common standard
deviations (Viberti (2010)).
In this example a sensitivity analysis shows that 100 realizations are enough to
provide a stable solution and reliable results. Realizations of input curves sampled
from d̃ provide the set of cost functions to be minimized and the related set of
petrophysical curves defining p̃. In Figure 5.9 we plot the different realizations and
the median of the petrophysical sets of curves.
From the previous QLI analysis we have a number of realizations of the petro-
physical variables representing basic properties of the formation such as porosity and
mineral volumetric contents. Thus, we can generate a set of realizations of elastic
properties logs through a Monte Carlo simulation by applying the rock physics model
to the set of petrophysical curves realizations from quantitative log interpretation. In
the MC simulation we add a random error to account for the uncertainty associated
to RPM. rom the PDFs of p̃ and the previously calibrated rock physics model, the
set of acoustic and elastic curves and the corresponding PDFs of velocities r̃ are then
Figure 5.9: Set of 100 realizations of petrophysical curves (gray curves), from left
to right: effective porosity, volume of clay, and volume of quartz (volume of silt is
computed by difference 1 minus the sum of effective porosity, clay, and quartz). The
pointwise median curve (P50) is displayed in red.
computed (Figure 5.10). In Figure 5.11 we show the full probability distributions
of porosity and P-wave velocity, and as an example at two given depth locations we
extract the corresponding histograms.
On the base of the set of 100 realizations of effective porosity, volume of clay and
VP /VS ratio, we can obtain 100 training sets sorted in t̃ to be used to generate 100
profiles of facies. In fact, from the previous steps of the methodology, we have obtained
100 realizations of petrophysical curves (QLI results, Figure 5.10) and 100 realizations
of rock physics curves (RPM results, Figure 5.10). We now perform 100 log-facies
Figure 5.10: Set of 100 realizations of elastic curves (gray curves), from left to right:
P-wave and S-wave velocity. The pointwise median curve (P50) is displayed in red.
classifications through MCSs (Figure 5.12, top), and estimate the posterior probability
on facies classification, in addition to the most probable facies profile. In particular
cluster analysis is applied to the 100 realizations to obtain 100 profiles of log-facies;
then the frequency of the facies, at each depth location of the profile, provides the
posterior probability of facies occurrence and the most likely facies estimation.
In Figure 5.12 (bottom), we show, as an example, 5 realizations of facies obtained
from Monte Carlo simulations: the thick high porosity sand layer is well identified in
all the profiles, whereas the thin layers at the top and at the bottom of the reservoir
are more uncertain. This means that in the upper and lower part of the reservoir the
log-facies classification can vary as a function of RPM and QLI input parameters in a
traditional deterministic workflow and the underestimation of the uncertainty could
Figure 5.11: Posterior probability density functions of porosity (left side) and P-wave
velocity (right side). For two depth locations z1 and z2 we also show the histogram
of the simulated values.
lead to a misclassification of the facies.

From the so-generated 100 profiles we can count, point by point along the profile,
the frequency of occurrence of each log-facies and infer the posterior probability dis-
tribution c, by normalizing, at each data level, the frequencies by the total number of
realizations (100). In Figure 5.13 we show the probability of log-facies obtained from
petro-elastic Monte Carlo realizations and the final most probable classification. We
notice that in the mid part of the profile the probability of having a HCT facies is very
high and the uncertainty is small, whereas in the thin layers zones the uncertainty is
much higher.
This is also reflected in the entropy curve that we plot together with the most
likely facies profile (Figure 5.13) in order to quantify the uncertainty in the discrete
probabilistic scenario so provided. Entropy is computed in base 3 since we have
discriminated 3 different facies, in order to get values between 0 (no uncertainty) and
1 (maximum uncertainty).
Figure 5.12: Set of 100 realizations of facies (top left) at well location and estimated
most likely facies profile (top right). On the bottom 5 realizations are shown: LCT
in green, MCT in brown, and HCT in yellow.
To interpret entropy results we comment here on the surface described by the

variation of the probabilities related to the three discriminated facies. Since the three
probabilities sum to 1, we can take any two as independent variables to map the
behavior of entropy for different values of the facies probabilities. Specifically:
h(cHCT , cLCT ) = −[cHCT log3 (cHCT )+cLCT log3 (cLCT )+(1−cHCT −cLCT ) log3 (1−cHCT −cLCT )]
(5.30)
with cHCT + cLCT ≤ 1, where cHCT is the probability of obtaining HCT at a given
depth, and cLCT is the probability associated with LCT; while cM CT = (1 − cHCT −
cLCT ). As previously explained the three probabilities are collected in the vector
Figure 5.13: Posterior probability distribution of facies estimated by Monte Carlo

simulation (left), most likely facies profile (middle) and associated entropy function
(right): LCT in green, MCT in brown, and HCT in yellow. On the right we show
two crossplots in petrophysical (top right) and petroelastic (bottom right) domain,
color coded by the associated entropy given by the probabilistic facies classification.
c = (cLCT , cM CT , cHCT ). When the three facies are equiprobable, we have maximum
entropy, whereas when the probability of a given facies is close to 1, then the entropy
tends to 0. The two crossplots on the right of Figure 5.13 help in clarifying the
information provided by entropy analysis. With Figure 5.8 in mind (right), it can be
easily seen that the transition areas between the different facies are characterized by
high entropy (yellow clouds in Figure 5.13) as expected: this fact mainly affects the
generalized high entropy associated to MCTs. On the other hand, the extreme LCT
and HCT facies show low entropy (blue clouds).
5.5 Application: second example

