Computers and Geotechnics 93 (2018) 87–103

Contents lists available at ScienceDirect

Computers and Geotechnics

journal homepage: www.elsevier.com/locate/compgeo

Research Paper

Settlement prediction for an embankment on soft clay

Nina Müthing ⇑, Chenyang Zhao, Raoul Hölter, Tom Schanz
Ruhr-Universität Bochum, Department of Civil and Environmental Engineering, Universitätsstr. 150, 44780 Bochum, Germany

a r t i c l e i n f o a b s t r a c t

Article history: A model prediction using FE modeling is performed for a trial embankment on soft clay in Ballina,
Received 23 January 2017 Australia. The comparison between the prediction and the in situ measurement exhibits significant dif-
Received in revised form 13 May 2017 ferences. A detailed analysis is performed to validate the model and determine necessary improvements.
Accepted 6 June 2017
Sensitivity analyses elaborate upon the most dominant constitutive parameters. Based on site measure-
ments, inverse analyses are performed to identify the optimum parameters. The inverse analysis
approach is validated regarding its dependency on the objective function and data incorporated. A com-
Keywords:
parison between the prediction and the back-calculated parameters confirms the importance of engi-
Embankment
Soft soil
neering judgment as well as the sensitivity of the modeling strategies to the input information.
Consolidation Ó 2017 Elsevier Ltd. All rights reserved.
Finite element modeling
Parametric study
Inverse analysis

1. Introduction and further prediction aspects, e.g., the horizontal displacements,

In Ballina, New South Wales, Australia, a full-scale test embank-
ment was constructed to serve as a National Field Testing Facility
(NFTF). It was implemented to study the settlement behavior of
soft, structured soils typically present along the east coast of Aus-
tralia. Therefore, the test site was equipped with extensive and ver-
satile in situ testing instrumentation, including that for the
measurement of pore pressures, vertical and horizontal deforma-
tions and soil pressures. Moreover, high-quality soil samples were
extracted and subjected to advanced laboratory testing at the Aus-
tralian Research Council Centre of Excellence for Geotechnical
Science and Engineering (CGSE) in Newcastle.
For the Embankment Prediction Symposium 2016 held by CGSE in
September 2016, asset owners, academic researchers and practi-
tioners were asked to submit predictions for the embankment
behavior obtained under different types of analyses, ranging from
hand calculations to complex 3D finite element analyses. The pre-
dictions were performed based on field and laboratory data pro-
vided by CGSE, summarized in [1].
Section 2 introduces the 2D finite element model used within
the present study, which is implemented using Plaxis 2D (version
2015). The model simulates the consolidation and creep processes
underneath the embankment on soft, estuarine clay and predicts
the embankment behavior in terms of transient settlements and
pore water pressure dissipation. More details on the model itself
⇑ Corresponding author.
E-mail address: nina.muething@rub.de (N. Müthing).
can be found in the prediction report [2].
To minimize the computational effort, sophisticated modeling
strategies are used: the geometrical conditions are reduced to an
equivalent 2D model, and the installed vertical drains are modeled
by a soil volume of equivalent permeability. A 2D/3D unit cell anal-
ysis validates the model regarding its dimensionality and ability to
adequately simulate the drainage behavior of the vertical drains.
Since the subsoil underneath the embankment includes layers
of natural, structured clay, particular focus is set on the choice of
a simple but adequate constitutive model as well as on the deter-
mination of appropriate constitutive parameters [3]. A validation
of the constitutive modeling strategy proves the eligibility of this
approach to simulate the complex compression behavior of struc-
tured clay with a less complex constitutive law.
Section 3 validates the prediction given in [2] by comparing its
results in terms of transient settlements and pore water dissipation
with the observed field response provided by CGSE. The compar-
ison reveals significant differences between the prediction and
real-life behavior of the embankment.
A detailed analysis is performed to validate the prediction
model and reveal necessary improvements. Within the present
study, the focus is set on the constitutive modeling strategies
and the determination of the constitutive parameters to obtain a
better understanding of the complex behavior of soft clays.
Fig. 1 illustrates the strategic approach for the model validation
and prediction improvement followed within the present study.
To improve the accuracy and reliability of a prediction, it is
essential to know which parameters most strongly influence the

http://dx.doi.org/10.1016/j.compgeo.2017.06.002
0266-352X/Ó 2017 Elsevier Ltd. All rights reserved.
88 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103

Measurements Numerical modeling approach Constitutive parameters

Geometry Physical
Settlements processes Lab data Field data
conditions

Pore water Staged Constitutive Initial

pressures construction assumptions parameters
Sensitivity
analysis

Relevant
Calculation
parameters
model
Model Inverse
validation analysis
Calibrated
parameters
Agreement? No
Parameter Uncertainty
validation analysis
Model Output
Yes response uncertainty

Fig. 1. Flowchart of the conceptual approach followed in this research.