The second case study is in the North Sea: the reservoir is located at approximately
1750m, has an average thickness of about 80m, and it is part of a complex fluvio-
deltaic system with sequences of sandstone and shale. Sandstones and shales layers are
relatively thin compared to the seismic resolution. Several works have been published
on a number of nearby fields in the North Sea (Mukerji et al. (2001a), Mukerji et al.
(2001b), and Avseth et al. (2005)).
A complete set of well logs is available for one well of the area. The well is slightly
deviated and passes through the main reservoir layer filled with oil. The set of sonic
logs, density, and a set of preliminary petrophysical curves performed with commer-
cial software are shown in Figure 5.14. We point out that three main mineralogical
fractions have been identified: clay (mainly illite), muscovite, and quartz. Porosity
is relatively high in clean sand in the upper part of the reservoir. The quality of
the sonic log is generally good except for few data samples where some high peaks
in P-wave velocity are recorded while no changes are recorded by S-wave velocity,
generating unrealistic values of Poisson ratio and VP /VS ratio.
Furthermore a preliminary facies classification has been performed based on sed-
imentological information, accurate depositional models, and core sample analysis.
A set of 8 facies has been identified in this reservoir: 1) marine silty-shale; 2) pro-
delta; 3) flood plain; 4) mouth bar; 5) distributary channel; 6) crevasse splay; 7)
tidal deltaic lobes; 8) tight. The distribution of porosity and clay content color coded
by facies classification (Figure 5.15) shows that some of these sedimentilogical facies
cannot be discriminated by petrophysical properties: for example marine silty-shale,
pro-delta and flood plain have similar distributions in terms of porosity and clay
content (Figure 5.15). As a consequence the elastic response of some facies is approx-
imately the same. For this reason, we applied the proposed methodology with two
different classifications: first we used the 8-facies sedimentological definition; then we
used a simplified 3-facies classification, namely sand, silty-sand (mixed), and shale,
to classify facies that can be recognized at seismic scale.
As in the previous example we show the different steps of the methodology. First
Figure 5.14: Well log dataset and preliminary petrophysical curves at well location,
from left to right: P-wave and S-wave velocity; density; porosity (total porosity in
blue, effective porosity in red); volumetric fraction curves (clay in green, quartz in
yellow, and muscovite in brown); and water saturation.
of all, a rock physics model is calibrated at the well location: in this case we applied
a stiff sand model (Mavko et al. (2009)). Both soft sand and stiff sand models
belong to the group of granular media models and are based on Hertz-Mindlin contact
theory: soft sand model extrapolates elastic property values to low porosities by
using modified Hashin-Shtrikman lower bound; stiff sand model uses modified Hashin-
Shtrikman upper bound, resulting in an increase of elastic properties values compared
to soft sand model. Similarly to the previous case, the calibration is performed
in wet conditions by comparing sonic logs data and rock physics model predictions
(Figure 5.16). The model parameters are: effective pressure P = 27MP a, critical
porosity of φ0 = 0.42, and coordination number n = 9. The overestimation of S-wave
velocity in stiff sand model is well known (Mavko et al. (2009)), for this reason we
applied a factor of 3/4 to reduce S-wave velocity predictions and match the data: this
correction is purely heuristic but it is often applied in real cases (a physical explanation
Figure 5.15: Preliminary facies classification: a set of 8 facies has been identified in
this reservoir: 1) marine silty-shale; 2) pro-delta; 3) flood plain; 4) mouth bar; 5)
distributary channel; 6) crevasse splay; 7) tidal deltaic lobes; 8) tight (left). Grouped
histogram of effective porosity (top right) and clay content (bottom right) as a func-
tion of facies classification.
of this reducing factor for shear wave can be found in Bachrach and Avseth (2008)).
Another method to achieve a good fit is to reduce the values of shear moduli of
mineralogical fractions but this could lead to unrealistic values for this mineralogy
and would require a recalibration of P-wave velocity. The direct comparison of rock
physics model predictions and sonic logs is shown in Figure 5.17: generally we have
a good agreement except for the data samples with high peaks in P-wave velocity. A
slight overestimation of velocity in the upper shaly layer is observable, but this could
be due to a different mineralogical composition of clay in the overcap layer on top of
the reservoir.
We then apply the uncertainty propagation methodology to petrophysical prop-
erties and rock physics elastic attributes. We generate a set of 50 realizations of
effective porosity, volume of clay and volume of quartz (volume of muscovite is com-
puted by difference to guarantee that the sum of the mineralogical fraction is 100%);
we then compute the corresponding set of P-wave and S-wave velocities and density
(Figure 5.17). The parameters used for the probabilistic formation evaluation analysis
are the same used in the previous case and summarized in Table 5.1.
Figure 5.16: Rock physics crossplots: (left) P-wave velocity versus effective porosity;
(right) S-wave velocity versus effective porosity color coded by volume of clay. Black
lines represent stiff sand model for different clay contents (from top to bottom: 10%,
20%, 30%, 40%, 50% and 60%).
Rock physics diagnostics (Figure 5.15) showed that, for this specific case, the use
of S-wave velocity could improve facies classification in the combined petro-elastic
domain. For this reason we decided to use in this case an extended dataset made by
effective porosity, volume of clay, volume of quartz, P-wave, and S-wave velocities.
We then performed the facies classification with 8 sedimentological facies. In such
application we expect more variability and higher entropy compared to the previous
example as the number of facies is higher and some facies cannot be discriminated
from the petrophysical and/or elastic point of view. The full set of 50 realizations is
shown in Figure 5.18: by comparing all the realizations we can detect some similar
features (such as layers made by facies 1, 2, and 3, and layers made by facies 4,
and 5 in the upper part) but as expected it is hard to discriminate between facies
with similar petro-elastic properties. The extracted statistics, in particular the etype
(ensemble average) and the maximum a posteriori (MAP) of facies, help to interpret
the results (Figure 5.18). The etype is the ensemble average of the set of models
and it is a continuous variable; the maximum a posteriori represents the facies that
maximizes the posterior probability estimated from the ensemble of realizations as
Figure 5.17: Set of 50 realizations of petrophysical and elastic curves (gray curves),
from left to right: effective porosity, volume of clay, and volume of quartz (volume of
silt is computed by difference 1 minus the sum of effective porosity, clay, and quartz),
P-wave velocity and S-wave velocity predictions. The pointwise median curve (P50)
is displayed in red. The dashed blue line represents sonic log data.
the facies frequency divided by the total number of realizations (Figure 5.19). In the
upper part of the profile we can recognize the shaly overcap made by facies 1, 2, and
3 and the thicker layer of the main reservoir made by facies 4 and 5. In the lower
part the variability in the classification is higher, which results in a sequence of thin
layers. However if we look at the whole set of realizations we observe that even in
the upper part of the reservoir the variability is quite high; this is confirmed by the
entropy profile (Figure 5.19). Differently from the previous case the overall entropy is
generally high: this can be due to the large number of facies in the classification but
also to the vertical heterogeneity of petrophysical and elastic properties (Figure 5.17).
To improve the facies classification at the well location, we introduced here a new
approach based on Markov chain methods (Krumbein and Dacey (1969)). Markov
chains are a statistical tool that has been used in geophysics to simulate facies se-
quences to capture the main features of the depositional process. Markov chains are
Figure 5.18: Set of 50 realizations of facies (left) at well location, etype (ensemble
average) and estimated most likely facies profile (top right). Color codes are the same
as in Figure 5.15.
based on a set of conditional probabilities that describe the dependency of the facies
value at a given location with the facies values at the locations above (upward chain)
or the locations below (downward chain). The chain is said to be first-order, if the
transition from one facies to another depends only on the immediately preceding fa-
cies. The transition matrix is built by assuming (or estimating from other wells in
the field or nearby fields) prior proportions and transition probabilities. The matrix
in our case is 8 by 8 elements: the terms on the diagonal of the transition matrix are
related to the thickness of the layers. The higher are the numbers on the diagonal, the
higher is the probability to observe no transition, and as a consequence the thicker
will be the layer.
In our application the posterior probability of facies estimated from logs of petro-
elastic properties is combined with the probabilities of the transition matrix (for
mathematical details, see Grana and Della Rossa (2010)) by a simple integration,
and facies values are sampled from the resulting distribution. In Figure 5.19 (right)
we show one realization of facies obtained by this technique: the use of Markov chain
allows us to account for vertical continuity and produces a more realistic classification
in cases where a high number of facies is identified. However the main limitation of
Figure 5.19: Facies probability estimated from ensemble and extracted statistics, from
left to right: probability of facies (first three plots); entropy, maximum a posteriori of
facies distribution and new facies classification performed by Markov chain integrated
approach. Color codes are the same as in Figure 5.15.
such a choice is due to the assumptions required to assemble the transition matrix.
In our case, for example, we do not have any information about the distribution and
thickness of facies 7 and 8 and the results we obtained only depend on the information
assumed from the depositional and sedimentological model.
We finally repeated the same application with a simplified classification: we as-
sumed 3 facies in the reservoir that we called for simplicity shale, silty-sand, and
sand. The reservoir characterization study performed on this field had shown that
these facies can be discriminated in the petro-elastic domain and represent the fa-
cies classification recognizable at seismic scale. Figure 5.20 shows the comparison of
the results obtained with the different classification. The methodology clearly works
better when a limited number of facies is identified (also by comparing the entropy
profiles of Figures 5.19 and 5.20); however we notice that there is a good agreement
between the two classifications as facies 1, 2, and 3 (in 8 facies definition) correspond
to shaly layers, and facies 4, and 5 (in 8 facies definition) correspond to sandy layers.
Figure 5.20: Comparison of two different classifications. On the left (first two plots)
we show the 8-facies classification: maximum a posteriori of facies and Markov chain
classification (color codes are the same as in Figure 5.15). On the right we show the
simplified 3-facies classification (namely seismic facies classification, last three plots):
posterior probability distribution, entropy, and maximum a posteriori of facies (shale
in green, silty-sand in brown, sand in yellow).
5.6 Discussion
The proposed methodology classifies log-facies at well location by integrating three
different disciplines: formation evaluation, rock physics modeling, and cluster anal-
ysis, by means of a statistical approach, that allows us to understand how plausible
the obtained interpretation is or, in other words, how large the uncertainties are in
the obtained solution. The assessment of uncertainty in geophysical inverse problems
has been addressed in several works (Jackson (1972), Jackson (1979), and Taran-
tola (2005)) but a standard application to well log analysis is in general neglected.
Classical deterministic workflows in formation evaluation analysis and, in particular,
log-facies classification, do not allow us to propagate the uncertainty in measured data
to the property estimates in a robust and comprehensive way. Statistical methods,
on the other hand, can incorporate all available information on the studied system
(observational data, theoretical predictions and prior knowledge) and this informa-
tion can be represented by probability distribution functions. According to this, the
solution is not unique, but rather it consists of a posterior probability distribution
function over the model domain space which describes the probability of a given
model being the closest to the true one.
The aim of this chapter is to enhance the classical approach by introducing Monte
Carlo simulations providing a proper way to account for a statistical view. Since
each step of the methodology strongly depends on the previous one, a robust uncer-
tainty analysis and classification from the very beginning (initial acquired dataset) is a
mandatory requirement for a safe and trustworthy probability distribution definition.
The input uncertainties represent the critical point of the methodology. Formation
evaluation is the aforementioned starting point and so it drives the remaining model-
ing. As we have already pointed out, the prior characterization of input uncertainties
for quantitative log interpretation relies on separating the different sources of errors
associated with tool measurements, environmental corrections, pre-processing com-
putations and model parameterizations. Systematic errors need to be corrected since
their propagation is meaningless. Random errors may not be neglected when the
number of measurements to be averaged is low and/or the measurement precision is
very poor.
Monte Carlo simulations try to address this uncertainty issue but a robust and
straightforward way to account for all the sources of uncertainty is still lacking. How-
ever, other open problems exist and two of the major ones are the possible correlation
existing between the uncertainty affecting the different log acquisitions and the vari-
ous reference volumes of investigation involved (vertical and lateral resolution). The
assignment of probability distribution functions to log measurements assumes a fixed
and certain value of the given nominal depth. A more robust approach should also
consider the vertical uncertainty related to the recorded depth of any log measure-
ments.
An important point of discussion is also related to the constrained number of
facies in the probabilistic workflow. In particular the methodology quantifies and
propagates the uncertainty through log-facies classification once the number of facies
has been a priori fixed. Monte Carlo simulations are applied to the multivariate
statistical technique chosen to discriminate a given number of log-facies. Different
petrophysical and elastic scenarios in the probabilistic training set could be better
described by a number of facies which is higher or lower with respect to the a priori
fixed one. A generalization of the methodology in order to take into account this
issue could also help for a more rigorous characterization.
The application of the entire workflow to real datasets seems to confirm its fea-
sibility and reliability. Log-facies probability and entropy evaluation quantify the
uncertainty in the classification and represent a probabilistic approach from the very
beginning of borehole data interpretation for seismic reservoir characterization. It is
worth mentioning that even more complex systems need to be tested and studied in
order to finally state all the potentiality of the methodology here developed.
It is also worth mentioning the choice of Monte Carlo simulations for accounting
the uncertainty propagation problem. First of all, Monte Carlo techniques calculate
the probability density of any functions of random variables and do not require any a
priori assumption on the type of distribution for the results. In fact, quantitative log
interpretation and rock physics modeling can be highly non-linear and the uncertainty
propagation problem cannot be solved analytically. Under this light, Monte Carlo
simulations provide a simple means by which uncertainties in inputs can be translated
into uncertainties in the calculated output properties. Moreover, this methodology
is very flexible, allowing different interpretation models to be built and uncertainties
tested in a robust way. Correlated as well as independent parameters can be handled
and all the possible constraints defining the problem can be taken into account in a
straightforward way. The downside to Monte Carlo simulations is that a large number
of simulations are typically required for meaningful statistics to be developed. So,
it is fundamental to understand the number of samplings needed in order to get
stable results. To conclude, by Monte Carlo simulations we have a powerful tool
to propagate uncertainties in the different steps of the methodology: the technique
can be applied to different studies, independently of the physical models (defining
formation evaluation and rock physics interpretation) and mathematical techniques
(for log-facies classification) chosen for the specific case.

As described in the application sections, in this work we performed log-facies clas-
sification following a hierarchical agglomerative clustering algorithm with a Ward’s
minimum variance linkage method. Other methods can be used to classify facies, such
as linear discriminant analysis, partitioning cluster algorithms or neural networks. In
particular, in petrophysics neural networks are used in unsupervised learning mode.
The different classification algorithms were previously investigated in order to rank
the obtained results, undertaking tests on known synthetic data sets of variable com-
plexity. The objective was to understand which methods are the most effective in gen-
erating a log-facies classification suitable for reservoir characterization purposes. In
particular, the proposed and here used hierarchical agglomerative Ward’s technique
has been evaluated by far the best, as it correctly classified most of the synthetic
data sets. K-means clustering, unsupervised neural networks, centroid linkage meth-
ods and unweighted average (UPGMA) with Euclidean distance algorithms showed
reasonably good classification rates, even though they were not suitable for arbitrary-
shaped data clouds. Moreover, no significant advantages were observed when Man-
hattan or Mahalanobis distance were used in substitution for the Euclidean distance
in the UPGMA algorithm. Other clustering algorithms (e.g. single linkage, complete
linkage, weighted average) have been proved unsuccessful since they were character-
ized by high misclassification rates. Given the above considerations on the ranking
and availability of the different classification algorithms, we here focused on Ward’s
hierarchical agglomerative method. However, the overall methodology works with any
appropriate classification algorithm. As previously mentioned, the main advantage
of the presented workflow is that the Monte Carlo simulations allow us to extend the
training dataset and to propagate the uncertainty from the input measured data to
the facies classification.
Uncertainty evaluation is important from a qualitative point of view to assess the
reliability of the estimated properties, but it can be used in quantitative modeling by
using the full probability distributions of the property. In reservoir modeling, many
geostatistical simulations methods (for example sequential indicator simulation and
sequential Gaussian simulation) require input prior distributions of the properties we
want to model, i.e. facies proportions and the prior distributions of rock properties
(such as porosity and net-to-gross). These distributions are generally assumed from
prior geological knowledge of the field or nearby fields; but in our case the distribu-
tions can be extracted directly from Monte Carlo simulations. Nevertheless we point
out that the PDFs derived from Monte Carlo simulations are depth dependent; then
stationary conditions should be first verified before extrapolating marginal distribu-
tions. In a practical reservoir modeling workflow (see for example Doyen (2007)), we
first generate reservoir facies models with sequential indicator simulations (or multi-
point geostatistics if a suitable training image is available) and then generate porosity
and other reservoir property models with sequential Gaussian simulation where the
prior distributions of reservoir properties are facies-dependent and are estimated from
the marginal distributions derived from Monte Carlo simulations in each facies. An
important point of our methodology is the integration of rock physics modeling. The
main advantage of this step is the link between facies classification and elastic prop-
erties, which is provided by the model calibrated at the well location. In fact, this
link guarantees that facies classification estimated from seismic amplitudes in seismic
characterization studies is consistent with log-facies classification at well location.
Instead of rock physics predictions, sonic logs could be used as well but the corre-
lation between sonic logs and petrophysical curves is not always optimal in well log
analysis. Furthermore, resolution of sonic logs is generally lower than the resolution
of well logs used in formation evaluation analysis. If the sonic logs are reliable and
correlated to petrophysical logs, the lower resolution of sonic logs would reduce the
entropy but at the same time increase the uncertainty in the output estimation. On
the other side rock physics models are applied to a set of petrophysical curves with
similar resolution, which results in a complete consistent petro-elastic dataset. The
drawback of such a choice is that the rock physics model approximation could be not
accurate in some interval layers, as pointed out in our applications, but this lack of
accuracy is partially compensated by Monte Carlo multiple realizations.
We finally point out that in our methodology we tried to include different sources
of uncertainty: log measurements, rock physics model approximations, heterogeneity
and natural variability of the rocks. However other factors could influence formation
evaluation analysis, rock physics modeling, and log-facies classification, such as for
example anisotropy, laminar and/or dipping layering, and well deviation. These is-
sues should be carefully evaluated in each case and a sensitivity analysis should be
performed to verify if they can be neglected or not.
5.7 Conclusion
The proposed probabilistic log-facies classification workflow is based on the integra-
tion of three different disciplines: quantitative log interpretation, rock physics model-
ing, and multivariate statistical techniques (cluster analysis) for log-facies classifica-
tion., Furthermore, this workflow focuses on the uncertainty propagation through the
three different steps, and it is based on Monte Carlo simulation method. The applica-
tion of the proposed methodology allows to generate several log-facies profiles at well
location and to estimate the most probable facies classification and the associated
uncertainty via entropy computation. The so obtained classification includes both
petrophysical properties and acoustic/elastic attributes to directly link log-facies to
seismic velocities used in facies simulation constrained by seismic information. This
workflow represents a rigorous starting point to propagate the well facies classification
and related uncertainty to the entire reservoir modeling. These obtained uncertainties
are more realistic inputs to reservoir modeling, rather than taking log data as fixed
and certain (so-called ”hard data” in geostatistics). The method has been illustrated
on real well log datasets in clastic environments, but it can be applied to different
scenarios.
Chapter 6
Pressure dependence of elastic

properties
6.1 Abstract
In time-lapse studies, the knowledge of the saturation and pressure effects on elastic
properties is a key factor. The relation between saturation changes and velocity vari-
ations is well known in rock physics and at seismic frequency it can be satisfactorily
described by Gassmann’s equations. The pressure effect still requires deeper inves-
tigations in order to be included in rock physics models for 4D studies. Theoretical
models of velocity-pressure relations often do not match lab measurements, or contain
empirical constants or theoretical parameters that are difficult to calibrate or do not
have a precise physical meaning. In this chapter, I present a new model to describe the
pressure sensitivity of elastic moduli for sandstones. This relation is then integrated
with a complete rock physics model to describe the relation between rock properties
(porosity and clay content) and dynamic conditions (saturation and pressure) and
elastic properties. Our new model is calibrated with lab measurements of room-dry
sand samples over a wide range of pressure variations and then applied to well data
to simulate different production scenarios. The complete rock physics model is then
used in time-lapse inversion to predict the spatial distribution of dynamic property
changes in the reservoir within a Bayesian framework.
159
CHAPTER 6. PRESSURE DEPENDENCE OF ELASTIC PROPERTIES 160
6.2 Introduction
Isotropic effective pressure is defined as the difference between the overburden pres-
sure, which is the integral of density from the surface to the depth of interest times
the acceleration due to gravity, and the pore pressure (Terzaghi (1943)). The effective
pressure, Pe , is taken as the difference between the lithostatic overburden pressure Po
and the hydrostatic pressure Pp (see for example, Mindlin (1949), Christensen and
Wang (1985)):
Pe = Po − Pp (6.1)
In traditional workflows, effective pressure - velocity transformations are calibrated

by using direct formation pressure measurements at the well, or by fitting lab mea-
surements. The effective pressure - velocity model can be a simple polynomial or
exponential relation or a more complex equation that also contains other rock prop-
erties, such as porosity and clay content. Eberhart proposed a more complex empirical
model to relate rock properties and isotropic effective pressure to elastic attributes
using the Han’s dataset (Han (1986)). More recent papers contain variants of those
equations (Dvorkin et al. (1996); Dutta (2002); Sayers et al. (2003)). In Gutierrez
et al. (2006) different models are, suggested the most appropriate relationships, and
discussed the calibration of the adjustable parameters, which appears to be critical.
All these papers contain relations calibrated on dry-rock velocities. MacBeth (2004)
proposed an analogous equation to link dry rock elastic moduli to effective pressure
with an exponential equation by fitting a set of lab measurements conducted on sand
and shaly sand dry samples. Multi-property relations are taken into account in the
model proposed by Saul and Lumley (2012). These two relations can be directly inte-
grated in rock physics models used in reservoir characterization to describe dry rock
elastic moduli variations as a function of effective pressure, the effect of fluid being
then modeled by Gassmann’s equation. However the empirical parameters depend
on porosity and mineralogy of the rock under study.
6.3 Physical model and application to lab data

The goal of this chapter is to find a reliable relation to link dry rock elastic moduli to
the effective pressure to be used in time-lapse studies. Most of the relations published
in literature so far are empirical and contain some parameters that must be fitted by
using lab measurements, but these empirical constants change as a function of porosity
and mineralogy.
Laboratory measurements show that velocities, as well as the elastic moduli, gen-
erally increase with effective pressure and the relation is generally non-linear (Han
(1986)). A relevant laboratory dataset of dry-rock velocity data is Han’s dataset,
from which we selected a subset of 32 samples whose clay content varies between 5%
and 25% (Figure 6.1). In the low effective pressure range, we observe a significant
increase in velocity, whereas for higher pressure values, velocity changes tend to be-
come smaller and exhibit an asymptotic behavior. Therefore, the relations between
pressure and velocity and pressure and elastic moduli can be expressed by an ex-
ponential function that reaches the asymptotic value when pressure becomes large.
However if we observe the measured data, we notice that even though each sample
has a similar exponential behavior, the slope of the curve, the initial value and the
asymptotic value are generally not the same. If we color code the data by porosity,
we observe that the curves tend to move downward as porosity increases: as a matter
of fact, if the clay content is approximately constant, generally the softer is the rock,
the lower is the velocity; if the clay content changes, the relation is more complex,
but still follows the general trend. The slope and the asymptotic value depend on
different parameters, such as initial porosity, lithology and number of cracks in the
rock. A similar behavior can be observed for the bulk modulus, as well as for S-wave
velocity and the shear modulus. For density, the behavior is generally more linear
and it reflects the porosity reduction due to increasing effective pressure. In Han’s
case no relevant variations in porosity have been observed and density was assumed
to be constant.
Multi-linear regression can be used to link elastic property changes to effective
35 0.3
30
0.25
Dry Bulk Modulus (GPa)

25
0.2
20
0.15
15
0.1
10
5 0.05
0 10 20 30 40 50
Effective pressure (MPa)
Figure 6.1: Han’s dataset: dry-rock bulk modulus versus effective pressure color coded
by porosity.
pressure, porosity and clay content, but these relations are generally difficult to cali-
brate due to the lack of exhaustive datasets. Alternatively we can reduce the number
of variables by introducing the concept of self-similarity (Dvorkin (2007)). In partic-
ular Dvorkin (2007) proposed to use a single variable given by the linear combination
of porosity and clay content: φ + avclay , where a is an empirical constant (Figure 6.2).
The petrophysical-properties dependence of velocity becomes more clear thanks to
the self-similarity concept.
Porosity+0.3Clay
35 0.3 35 0.35
30 30
0.25 0.3
25 25
0.2 0.25
20 20
0.15 0.2
15 15
0.1 0.15
10 10
5 0.05 5 0.1
0 10 20 30 40 50 0 10 20 30 40 50
Effective pressure (MPa) Effective pressure (MPa)
Figure 6.2: Han’s dataset: dry-rock bulk modulus versus effective pressure. The color
represents porosity (left) and a linear combination of porosity and clay content(right).
The two relations that better describe the exponential behavior are: Eberhart-
Phillips and MacBeth relations. Eberhart-Phillips’s relation links velocity and pres-
sure and it contains porosity and clay content terms:
√
VP = α − βφ − γ vclay + δ(Pe − e−εPe ) (6.2)
where α, β, γ, δ, ε are empirical parameters. This equation can be divided into two
√
terms: the first term α − βφ − γ vclay does not depend on pressure, the second is the
pressure-dependent exponential component. In other words the slope of the curves
obtained from this model is assumed to be the same and independent of porosity
and lithology, and the curves differ only for the changes in the first term, that shifts
the curve upward and downward depending on porosity and clay content. If we fit
each sample independently (Figure 6.3, left), the fitted parameters will depend on the
specific sample; on the other hand, if we fit all the samples simultaneously, the slope
of the curves does not change: only the initial point and the asymptotic value are
shifted (Figure 6.3, right). If we fit all the data simultaneously we obtain the following
values α = 5.77, β = 6.94, γ = 1.73, δ = 0.446, ε = 16.7 as in Eberhart-Phillips et al.
(1989).
5 0.3 5 0.3
4.5 4.5
0.25 0.25
P−wave velocity (km/s)
P−wave velocity (km/s)
4 4
0.2 0.2
3.5 3.5
0.15 0.15
3 3
0.1 0.1
2.5 2.5
2 0.05 2 0.05
0 10 20 30 40 50 0 10 20 30 40 50
Figure 6.3: Eberhart-Phillips relation: on the left each sample has been fitted inde-
pendently, on the right all samples have been fitted all together.
MacBeth’s relation links dry rock bulk modulus and effective pressure
K∞
Kdry (Pe ) = (6.3)
1 + AK e−Pe /PK
where K ∞ , AK , and PK are empirical parameters: K ∞ represents the asymptotic

value, whereas AK and PK are related to the slope. For the physical derivation
of Eq. 6.3 we refer the reader to MacBeth (2004). When we fit this relation inde-
pendently for each sample (Figure 6.4), we obtain different values for the empirical
parameters: if we plot these values versus porosity φ or a linear combination of poros-
ity and clay content such as φ + avclay , a being for instance equal to 0.3, we observe
a clear dependence for K ∞ and AK (Figure 6.5).
Porosity
35 0.3
30
0.25
Dry bulk modulus (GPa)
25
0.2
20
0.15
15
0.1
10
5 0.05
0 10 20 30 40 50
Figure 6.4: MacBeth relation (bulk modulus): each sample has been fitted indepen-
dently.
We also observe that K ∞ and AK are not independent. As a matter of fact, if K0

is the measured value at a given effective pressure P0 , then AK can be expressed as
a function of K ∞ and K0 :
K∞
, AK =
K0 = Kdry (Pe = P0 ) (6.4)
K ∞ − K0
To modify the Eq. 6.3 to include petrophysical-properties dependence, I used here
the concept of self similarity to express K ∞ as a linear function of φ + avclay (with
a = 0.3, Figure 6.6) and introduced Eq. 6.4. The modified equation then becomes:
K∞
Kdry (Pe ) = K∞ (P0 −Pe )/PK
, K ∞ = λ1 (φ + avclay ) + λ2 (6.5)
1 + K∞ −K0
e
35 30 16
30 25 14
∞
Parameter AK
Parameter PK
Parameter K
25 20 12
20 15 10
15 10 8
10 5 6
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
Porosity Porosity Porosity
35 30 16
30 25 14
∞
Parameter AK
Parameter PK
Parameter K
25 20 12
20 15 10
15 10 8
10 5 6
0.1 0.2 0.3 0.4 0.5 0.1 0.2 0.3 0.4 0.5 0.1 0.2 0.3 0.4 0.5
Porosity+0.3Clay Porosity+0.3Clay Porosity+0.3Clay
Figure 6.5: Sensitivity analysis on fitted parameters on MacBeth relation: (top) fitted
parameters versus porosity; (bottom) fitted parameters versus linear combination of
porosity and clay content.
where λ1 and λ2 are empirical parameters that must be fitted using lab mea-
surements. When we apply this equation to Han’s dataset, the results are quite
satisfactory (Figure 6.7). Similar results have been obtained for the shear modulus
(Figure 6.8). The modified equation for the shear modulus is:
µ∞
µdry (Pe ) = µ∞ , µ∞ = λ3 (φ + avclay ) + λ4 (6.6)
1 + µ∞ −µ0 e(P0 −Pe )/Pµ
where µ∞ is a linear function of φ + avclay , Pµ , λ3 and λ4 are empirical parameters,

µ0 is the shear modulus at the initial pressure P0 .
Pressure (MPa)
45 50 45
measurements P=50MPa
40 45 Linear regression
40
35 40 K=−95.9(φ+0.3 vclay)+41.6
35

30 35
30
25 30
25
20 25
20
15 20
15
10 15
5 10 10
0 5 5
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
Porosity+0.3Clay Porosity+0.3Clay
Figure 6.6: Self similarity concept: (left) measured bulk moduli versus a linear com-
bination of porosity and clay content color coded by effective pressure; (right) bulk
modulus at 50 MPa versus a linear combination of porosity and clay content and
linear fitting.
Porosity
35 0.3
30
0.25
25
0.2
20
0.15
15
0.1
10
5 0.05
0 10 20 30 40 50
Figure 6.7: Modified MacBeth relation (bulk modulus).

Porosity
0.3
25
Dry shear modulus (GPa) 0.25
20
0.2
15
0.15
10
0.1
0.05
0 10 20 30 40 50
Figure 6.8: Modified MacBeth relation (shear modulus).
Porosity Porosity
30 0.3 22 0.3
20
25 0.25 0.25
18
Dry shear modulus (GPa)
16
20 0.2 0.2
14
12
15 0.15 0.15
10
8
10 0.1 0.1
5 0.05 4 0.05
0 10 20 30 40 50 0 10 20 30 40 50
Figure 6.9: Modified MacBeth relation (set of five samples): bulk modulus (left) and
shear modulus (right).
To clarify the application we show the results by only using 5 samples (Figure 6.9).
In the next section we show how to apply the new equation in a complete rock physics
workflow. As shown in the previous chapters, in seismic reservoir characterization we
want to establish a set of rock physics relations to link reservoir properties with elastic
properties. We now extend these relations to time-lapse studies. The input data of
the rock physics model are: effective porosity, clay content, fluid saturations and
effective pressure. We first compute the elastic moduli of the rock frame (matrix)
using Voigt-Reuss-Hill average, we then compute the elastic moduli of the dry rock
using a suitable rock physics model (empirical relations, granular media models, or
inclusion models), we then apply the pressure effect using the modified MacBeth’s
equations (Eqs. 6.5 and 6.6), and we finally model the fluid effect using Gassmann’s
equations to obtain the elastic moduli of the saturated rock under assigned pressure
conditions.
6.4 Application to log data and sensitivity

In this section we want to explore the combined effect of pressure and saturation
changes on well log data and the corresponding synthetic seismograms.
We consider a complete set of well log data from an oil field in the North Sea.
Sonic logs and petrophysical curves from the formation evaluation analysis are shown
in Figure 6.10. Insitu reservoir conditions are: effective pressure 20 MPa and gas
saturation given by the saturation curve. We then design 8 different production
scenarios mimicking different production mechanism, such as gas injection, water
injection and depletion:
1. Effective pressure decreases (Pe = 10 MP a) and saturation is the initial satu-

ration;
2. Effective pressure increases (Pe = 25 MP a) and saturation is the initial satu-

ration;
3. Effective pressure equals Pe = 20 MP a and rock is water-saturated;

4. Effective pressure equals Pe = 20 MP a and rock is gas-saturated;
5. Effective pressure decreases (Pe = 10 MP a) and rock is water-saturated;
6. Effective pressure decreases (Pe = 10 MP a) and rock is gas-saturated;
7. Effective pressure increases (Pe = 25 MP a) and rock is water-saturated;
8. Effective pressure increases (Pe = 25 MP a) and rock is gas-saturated.
Quartz Porosity
Well log
2020 Clay Water saturation
Rock physics model
Silt
2040
2060
Depth (m)
2080
2100
2120
2140
2160
3000 4000 5000 1500 2000 2500 2.2 2.4 2.6 2.8 0 0.5 1 0 0.5 1
P−wave velocity (m/s) S−wave velocity (m/s) Density (g/cm3) Mineral fractions (v/v) Pore fluid (v/v)
Figure 6.10: Well log data: from left to right P-wave velocity, S-wave velocity, density
(well log in blue and rock physics model in red), and volumetric fractions (volume
of quartz in yellow, clay in green, silt in black, effective porosity in red, and water
saturation in blue).
We first applied a rock physics model (stiff sand model) and the performed the
fluid and pressure substitution on rock physics model prediction curves (Figure 6.10).
The effects on elastic properties are shown in Figures 6.11 (P-wave velocity), 6.12
(S-wave velocity), and 6.13 (density). The effects on seismic properties are computed
using a convolutional model and are shown in Figures 6.14 (short offsets), and 6.15
(long offsets).
P=10 MPa P=25 MPa P=20 MPa (insitu) P=20 MPa (insitu) P=10 MPa P=10 MPa P=25 MPa P=25 MPa
sw insitu sw insitu sw=1 sw=0 (gas) sw=1 sw=0 (gas) sw=1 sw=0 (gas)
2020
2040
2060
Depth (m)
2080
2100
2120
2140
2160
2500 3500 4500 2500 3500 4500 2500 3500 4500 2500 3500 4500 2500 3500 4500 2500 3500 4500 2500 3500 4500 2500 3500 4500
V (m/s) V (m/s) V (m/s) V (m/s) V (m/s) V (m/s) V (m/s) V (m/s)
p p p p p p p p
Figure 6.11: P-wave velocities computed through the rock physics model in 8 different
scenarios (in-situ condition in red, and production scenarios in black).
2020
2040
2060
Depth (m)
2080
2100
2120
2140
2160
1000 2000 3000 1000 2000 3000 1000 2000 3000 1000 2000 3000 1000 2000 3000 1000 2000 3000 1000 2000 3000 1000 2000 3000
V (m/s) V (m/s) V (m/s) V (m/s) V (m/s) V (m/s) V (m/s) V (m/s)
s s s s s s s s
Figure 6.12: S-wave velocities computed through the rock physics model in 8 different
production scenarios (in-situ condition in red, and production scenarios in black).
2020
2040
2060
Depth (m)
2080
2100
2120
2140
2160
2 2.5 3 2 2.5 3 2 2.5 3 2 2.5 3 2 2.5 3 2 2.5 3 2 2.5 3 2 2.5 3
Rho (g/cm3) Rho (g/cm3) Rho (g/cm3) Rho (g/cm3) Rho (g/cm3) Rho (g/cm3) Rho (g/cm3) Rho (g/cm3)
Figure 6.13: Density computed through the rock physics model in 8 different produc-
tion scenarios (in-situ condition in red, and production scenarios in black).
Seismograms Seismograms Seismograms Seismograms Seismograms Seismograms Seismograms Seismograms

1
1.02
1.04
1.06
Time (s)
1.08
1.1
1.12
0 25 50 75 100 0 25 50 75 100 0 25 50 75 100 0 25 50 75 100 0 25 50 75 100 0 25 50 75 100 0 25 50 75 100 0 25 50 75 100

Offset (m) Offset (m) Offset (m) Offset (m) Offset (m) Offset (m) Offset (m) Offset (m)
Figure 6.14: Synthetic seismograms in 8 different production scenarios (short offset).

Figure 6.15: Synthetic seismograms in 8 different production scenarios (long offset).
6.5 Conclusion
The proposed new model describes the behavior of bulk and shear moduli over a wide
range of effective pressures by including the porosity and lithology effects through the
self-similarity concept. The new model is semi-empirical, it assumes funiform fluid
distribution, and it must be fitted using laboratory measurements of rock samples in
dry conditions. The model equation is represented by an exponential relation that
shows the increase of dry elastic (bulk and shear) moduli as a function of effective
pressure. The fluid effect can be then introduced using the traditional Gassmann’s
equations. The new model also improves the accuracy of fluid substitution results
because the dry-rock moduli is a key input parameter to Gassmanns fluid substitu-
tion equations. Our new velocity-pressure can be used for improved prediction of
rock properties as a function of effective pressure and for time-lapse inversion and
interpretation.
Chapter 7
Bayesian time-lapse inversion
7.1 Abstract
Traditionally static reservoir models are obtained as a solution of an inverse problem
where reservoir properties, such as porosity and lithology, are estimated from seismic
data. With the emergence of time-lapse reservoir models, we can integrate static and
dynamic reservoir properties in the seismic reservoir characterization workflow. Here,
we propose a methodology to jointly estimate rock properties, such as porosity, and
dynamic property changes, such as pressure and saturation changes, from time-lapse
seismic data. The methodology is based on a full Bayesian approach to seismic inver-
sion and can be divided into two steps. First we estimate the conditional probability
of elastic property and their relative changes, then we estimate the posterior prob-
ability of rock properties and dynamic property changes. We applied the proposed
methodology to a synthetic reservoir study where we created synthetic seismic survey
for a real dynamic reservoir model including pre-production and production scenar-
ios. The final result is then a set of point-wise probability distributions that allow us
to predict the most probable reservoir models at each time step and to evaluate the
associated uncertainty.
173
CHAPTER 7. BAYESIAN TIME-LAPSE INVERSION 174
7.2 Introduction
Seismic data, repeatedly acquired at multiple times, monitor the changes in seismic
response during the production phase. Reservoir models are driven by the need to in-
clude these time-lapse seismic conditioning data in the reservoir fluid flow simulator,
in order to assess the predictions of dynamic property changes (in particular hy-
drocarbon saturations and pressure changes), compared to the pre-production static
model and initial conditions (see Landrø (2001), and Landrø et al. (2003)). Due to
the low resolution and degree of accuracy of seismic data, uncertainty quantification
in pressure and saturation changes plays a key role, which calls for a probabilistic
approach to time-lapse seismic inversion (see Buland and El Ouair (2006), and Trani
et al. (2011)).
Mostly, time-lapse seismic data has been used as a reservoir monitoring tool. If
the repeatability of the seismic data is sufficiently good, time-lapse seismic data will
contain small uncertainty and can be used to determine changes in the earth that
are related to reservoir production. For to this reason, time-lapse seismic data have
recently been used as an additional set of dynamic data for history matching in
reservoir applications (see for example Aanonsen et al. (2003) and Dong and Oliver
(2005)). A lot of efforts have been done in the direction of integration of production
data and time-lapse seismic data (see for example Huang et al. (1997), Kretz et al.
(2004) and Landa and Horne (1997)). Direct inversion methods based on first order or
second order approximations of the analytical expression of pressure and saturation
changes as a function of reflectivity coefficients have been presented as well (see
Landrø (2001), Landrø et al. (2003), Dadashpour et al. (2008) and Trani et al. (2011)).
In this work we propose a Bayesian inversion to estimate the posterior distribution
of pressure and saturation changes from time-lapse seismic data. The problem is split
into two sub-problems: the estimation of relative changes in elastic properties from
seismic differences and the estimation of saturation and pressure changes from elastic
property changes. First, the posterior distribution of relative changes in elastic prop-
erties is estimated from time-lapse partial-stacked seismic data using the Bayesian
time-lapse inversion proposed by Buland and El Ouair (2006)), that combines the
convolutional model with Aki-Richards linearized approximation of Zoeppritz equa-

tions (see Buland and Omre (2003)). Then the likelihood of saturation and pressure
changes given relative elastic property changes is estimated using rock physics models
and kernel density estimation methods. Finally the two probabilities are combined
to estimate the posterior distribution of reservoir property changes given time-lapse
seismic data. In order to improve reservoir property predictions, we propose a simulta-
neous Bayesian inversion of the pre-production seismic survey and time-lapse seismic
differences, to account for the correlation between static properties and changes in
dynamic properties. The methodology has been applied to a real reservoir model
with synthetic seismic data. The seismic inversion for elastic properties is performed
combining the approaches presented in Buland and Omre (2003) and Buland and El
Ouair (2006), whereas we perform the inversion for reservoir properties by using rock
physics models and non-parametric probabilistic techniques. Time-lapse seismic data
are corrected for time-shifts due to reservoir production, however this information
could be made probabilistic (Avseth et al. (2012)) and integrated in the inversion
workflow.
7.3 Methodology
In seismic reservoir characterization, we aim to solve the inverse problem of assessing
a model m given the data Sbase
Sbase = Gm + es (7.1)
where Sbase represents the base seismic survey, G is the forward linear model, m is
the elastic parameter, e.g. the logarithm of elastic impedances IP and IS , and es is a
random error. The linear operator G can be written as G = WAD where W is the
matrix containing the wavelets, A represents the reflectivity coefficients computed by
Aki-Richards approximation and D is a differential matrix (as in Buland and Omre
(2003)). In time-lapse studies we aim to solve the inverse problem of assessing ∆m
given ∆S with
∆S = G∆m + ǫs (7.2)
where ∆S is the difference between the repeated seismic survey and the base seismic
survey ∆S = Srep −Sbase , G is the same forward linear model as in Eq. 7.1, ∆m is the
model parameter consisting of the logarithm of relative changes in elastic properties,
and ǫs is a random error. The variable ∆m can be written as (Buland and El Ouair
(2006)):  rep 
IP
ln I base
∆m =  Prep
IS
. (7.3)
ln I base
S
Once we have estimated the elastic properties, we then aim to estimate the static
rock properties R, such as porosity, from the model m as a solution of another
inverse problem
m = fRP M (R) + er (7.4)
and the changes in reservoir properties ∆D, such as saturation and pressure changes,
from the model ∆m as
∆m = gRP M (∆D) + ǫr (7.5)
where the functions fRP M and gRP M are rock physics models and er and ǫr are random
errors. Different petro-elastic models fRP M , e.g. empirical relations, granular media
models, or inclusion models (Mavko et al. (2009)) can be used, and these are generally
calibrated with well data. The rock physics model gRP M relating elastic property
changes to pressure and saturation changes generally combines an empirical equation
with analytical models. In our work we use Gassmann’s equation for the saturation
effect (Mavko et al. (2009)), whereas for the pressure effect we use a modified version
of MacBeth’s relation (Mavko et al. (2009)) where we include the effect of initial
reservoir properties R.
By combining the formulations in Buland and Omre (2003) and Buland and El
Ouair (2006) we can write:
" # " #" # " #
Sbase G 0 m es
= + , (7.6)
∆S 0 G ∆m εs
and " # " #!

R m
= FRP M + eRP M , (7.7)
∆m ∆m
where FRP M is the rock physics model that include fRP M and gRP M and eRP M is the
error associated to it.
We propose here a Bayesian approach to solve simultaneously the combined inverse
problem (Eqs. 7.6 and 7.7). For the first step (seismic inversion, Eq. 7.6), we assume
a Gaussian prior distribution for the joint variable y = [m, ∆m]T . We then estimate
the likelihood function using a linearized seismic forward model consisting of the
convolution between the wavelet and the Aki-Richards linearized approximation of
Zoeppritz equations (see Buland and Omre (2003), and Buland and El Ouair (2006)).
Finally, we estimate the posterior distribution of the joint vector y|x, where x =
[Sbase , ∆S]T using Bayes’ rule (Tarantola (2005)):
" # " #! " # " #! " #!
m Sbase Sbase m m
P (y|x) = P | ∝P | P
∆m ∆S ∆S ∆m ∆m
(7.8)
T
In particular, if we assume that the joint vector y = [m, ∆m] is prior distributed
according to a Gaussian model:
" # " # " #!
m µm Σm 0
y= ∼N , (7.9)
∆m µψ 0 Σψ
then the posterior distribution y|x where x = [Sbase , ∆S]T is Gaussian:

y|x ∼ N µy|x , Σy|x (7.10)
and the expressions of the conditional mean µy|x and variance Σy|x is analytical.
For the second step (rock physics inversion, Eq. 7.7), if R represents the rock
properties in the reservoir, such as porosity, and ∆D represents the changes in dy-
namic properties in the reservoir, such as pressure and saturation changes, then from
rock physics theory (Mavko et al. (2009)) we know that the joint distribution of
w = [R, ∆D]T depends on y, i.e. on the initial elastic properties m and their relative
changes ∆m. Using Gassmann’s equation applied to well log data and the modified
MacBeth’s relation fitted on core measurements describing different geological and
production scenarios, we can estimate the likelihood function P (w|y):

" # " #! " # " #! " #!
R m m R R
P (w|y) = P | ∝P | P
∆D ∆m ∆m ∆D ∆D
(7.11)
The conditional distribution in Equation 7.11 is generally non-Gaussian and can-
not be described by parametric distributions. For this reason we estimated the likeli-
hood by using kernel density estimation (Silverman (1986)) where the kernel function
is the Epanechnikov kernel.
From the first step (seismic inversion) we obtained the posterior probability of
elastic properties and their relative changes conditioned by seismic data and their
differences P (y|x), whereas from the second step (rock physics) we obtained the
posterior probability of reservoir properties and their changes conditioned by elas-
tic properties P (w|y). Now we combine the two probabilistic information using
Chapman-Kolmogorov equation (Papoulis (1984)):
" # " #!
R Sbase
Z
P | = P (w|x) = P (w|y) P (y|x)dy (7.12)
∆D ∆S Rn
In order to compute the integral in Eq. 7.12, we use uniform discretizations of the
model spaces of x, y, and w, evaluate the probability at each point of the discretized
space, and compute the matrix product of the corresponding probability matrices.
From Eq. 7.12,, we can estimate the most likely model (by taking a point-wise sta-
tistical estimator, such as the mean or the median), or simulate different realizations
using geostatistical methods, such as probability field simulations.
7.4 Application: synthetic case

We tested the methodology using three different synthetic models: 1) a well log
dataset representing a three layer model, 2) a real well log dataset with synthetic
seismic, and 3) a 2D section of a real reservoir model with synthetic seismic.
In the first example we simply invert for elastic properties and their relative
changes the synthetic partial-stacked seismograms of a three layer model. The geolog-
ical model represents a thick sand layer embedded in two shaly layers. Both rock and
elastic properties are assumed to be constant in each layer and there is no error on the
seismogram. We then apply the Bayesian inversion proposed in the previous sections
where we combine the approaches presented by Buland and Omre (2003) and Bu-
land and El Ouair (2006). The elastic model is given by P-impedances, P-impedance
relative change, and S-impedance relative change. The prior model is the Gaussian
distribution in Eq. 7.9 where we assume that the mean value of P-impedance is a
low frequency model of the actual log of P-impedance (filtered to 4 Hz), the mean
value of the relative changes is equal to 0 everywhere, and the correlation between
elastic properties and their changes is 0. In Figure 7.1 we show the actual data and
the results of the inversion: the mean values and the percentiles P10 and P90 of the
point-wise posterior distributions.
P−impedance relative change S−impedance relative change P−impedance

2000
Estimated model
True model
Prior model
2020
2040
Depth (m)
2060
2080
2100
2120
0.8 1 1.2 1.4 0.9 1 1.1 1.2 4000 6000 8000
3 3 3
P−impedance (m/s g/cm ) S−impedance (m/s g/cm ) P−impedance (m/s g/cm )
Figure 7.1: Combined Bayesian inversion results for elastic properties and elastic
properties relative changes for a three-layer wedge model: estimated P-impedance
relative change, S-impedance relative change and P-impedance base model (black
curves represent the prior model, blue curves the actual model, solid red curves rep-
resent the mean values and dotted red curves percentiles represent P10 and P90).
In the second example, we use a real set of well logs coming from a well dataset in
the North Sea. At the well location we created synthetic partial-stacked seismograms
at different times to mimic realistic time-lapse data. Time-lapse seismic data are ob-
tained from pseudo-logs mimicking production scenarios as presented in Chapter 6. In
the following we present a production scenario where effective pressure increases due
water injection and oil production. The inversion is performed for elastic properties
(impedances and relative changes) and reservoir properties (porosity and saturation
and pressure changes). The full posterior distribution is show as background color in
Figure 7.2.
We then performed the inversion of elastic properties in terms of rock properties
(porosity and changes in dynamic properties (saturation and pressure) for two differ-
ent cases: in the first case we used a highly informative prior distribution for pressure
and saturation changes (Figure 7.3) estimated from the pseudo-wells used to create
synthetic seismic; in the second case we use a poorly informative prior distribution
for pressure and saturation changes to include other production scenarios in addition
to the real one.
P−impedance relative change S−impedance relative change P−impedance (base survey) Probability
2000 0.3
2020
0.25
2040
2060
0.2
2080
Depth (m)
2100 0.15
2120
0.1
2140
2160
0.05
2180
2200 0
0.9 1 1.1 1 1.1 6000 8000 10000 12000
3 3 3
P−impedance (m/s g/cm ) S−impedance (m/s g/cm ) P−impedance (m/s g/cm )
properties relative changes for a real well log dataset: estimated posterior distributions
of P-impedance relative change, S-impedance relative change and P-impedance base
model (black curves represent the actual model.
The results obtained for the first case are shown in Figure 7.4. For the second case
we used a uniform distribution for pressure, a bimodal distribution for porosity and a
non-parametric distribution for saturation (Figure 7.5). The corresponding results are
shown in Figure 7.6. The comparison of Figures 7.4 and 7.6 show that the uncertainty
in the posterior distribution generally increases when the prior distribution is less
informative.
Histrogram of water saturation changes Histrogram of effective pressure changes Histrogram of effective porosity
500 700 250
600
400 200
500
Frequency
300 150
400
300
200 100
200
100 50
100
0 0 0
0 0.5 1 0 2 4 6 0 0.1 0.2 0.3 0.4
Water saturation (v/v) Effective pressure (MPa) Effective porosity (v/v)
Figure 7.3: Highly informative prior distribution for pressure and saturation changes.
Water saturation change Effective pressure change Effective porosity Probability

2000 0.2
Predicted
2020 Measured 0.18
2040 0.16
2060 0.14
2080 0.12
Depth (m)
2100 0.1
2120 0.08
2140 0.06
2160 0.04
2180 0.02
2200 0
0 0.5 1 0 5 10 0 0.2 0.4
properties relative changes for a real well log dataset : estimated posterior distribu-
tions of P-impedance relative change, S-impedance relative change and P-impedance
base model (black curves represent the actual model).
Histrogram of water saturation changes Histrogram of effective pressure changes Histogram of effective porosity
2500 2500 2500
2000 2000 2000

Frequency
1500 1500 1500
1000 1000 1000
500 500 500
0 0 0
0 0.5 1 0 2 4 6 0 0.1 0.2 0.3
Figure 7.5: Poorly informative prior distribution for pressure and saturation changes.
Water saturation change Effective pressure change Effective porosity

2000
Predicted
2020 Measured
2040
2060
2080
Depth (m)
2100
2120
2140
2160
2180
2200
0 0.2 0.4 0.6 0.8 0 2 4 6 0 0.2 0.4
properties relative changes for a real well log dataset assuming a poorly informative
prior distribution: estimated posterior distributions of water saturation change, ef-
fective pressure change and porosity base model (black curves represent the actual
model.
The methodology has then been tested using a real reservoir model with synthetic
time-lapse seismic data. The input reservoir model represents a clastic reservoir
with sand and shale. The high-porosity sand layer at the top of the reservoir is
filled by gas, whereas the mid section of shaly sand contains oil. The production
mechanism consists of injecting water in the lower part of the reservoir and producing
hydrocarbon in the upper part. From the reservoir fluid flow simulator we extracted
a 2D section (Figure 7.7) at two different time steps: 2012 (pre-production) and 2017
(after 4 years of production).
Figure 7.7: Fluid saturations and effective pressure before (top) and after production
(mid). Bottom left, effective porosity; bottom right, net-to-gross.
For both time steps we computed the corresponding elastic properties using a
rock physics model and synthetic partial stacked seismic data with three different
angle stacks: near (10o), mid (20o ), and far (30o ) using a convolutional model. The
velocity and density models used to calculate the seismic dataset were computed using
a rock physics model that combines: stiff sand model, Gassmann fluid substitution
and modified MacBeth’s relation. We finally applied a correction to seismic data to
account for time-shifts using a warping technique (Figure 7.8). Time-lapse differences
are shown in Figure 7.9.
Base survey Repeated survey

Near (10o) Near (10o)
1100 1100
1150 1150
Time (ms)
1200 1200
1250 1250
50 100 150 50 100 150

Mid (20o) Mid (20o)
1100 1100
1150 1150
Time (ms)
1200 1200
1250 1250
50 100 150 50 100 150

Far (30o) Far (30o)
1100 1100
1150 1150
Time (ms)
1200 1200
1250 1250
50 100 150 50 100 150

Trace number Trace number
Figure 7.8: Base (left) and repeated (right) seismic survey (perfect data): from top
to bottom near (10o ), mid (20o ), and far (30o ).
Repeated survey − Base survey

Near (10o)
1100
Time (ms) 1150
1200
1250
20 40 60 80 100 120 140 160

o
Mid (20 )
1100
1150
Time (ms)
1200
1250
20 40 60 80 100 120 140 160

o
Far (30 )
1100
1150
Time (ms)
1200
1250
20 40 60 80 100 120 140 160

Trace number
Figure 7.9: Time-lapse seismic differences: from top to bottom near (10o ), mid (20o ),
and far (30o ).
In this application the model parameter m is P-impedance, ∆m is the variable

given in Eq. 7.3, R is effective porosity, and ∆D represents changes in effective
pressure, oil saturation and gas saturation. To build the prior model of [m, ∆m]T
we filtered the true impedance model (m) using a maximum frequency of 4 Hz,
while for ∆m we assumed a prior mean of 0 everywhere. The prior correlation
between m and ∆m is 0. The mean values of the marginal distributions of the
point-wise posterior probabilities of elastic properties and their relative changes (1 -
mrep /mbase ) are shown in Figure 7.10. The corresponding absolute changes assuming
that the initial model is correct, are shown in Figure 7.11. The rock physics likelihood
has been estimated using well log data and core measurements and combining Monte
Carlo simulations and rock physics models (Gassmann fluid substitution and modified
MacBeth’s relation). In particular the pressure-velocity relation has been established
using lab measurements at different effective pressure conditions (Chapter 6).
In Figure 7.12, we display the mean of the point-wise posterior distributions of
porosity, and saturation and pressure changes (Eq. 7.12). The high porosity layer at
the top of the reservoir and the changes in fluid contacts are well detected by the
inversion results, whereas pressure changes are slightly underestimated close to the
well locations. Since the methodology is a full Bayesian inversion, for each point in the
seismic grid we can evaluate the posterior distribution of each estimated property. In
Figure 7.13, we show the posterior distributions of the predicted parameters at a given
location in the seismic grid. The uncertainty in saturation changes is larger than for
other properties since the fluid effect on elastic properties is generally less detectable.
The posterior distribution of porosity is bimodal since it reflects the overall behavior
of porosity in a mixture of sand and shale.
Figure 7.10: Mean values of inverted elastic properties and elastic properties rela-
tive changes: from top to bottom P-impedance relative change, S-impedance relative
change, and P-impedance estimation from base survey.
Figure 7.11: Mean values of inverted elastic properties and elastic properties absolute
changes: from top to bottom P-impedance absolute change, S-impedance absolute
change, and P-impedance estimation from base survey.
Figure 7.12: Mean values of inverted reservoir properties and dynamic properties
relative changes: from top to bottom gas saturation change, oil saturation change,
effective pressure change, and porosity estimation.
Posterior probability Posterior probability

0.2 0.2
0.15 0.15
Probability
Probability
0.1 0.1
0.05 0.05
0 0
0 0.5 1 0 0.5 1
Water saturation change Oil saturation change
Posterior probability Posterior probability
0.2 0.2
0.15 0.15
Probability
Probability
0.1 0.1
0.05 0.05
0 0
−5 −2.5 0 2.5 5 0 0.1 0.2 0.3
Pressure change (MPa) Porosity
Figure 7.13: Examples of point-wise posterior distributions of reservoir properties

and dynamic properties relative changes from top-left to bottom-right gas saturation
change, oil saturation change, effective pressure change, and porosity.
7.5 Application: real case

We applied the inversion methodology to the Norne dataset, a real case study in the
Barentz Sea. It is an oil field and oil production started in 1997. Sea depth in the area
is about 380 m and the main reservoir is located at a deep of 2500-2700 m. Norne
field consists of two oil compartments; Norne main structure (Norne C, D and E
segments), which contains 97% of the oil in place, and the north-east segment (Norne
Gsegment). Total hydrocarbon column is 135 m which contains 110 m oil and 25
m gas. The Norne main structure is relative flat. The reservoir is subdivided into
four different formations. Hydrocarbons in this reservoir are located in the Lower to
Middle Jurassic sandstones. The reservoir sandstones are dominated by fine-grained
and well to very well stored arenites. Most of the sandstones are good reservoir rocks.
The porosity is in the range of 25-30%. The average reservoir thickness is 200 m.
Figure 7.14: Main structure of the geocellular model (blue, top layer in red) and
seismic survey geometry (black).
Figure 7.15: Active cells of the geocellular model (blue, top layer in red). Seismic
survey geometry is shown in black for comparison.
The dataset includes well log data, a base seismic survey (2001) and three repeated
seismic surveys in 2003, 2004 and 2006. Rock physics and petrophysics of the Norne
main field are based on data from two exploration wells: 6608/10-2 (well E2 in the
following) and 6608/10-3 (well F3 in the following). The geocellular model and the
dynamic model are available as well. Seismic acquisition geometry and the main
structure of the irregular reservoir grid are shown in Figure 7.14. In Figure 7.15 we
only show the active cells of the reservoir grid (cells that contribute to fluid flow).
Well logs are shown in Figure 7.16, whereas the rock physics diagnostics are shown in
Figure 7.17. The applied rock physics model is the contact cement model (see Avseth
et al. (2001)). Time-laspe seismic data are shown in Figures 7.18 and 7.19.
2750
2800
Depth (m)
2850
2900
Porosity
2950
Water saturation
2000 4000 6000 1000 2000 3000 1 2 3 0 0.5 1 0 0.5 1

P−wave velocity (m/s) S−wave velocity (m/s) 3 Clay content (v/v) Porosity and saturation (v/v)
Density (g/cm )
Figure 7.16: Well log data (well E2): from left to right P-wave velocity, S-wave
velocity, density, clay content, porosity and water saturation.
color = clay content color = clay content

4600 3000
0.9 0.9
4400
0.8 0.8
4200
0.7 0.7
P−wave velocity (m/s)
S−wave velocity (m/s)

4000 2500
0.6 0.6
3800
0.5 0.5
3600
0.4 0.4
3400 0.3 2000 0.3
3200 0.2 0.2
3000 0.1 0.1
2800 1500
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
Porosity (v/v) Porosity (v/v)
Figure 7.17: Well log data (well E2): crossplot of P-wave velocity versus porosity
(left) and S-wave velocity versus porosity (right) color coded by clay content.
Figure 7.18: Seismic data differences in 2003: from left to right, near, mid and far.
Figure 7.19: Seismic data differences in 2006: from left to right, near, mid and far.
Figure 7.20: Predicted gas saturation in 2003.
Figure 7.21: Predicted gas saturation in 2006.

Figure 7.22: Predicted fluid pressure in 2003.
Figure 7.23: Predicted fluid pressure in 2006.

For this application we did not invert simultaneously for static and dynamic prop-
erties but only for dynamic property changes since the initial reservoir model was
already provided in the dataset and obtained with a similar technique. We then per-
formed the Bayesian time-lapse inversion of time-lapse seismic differences for changes
in pressure and saturation. The inversion was performed in time-domain on a reg-
ular grid. We then converted the data from time to depth using the velocity model
provided with the dataset and then we mapped (and upscaled) the regular seismic
grid into the geocellular model using 3-dimensional interpolation. We assumed that
the initial model for pressure and saturation was known and we added the predicted
changes to obtain estimations of dynamic properties at the corresponding time-steps
(2003, 2004, and 2006). To avoid non-physical results, such as saturations greater
than 1, we applied cut-off values to preserve physical ranges. In Figures 7.20 and
7.21 we show the predictions for gas saturation in 2003 and 2006 where we observe a
decrease of gas saturation in the upper part of the reservoir and a shift in the gas-oil
contact. In Figures 7.22 and 7.23 we show the corresponding predictions for pore
pressure, where we observe two different pressure behaviors in two different regions
of the reservoir.
7.6 Conclusion
We presented a Bayesian inversion approach for assessing reservoir properties and
their changes from time-lapse data. The simultaneous inversion of base seismic sur-
vey and tile-lapse seismic differences accounts for the correlation between static prop-
erties (e.g. porosity, net-to-gross) and changes in dynamic properties (saturations
and pressure), and the full Bayesian approach allows us to quantify the associated
uncertainty. This approach is different from the usual one where 3D seismic data is
used to first characterize static reservoir properties and then time-lapse seismic data
is used to estimate saturation and pressure changes. Here, instead we make use of the
rock physics relations between pressure/saturation changes and porosity/lithology, to
jointly invert for static and dynamic properties. The use of non-parametric distri-
butions allows us to describe the non-Gaussian behavior of dynamic properties, in
particular saturations. The mean, median or most probable values of the estimated
distributions of reservoir properties can be used in seismic history matching to im-
prove reservoir model predictions and assess their uncertainty. The application to a
real reservoir model with synthetic data showed promising results.
Chapter 8
Final remarks
The main achievement of this thesis consists of a set of probabilistic methodologies to

estimate rock properties in the subsurface. All the methods presented in this thesis
combine seismic inversion with rock physics modeling and geostatistics. Rock physics
models allow us to study the relations between rock properties and elastic attributes,
specifically the physical models that link porosity, lithology and fluid properties with
velocities of elastic waves. However, spatial variations and inter-dependence between
different attributes are complex and cannot be described using simple deterministic
functions. Geostatistics provides a framework to represent complex reservoir het-
erogeneities and model complicated relationships between reservoir attributes and is
particularly useful when we have incomplete data information about the reservoir.
The first Bayesian inversion method presented in Chapter 2 is particularly suitable
in exploration phases where few data are available, whereas the inversion methods
presented in Chapters 3 and 4 can be applied in advanced stages where the goal of the
study is to obtain high detailed reservoir models. Similar methodologies have been
applied in the past to elastic inversion problems, the introduction of rock physics
modeling and statistical relations allows us to link the results of elastic inversion with
static and dynamic reservoir models used in fluid flow simulators. Another statistical
methodology is presented in Chapter 5 to account for the uncertainty of well log data
and integrate formation evaluation results with rock physics relations. Finally, for a
198
CHAPTER 8. FINAL REMARKS 199
full integration of geophysics and reservoir engineering, a Bayesian inversion of time-

lapse seismic data is presented in Chapter 7, in order to estimate not only the initial
static reservoir model, but also changes in dynamic properties, such as pressure and
saturation. This goal is achieved by integrating seismic inversion techniques with a
rock physics model describing pressure-velocity relations in different reservoir facies
as shown in Chapter 6.
Most of the methods presented in this thesis have been applied to clastic reservoirs.
A possible research direction for the future is the application of these methodologies
to unconventional reservoirs, in particular CO2 sequestration problems. A second
important future development focuses on the integration of the inversion results in
the reservoir modeling workflow. Time-lapse inversion results could be integrated in
seismic history matching to improve the reservoir description and reduce the uncer-
tainty. A third innovative research topic is the integration of other types of data and
disciplines, such as electromagnetic data and geomechanical models.
Bibliography
Aanonsen, S.I., I. Aavatsmark, T. Barkve, A. Cominelli, R. Gonard, O. Gosselin, M.

Kolasin- ski,and H. Reme, H., 2003, Effect of scale dependent data correlations in
an integrated history matching loop combining production data and 4D seismic
data: Proceedings of SPE Reservoir Simulation Symposium, SPE 79665.
Aki, K., and P. G. Richards, 1980, Quantitative seismology: W. H. Freeman & Co.
Alspach, D. L., and H. W. Sorenson, 1972, Nonlinear Bayesian estimation using

Gaussian sum approximation: IEEE Transaction of Automata Control, 17, 439–
448.
Avseth, P., T. Mukerji, A. Jørstad,, G. Mavko, and T. Veggeland, 2001, Seismic

reservoir mapping from 3-D AVO in a North Sea turbidite system: Geophysics, 66,
1157–1176.
Avseth, P., T. Mukerji, and G. Mavko, 2005, Quantitative seismic interpretation:

Cambridge University Press.
Avseth, P., N. Skjei, and Å. Skålnes, 2012, Rock physics modelling of 4D time-
shifts and time-shift derivatives using well log data a North Sea demonstration:
Geophysical Prospecting, 1–11.
Bachrach, R., 2006, Joint estimation of porosity and saturation using stochastic rock
physics modeling: Geophysics, 71, O53–O63.
Bachrach, R., and P. Avseth, 2008, Rock physics modeling of unconsolidated sands:
Accounting for nonuniform contacts and heterogeneous stress fields in the effective
200
BIBLIOGRAPHY 201
media approximation with applications to hydrocarbon exploration: Geophysics,

73, E197–E209.
Backus, G. E., 1962, Long-wave elastic anisotropy produced by horizontal layering:

Journal of Geophysical Research, 67, 4427–4440.
Batzle, M. L., and Z. Wang, 1992, Seismic properties of pore fluids: Geophysics, 57,
1396–1408.
Bornard, R., F. Allo, T. Coleou, Y. Freudenreich, D. H. Caldwell, J. G. Hamman,

2005, Petro- physical seismic inversion to determine more accurate and precise
reservoir properties: Proceedingf of SPE EUROPEC Conference, SPE 94144.
Bortoli, L. J., F. Alabert, A. Haas, and A. Journel, 1992, Constraining stochastic

images to seismic data: Stochastic simulation of synthetic seismograms, in Soares,
A. Ed., Geostatistics Troia 1992, Kluwer Academic Publishers, 325337.
Bosch, M., C. Carvajal, J. Rodrigues, A. Torres, M. Aldana, and J. Sierra, 2009,

Petrophysical seismic inversion conditioned to well-log data: Methods and applica-
tion to a gas reservoir: Geophysics, 74 (2), O1–O15.
Bosch, M., T. Mukerji, and E. F. González, 2010, Seismic inversion for reservoir prop-
erties combining statistical rock physics and geostatistics: A review: Geophysics,
75 (5), 75A165–75A176.
Bourbie, T., O. Coussy, and B. Zinszner, 1988, Acoustics of porous media: CRC
Press.
Buland, A., and H. Omre, 2003, Bayesian linearized AVO inversion: Geophysics, 68,
185–198.
Buland, A., and Y. El Ouair, 2006, Bayesian time-lapse inversion: Geophysics, 71

(3), R43–R48.
Buland, A., O. Kolbjørnsen, R. Hauge, O. Skjæveland, and K. Duffaut, 2008,

Bayesian lithology and fluid prediction from seismic prestack data: Geophysics,
73 (3), C13–C21.
BIBLIOGRAPHY 202
Caers, J., and T. B. Hoffman, 2006, The Probability Perturbation Method: A New
Look at Bayesian Inverse Modeling: Mathematical Geology, 38 (1), 81–100.
Castro, S. A., J. Caers, C. Otterlei, H. Meisingset, T. Hoye, P. Gomel, and E. Zachari-

assen, 2009, Incorporating 4D seismic data into reservoir models while honoring
production and geologic data: A case: The Leading Edge, 28, 1498.
Christensen, N. and H. F. Wang, 1985, The influence of pore pressure and confining
pressure on dynamic elastic properties of Berea sandstone: Geophysics, 28, 207-213
Coleou, T., F. Allo, R. Bornard, J. G. Hamman, D.H. Caldwel, 2005 Petrophysical

seismic inversion: SEG Expanded Abstract, 1355–1359.
Dadashpour, M., M. Landrø, and J. Kleppe, 2008, Nonlinear inversion for estimat-
ing reservoir parameters from time-lapse seismic data: Journal of Geophysical
Engineering, 5, 54-66.
Darling, T., 2005, Well Logging and Formation Evaluation: Elsevier Inc.
Dempster, A. P., N. M. Laird, D. B. Rubin, 1977, Maximum likelihood from incom-

plete data via the EM algorithm: Journal of the Royal Statistical Society, Series B
(Methodological), 39 (1), 1–38.
Deutsch, C., and A. G. Journel, 1992 GSLIB: Geostatistical software library and
user’s guide: Oxford University Press.
Dong, Y., and D. S. Oliver, 2005, Quantitative use of 4D seismic data for reservoir
description: SPE Journal 10 (1), 91-99.
Dovera, L., and E. Della Rossa, 2011, Multimodal ensemble Kalman filtering using
Gaussian mixture models: Computational Geosciences, 15 (2), 307–323.
Doyen, P., 1988, Porosity from seismic data: A geostatistical approach: Geophysics,
53, 1263-1275.
Doyen, P., 2007, Seismic reservoir characterization: European Association of Geosci-

entists & Engineers.
BIBLIOGRAPHY 203
Dubrule, O., 2003, Geostatisics for seismic data integration in earth models:
SEG/EAGE Distinguished Instructor Series.
Dutta, N. C., 2002, Geopressure prediction using seismic data: current status and
the road ahead: Geophysics, 67, 20122041.
Dvorkin, J., A. Nur, and H. Yin, 1994, Effective Properties of Cemented Granular
Materials: Mechanics of Materials, 18, 351–366.
Dvorkin, J., and A. Nur, 1996, Elasticity of high-porosity sandstones: Theory for two
North Sea data sets: Geophysics, 61, 1363–1370.
Dvorkin, J., A. Nur, and C. Chaika, 1996, Stress sensitivity of sandstones: Geophysics,
61, 444–455.
Dvorkin, J., G. Mavko, and B. Gurevich, 2007, Fluid substitution in shaley sediment
using effective porosity: Geophysics, 72, O1–O8.
Dvorkin, 2007, Self-similarity in rock physics: The Leading Edge, 26, 946–950.
Eberhart-Phillips, D., D. H. Han, and M. D. Zoback, 1989, Empirical relationships

among seismic velocity, effective pressure, porosity, and clay content in sandstone:
Geophysics, 51, 82–89.
Eidsvik, J., P. Avseth, H. Omre, T. Mukerji, and G. Mavko, 2004a, Stochastic reser-
voir characterization using prestack seismic data: Geophysics, 69, 978–993.
Eidsvik, J., T. Mukerji, and P. Switzer, 2004b, Estimation of geological attributes

from a well log: an application of hidden Markov chains: Mathematical Geology,
36, 379–397.
Ellis, D. V., and J. M. Singer, 2007, Well logging for earth scientists: Springer.
Fylling, A., 2002, Quantification of petrophysical uncertainty and its effect on in-
place volume estimates: numerous challenges and some solutions: Proceedings of
SPE Annual Technical Conference and Exhibition, SPE 77637.
BIBLIOGRAPHY 204
Gal, D., J. Dvorkin, and A. Nur, 1998, A physical model for porosity reduction in
sandstones: Geophysics, 63, 454–459.
Gallop, J., 2006, Facies probability from mixture distributions with non-stationary
impedance errors: 76th Annual International Meeting, SEG Expanded Abstract,
1801–1805.
Gilardi, N., S. Bengio, and M. Kanevski, 2002, Conditional Gaussian mixture mod-
els for environmental risk mapping: Proceedings of IEEE Workshop on Neural
Networks for Signal Processing, 777–786.
González, E. F., T. Mukerji, and G. Mavko, 2008, Seismic inversion combining rock
physics and multiple-point geostatistics: Geophysics, 73, R11–R21.
Goovaerts, P., 1997, Geostatistics for Natural Resources Evaluation: Oxford Univer-
sity Press.
Grana, D., E. Della Rossa, and C. D’Agosto, 2009, Petrophysical properties estima-
tion in a crosswell study integrated with statistical rock physics: SEG Expanded
Abstracts, 28, 1805–1809, doi: 10.1190/1.3255203.
Grana, D., and E. Della Rossa, 2010, Probabilistic petrophysical-properties estima-

tion integrating statistical rock physics with seismic inversion: Geophysics, 75 (3),
O21–O37, doi: 10.1190/1.3386676.
Grana, D., and J. Dvorkin, 2011, The link between seismic inversion, rock physics,
and geostatistical simulations in seismic reservoir characterization studies: The
Leading Edge, 30 (1), 54–61, doi: 10.1190/1.3535433.
Grana, D., M. Pirrone, and T. Mukerji, 2012a, Quantitative log interpretation and un-
certainty propagation of petrophysical properties and facies classification from rock-
physics modeling and formation evaluation analysis: Geophysics, 77 (3), WA45–
WA63, doi: 10.1190/geo2011-0272.1.
BIBLIOGRAPHY 205
Grana, D., T. Mukerji, J. Dvorkin, and G. Mavko, 2012b, Stochastic inversion of facies
from seismic data based on sequential simulations and probability perturbation
method: Geophysics, 77 (4), M53–M72, doi: 10.1190/geo2011-0417.1.
Grana, D., T. Mukerji, L. Dovera, and E. Della Rossa, 2012, Sequential Simulations of
Mixed Discrete-Continuous Properties: Sequential Gaussian Mixture Simulation:
in P. Abrahamsen et al. (eds.), Geostatistics Oslo 2012, Quantitative Geology and
Geostatistics 17, doi: 10.1007/978-94-007-4153-9-19.
Grana, D., T. Mukerji, 2012, Sequential Bayesian Gaussian mixture linear inversion
of seismic data for elastic and reservoir properties estimation: SEG Expanded
Abstract, 1–5, doi: 10.1190/segam2012-0095.1.
Grana, D., E. Paparozzi, S. Mancini, and C. Tarchiani, 2013, Seismic driven proba-
bilistic classification of reservoir facies for static reservoir modelling: a case history
in the Barents Sea: Geophysical Prospecting, 61 (3), 613–629, doi: 10.1111/j.1365-
2478.2012.01115.x.
Gunning, J., and M. Glinsky, 2007, Detection of reservoir quality using Bayesian
seismic inversion: Geophysics, 72, R37–R49.
Gutierrez, M. A., N. R. Braunsdorf, and B. A. Couzens, 2006, Calibration and ranking

of pore-pressure prediction models: The Leading Edge, 25, 15161523.
Haas, A. and O. Dubrule, 1994, Geostatistical inversion - A sequential method of

stochastic reservoir modeling constrained by seismic data: First Break, 12, 561–
569.
Han, D., 1986, Effects of porosity and clay content on acoustic properties of sandstones
and unconsolidated sediments: PhD thesis, Stanford University.
Hansen, T. M., A. G. Journel, A. Tarantola, and K. Mosegaard, 2006, Linear inverse

Gaussian theory and geostatistic: Geophysics, 71, R101–R111.
BIBLIOGRAPHY 206
Hansen, T. M., K. S. Cordua, and K. Mosegaard, 2012, Inverse problems with non-
trivial priors: efficient solution through Sequential Gibbs Sampling: Computational
Geosciences, 1–19.
Hasselblad, V., 1966, Estimation of Parameters for a Mixture of Normal Distributions:

Technometrics, 8 (3), 431–444.
Hastie, T., and R. Tibshirani, 1996, Discriminant analysis by gaussian mixtures:

Journal of Royal Statistical Society, Series B, 58 (1), 155–176.
Hastie, T., R. Tibshirani, and J. Friedmann, 2002, The elements of statistical learning:
Springer.
Heidari, Z., C. Torres-Verdin, and W. E. Preeg, 2010, Improved estimation of min-

eral and fluid volumetric concentrations in thinly-bedded and invaded formations,
Transactions of the SPWLA 51st Annual Symposium.
Hill, T., and P. Lewicki, 2006, Statistics. Methods and Applications: StatSoft Inc.
Huang, X., L. Meister, and R. Workman, 1997, Reservoir characterization by inte-

gration of time-lapse seismic and production data: Proceedings of SPE Annual
Technical Conference and Exhibition, SPE 38695.
Jackson, D. D., 1972, Interpretation of Inaccurate Insufficient and Inconsistent Data,

Geophysical Journal of the Royal Astronomical Society, 28, 97–109.
Jackson, D. D., 1979, The use of a priori data to resolve non-uniqueness in linear
inversion, Geophysical Journal of the Royal Astronomical Society, 57, 137–157.
Journel, A. G., and J. J. Gomez-Hernandez, 1989, Stochastic imaging of the Wilm-

ington clastic sequence: SPE Formation Evaluation, 8 (1), 33–40, SPE 19857-PA.
Journel, A. G., 2002, Combining knowledge from diverse sources: an alternative to

traditional conditional independence hypothesis: Mathematical Geology, 34 (5),
573–596.
BIBLIOGRAPHY 207
Kaufman, L., and P. J. Rousseeuw, 1990, Finding groups in data. An introduction to

cluster analysis: Wiley series in probability and statistics.
Kennedy, J., P., Pujiyono, A. Cox, and R. Aldred, 2010, Using quantified model based
petrophysical uncertainty to aid in conflict resolution: Transactions of the SPWLA
51st Annual Symposium.
Kretz, V., M. le Ravalec-Dupin, and F. Roggero, 2004, An integrated reservoir char-

acterization study matching production data and 4D seismic: SPE Reservoir Eval-
uation & Engineering 7 (2), 116122.
Krishnan, S., 2008, The Tau Model for Data Redundancy and Information Combi-
nation in Earth Sciences: Theory and Application, Mathematical Geology, 40 (6),
705–727.
Krumbein, W. C., and M. F. Dacey, 1969. Markov Chains and Embedded Markov
Chains in Geology: Mathematical Geology, 1 (1), 79–96.
Lake, L. W., and S. Srinivasan, 2004, Statistical scale-up of reservoir properties:

concepts and applications: Journal of Petroleum Science & Engineering, 44, 27–
39.
Landa, J. L., and R. N.Horne, 1997, A procedure to integrate well test data, reser-
voir performance history and 4-D seismic information into a reservoir description:
Proceedings of SPE Annual Technical Conference and Exhibition, SPE 38653.
Landrø, M., 2001, Discrimination between pressure and fluid saturation changes from
time-lapse seismic data: Geophysics, 66, 836844.
Landrø, M., and Ø. Kvam, 2002, Pore pressure estimation - What can we learn from
4D?: CSEG Recorder.
Landrø, M., H. H. Veire, K. Duffaut, and N. Najjar, 2003, Discrimination between

pressure and fluid saturation changes from marine multicomponent time-lapse seis-
mic data: Geophysics, 68, 15921599.
BIBLIOGRAPHY 208
Larsen, A. L., M. Ulvmoen, H. Omre, and A. Buland, 2006, Bayesian lithology/fluid

prediction and simulation on the basis of a markov-chain prior model: Geophysics,
71, R69–R78.
Liu, Y., 2003, Downscaling seismic data into a geologically sound numerical model:
PhD thesis, Stanford University.
MacBeth, C., 2004, A classification for the pressure-sensitivity properties of a sand-

stone rock frame: Geophysics 69, 497–510.
Mavko, G., and T. Mukerji, 1998, A rock physics strategy for quantifying uncertainty
in common hydrocarbon indicators, Geophysics 63, 1997.
Mavko, G., T. Mukerji, and J. Dvorkin, 2009, The rock physics handbook: Cambridge
University Press.
Merletti, G. and C. Torres-Verdin, 2006, Accurate detection and spatial delineation of

thin-sand sedimentary sequences via joint stochastic inversion of well logs and 3D
pre-stack seismic amplitude data: Proceedings of SPE Annual Technical Conference
and Exhibition, SPE 102444-PA.
Mindlin, R. D., 1949, Compliance of elastic bodies in contact: Applied Mechanics,

16, 259-268
Mosegaard K., and A. Tarantola, 1995, Monte Carlo sampling of solutions to inverse
problems: Journal of Geophysical Research, 100, 12431–12447.
Mukerji, T., A. Jørstad, P. Avseth, G. Mavko, and J. R. Granli, 2001a, Mapping

lithofacies and pore-fluid probabilities in a North Sea reservoir: Seismic inversions
and statistical rock physics: Geophysics, 66, 988–1001.
Mukerji, T., P. Avseth, G. Mavko, I. Takahashi, and E. F. González, 2001b, Statistical

rock physics: Combining rock physics, information theory, and geostatistics to
reduce uncertainty in seismic reservoir characterization: The Leading Edge, 20,
313–319.
BIBLIOGRAPHY 209
Nur, A., and Z. Wang, 1989, Seismic and acoustic velocities in reservoir rocks, Volume
1, Experimental studies: SEG Geophysics reprint series.
Oliver, D., 2008, Inverse Theory for Petroleum Reservoir Characterization and History
Matching: Cambridge University Press.
Papoulis, A., 1984, Probability, random variables and stochastic processes: McGraw-
Hill.
Poupon, A., and J. Leveaux, 1971, Evaluation of water saturation in shaly formations,
Transactions of the SPWLA 12th Annual Symposium.
Remy, N., A. Boucher, and J. Wu, 2009, Applied geostatistics with SGeMS: Cam-
bridge University Press.
Reynolds D. A., T. F. Quatieri, and R. B. Dunn, 2000, Speaker Verification Using

Adapted Gaussian Mixture Models: Digital Signal Processing, 10 1-3, 19–41.
Rimstad, K., and H. Omre, 2010, Impact of rock-physics depth trends and Markov
random fields on hierarchical Bayesian lithology/fluid prediction: Geophysics,
75(4), R93–R108.
Sams, M. S., D. Atkins, N. Said, E. Parwito and P. van Riel, 1999, Stochastic in-
version for high resolution reservoir characterisation in the Central Sumatra Basin:
Proceedings of SPE Annual Technical Conference and Exhibition, SPE 57260.
Sams M. S., I. Millar, W. Satriawan, D. Saussus and S. Bhattacharyya, 2011, Inte-

gration of geology and geophysics through geostatistical inversion: A case study:
First Break, 29 (8), 4756.
Saul, M., and Lumley D., 2012, A new velocity-pressure-compaction model for unce-
mented sediments: SEG Expanded Abstracts, 1-5.
Sayers, C. M., T. J. H. Smith, C. van Eden, R. Wervelman, T. Fitts, J. Bingham,

K. McLachlan, P. Hooyman, S. Noeth, and D. Mardhiri, 2003, Use of reflection to-
mography to predict pore pressure in overpressured reservoir sands: SEG Expanded
Abstracts, 22, 13621365.
BIBLIOGRAPHY 210
Sengupta, M., and R. Bachrach, 2007, Uncertainty in seismic-based pay volume es-
timation: Analysis using rock physics and bayesian statistics: The Leading Edge,
24, 184–189.
Shannon, C. E., 1948, A mathematical theory of communication: Bell System Tech-

nical Journal, 27, 379–423.
Silverman, B. W., 1986, Density estimation for statistics and data analysis. mono-
graphs on statistics and applied probability: Chapman & Hall.
Spikes, K., T. Mukerji, J. Dvorkin, and G. Mavko, 2008, Probabilistic seismic inver-
sion based on rock physics models: Geophysics, 72, R87–R97.
Stolt, R. H., and A. B. Weglein, 1985, Migration and inversion of seismic data: Geo-
physics, 50 24582472.
Stright, L., A. Bernhardt, A. Boucher, T. Mukerji, and R. Derksen, 2009, Revisiting

the use of seismic attributes as soft data for subseismic facies prediction: Proportion
versus probabilities: The Leading Edge, 28 (12), 1460–1469.
Sung H. G., 2004, Gaussian Mixture Regression and Classification: PhD Thesis, Rice
University.
Tarantola, A., 2005, Inverse problem theory: SIAM.
Terzaghi, K., 1943, Theoretical Soil Mechanics: John Wiley & Sons.
Theys, P., 1991, Log data acquisition and quality control, Editions Technip.
Theys, P., 1994, A serious look at repeated sections, Transactions of the SPWLA
30th Annual Symposium.
Theys, P., 1997, Accuracy. Essential information for a log measurement, Transactions
of the SPWLA 38th Annual Symposium.
Trani, M., R. Arts, O. Leeuwenburgh, and J. Brouwer, 2011, Estimation of changes in

saturation and pressure from 4D seismic AVO and time-shift analysis: Geophysics,
76, 1-17.
BIBLIOGRAPHY 211
Verga, F., D. Viberti, and M. Gonfalini, 2002, Uncertainty evaluation in well log-
ging: analytical or numerical approach?, Transactions of the SPWLA 43rd Annual
Symposium.
Viberti, D., 2010, A rigorous mathematical approach for petrophysical estimation,

Americal Journal of Applied Sciences, 7 (11), 1509–1516.
Ulvmoen, M., and H. Omre, 2010, Improved resolution in Bayesian lithology/fluid

inversion from prestack seismic data and well observations: Part 1 - Methodology:
Geophysics, 75 (2), R21-R35.
Wang, Z., and A. Nur, 1992, Seismic and acoustic velocities in reservoir rocks, Volume
2, Theoretical and model studies: SEG Geophysics reprint series.
Wang, Z., and A. Nur, 2000, Seismic and acoustic velocities in reservoir rocks, Volume
3, Recent development: SEG Geophysics reprint series.
Ward, J. H. Jr., 1963, Hierarchical grouping to optimize an objective function, Journal

of American Statistical Association, 58, 236–244.
Wilschut, F., E. Peters, K. Visser, P. A. Fokker, and P. M. E. van Hooff, 2011, Joint
History Matching of Well Data and Surface Subsidence Observations Using the
Ensemble Kalman Filter: A Field Study: Proceedings of SPE Reservoir Simulation
Symposium, 141690-MS.

SRB 136 Aug13 Grana

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

SRB 136 Aug13 Grana

Enviado por

Direitos autorais:

Formatos disponíveis

BAYESIAN INVERSION METHODS

FOR SEISMIC RESERVOIR CHARACTERIZATION

(Gary Mavko) Principal Adviser

I certify that I have read this dissertation and that, in my opinion, it

I certify that I have read this dissertation and that, in my opinion, it

Approved for the University Committee on Graduate Studies

This dissertation addresses mathematical methodologies for seismic reservoir charac-

2 Probabilistic reservoir properties estimation 10

4 Stochastic inversion for facies modeling 67

5 Statistical methods for log evaluation 113

6 Pressure dependence of elastic properties 159

7 Bayesian time-lapse inversion 173

8 Final remarks 198

5.1 Standard deviations associated to log measurements. . . . . . . . . . 140

2.1 Flowchart of the probabilistic petrophysical properties estimation. . . 13

4.1 Workflow of stochastic inversion. . . . . . . . . . . . . . . . . . . . . 72

5.1 Flowchart of the probabilistic petrophysical properties estimation. . . 118

1.1 Motivation and objectives

1.2 Chapter description

input seismic dataset. This method is based on a stochastic optimization technique,

1.3 Inversion methods

1.4 Bayesian approach

Probabilistic reservoir properties

petrophysical variables and litho-fluid classes in a multiproperties, multiscale model.

Prestack seismic data

Well data &

Petrophysics R (ĭ, C, sw)

Litho fluid class

Figure 2.1: Flowchart of the probabilistic petrophysical properties estimation.

2.3 Probabilistic approach to petrophysical inver-

2.3.1 Statistical rock physics modeling

m = fRP M (R) + ε (2.1)

this assumption we can state that

P (m|R) = N(m; µ(R), Σε ) (2.3)

with mean and covariance of each component given by

where λk are the weights of the conditional distribution:

πk N(m; µkm , Σkm )

for each given m.

2.3.2 Probabilistic upscaling

where n is the dimension of mf (n = 2 in the case of Ip and Is ). This expression is

2.3.3 From seismic to petrophysics inversion

µℓ|S = µℓ + ΣTS,ℓ (ΣS )−1 (S − µS ) (2.20)

2.4 Real case application

2.4.1 Methodology implementation

[Vp , Vs , ρ] = fRP M (φ, sw, C) + ε, (2.25)

f). This last conditional probability is then combined by means of Chapman-Kolmogorov

at each vertical position z.

and we integrate it with a Markov chain model by means of Eq. 2.24.

2.4.2 Data application

P−impedance (m/s g/cm3)

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35

ρ (g/cm3 ) K (GP a) µ (GP a)

We then generate pseudo petrophysical curves in two steps. We firstly generate

0.7 0.7 2.5

0.2 0.2 0.5

Figure 2.5: Prior distribution of petrophysical variables. (top) 2D marginal distribu-

4.5 4.5 4.5

3.5 3.5 3.5

2.5 2.5 2.5

1.5 1.5 1.5

0.5 0.5 0.5

Stiff sand model

µsat = µdry . (2.38)