model response. Therefore, in a first step, a parametric study using
metamodeling strategies to perform global sensitivity analyses is
used to elaborate upon the most relevant constitutive parameters.
However, it has to be considered that the predominance of certain
model parameters might be caused by model assumptions. A close
evaluation of this enables the assessment of the method and high-
lights what has to be considered when conducting this type of
analysis.
With the knowledge of the relevant model parameters, inverse
analyses using a Generic Algorithm (GA) for parameter identifica-
tion can be performed. The aim of this inverse analysis is to iden-
tify an optimum parameter set to achieve a good fit between the
predicted and measured values. This can be particularly useful
when only limited experimental data on the soil properties are
available and/or a sophisticated constitutive model requiring
advanced laboratory tests is applied. Of course, the inverse analysis
itself and its output are strongly dependent on the input informa-
tion provided for the parameter optimization. Here, particularly,
the influence of the chosen objective function for the optimization
in terms of type, amount and quality of data incorporated needs to
be evaluated. Thus, a step-wise inverse analysis approach consid-
ering different objective functions is preferred and executed in
the present study. This strategy allows for the assessment of the
tool of inverse analysis regarding its dependency on the choice of
objective function and data incorporated. Consequently, the
importance of engineering judgment and the sensitivity of the
method to input information can be illustrated.
After identifying an optimum parameter set, a critical examina-
tion of the identified parameter values needs to be conducted.
Through a comparison of the initial and optimum parameter val-
ues, a physical justification needs to be achieved.
Moreover, the initial as well as final predictions are based on
parameter values that are not precisely known but that are mean
values with ranges estimated by engineering judgment. To observe
the impact of parameter uncertainty and identify the most domi-
nant parameter ranges on the model output, an uncertainty analy-
sis needs to be executed. A comparison of the model response
uncertainty between the initial and optimum parameters will help
provide a better understanding and assessment of the performed
improvement strategy.
2. Finite element model for embankment prediction
The test embankment is simulated by both a 2D plane strain
model and a 3D Finite-Element (FE) model. Plaxis 2D (version
2015) and Plaxis 3D (version AE) are used for conducting the 2D
and 3D simulations, respectively.
2.1. 2D FE model
The geometry and mesh discretization of the 2D FE-model are
shown in Fig. 2. The soil deposit present prior to embankment con-
struction is simplified into 6 horizontal layers. The working plat-
form and horizontal drainage layers are installed prior to the
embankment construction. A total of 22 wick drains with a spacing
of 1.2 m are used in this test embankment, and the drains are
installed up to 15 m beneath the ground surface. According to
the in situ investigation, the ground water level is assumed to be
0.72 m below the ground surface.
Based on the trial analysis, the mesh discretization is deter-
mined such that model responses are mesh independent. Specifi-
cally, Fig. 3 shows the relation between the model response and
the number of elements. The point that is located at the center of
the 2D model and 4 m below the surface is chosen to check the
excess pore pressure after the embankment construction is com-
pleted. The model responses using 14,100 elements are taken as
the reference value (42 kPa excess pore pressure) to calculate the
normalized value of the other cases. As seen, when the number
of elements increases from 14,100 to 18,200, the variation in the
model response is less than 1%; hence, models with 14,100 ele-
ments are used in this research, and the total number of nodes is
approximately 114,000.
The bottom of the model is assumed to be fixed, with no defor-
mation allowed. At the left and right boundaries, the horizontal
displacements are fixed, whereas vertical displacements are

1.05
Normalized excess pore pressure [-]
1.04
1.03
1.02
1.01
1 4.
0.99
4 6 8 10
Number of elements [×1000]
Fig. 3. Mesh sensitivity study; the excess pore pressures are normalized based on
the result using 14,100 elements (42 kPa).
allowed. The top of the model is free to deform in both the horizon-
tal and vertical direction. Regarding the hydraulic boundary condi-
tion, excess pore water pressure is allowed to dissipate through all
boundaries of the model.
2.2. 3D FE model
Accounting for the fact that the working platform layer is not
constructed symmetrically in reality, a three-dimensional FE model
is conducted in the present study for the validation of the 2D plane
strain model [4]. The geometry and mesh discretization of the 3D
model are presented in Fig. 4. The cross section on the right-hand
side of the embankment is identical to the 2D plane strain model.
For the cross section on the left-hand side of the embankment, the
working platform is wider compared to the 2D model.
In this 3D model, the total number of nodes and elements is
approximately 210,000 and 150,000,respectively. The mechanical
as well as hydraulic model boundary conditions are in analogy to
the 2D model.
2.3. Construction of the embankment
With reference to the report on the in situ situation provided by
CGSE, the sequential construction of the embankment is numeri-
N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103 89
Fig. 2. Geometry and FE discretization of the 2D model.
cally divided into five construction steps. All construction steps
are conducted as consolidation analysis to guarantee an adequate
modeling of the underlying physical processes. Then, a long-term
consolidation analysis is conducted, from which the prediction of
the soil deformation and pore water pressure dissipation is
derived. The detailed sequence of the numerical staged construc-
tion steps is given as follows:
1. Preparation of the working platform (3 days)
2. Short consolidation phase (15 days)
3. Preparation of the drainage layer (8 days)
Drainage installation (13 days)
5. Construction of the embankment/earth fill (10 days)
6. Long-term consolidation (about 1000 days)
12 14 16 18 20
2.4. Modeling of the drainage system
The test embankment was constructed using two different
types of vertical drains: conventional wick drains and biodegrad-
able jute drains. However, in the present study, for the sake of sim-
plicity, both drain systems were assumed to exhibit a similar
drainage behavior and thus were modeled in a uniform manner.
The measurement results from the in situ site show that the
embankment deformation is equivalent for both drainage areas.
Therefore, the assumption was confirmed to be acceptable.
During the installation of the wick drains, the soil in the near
field around the drains is disturbed and compacted in the horizon-
tal direction. This area is usually referred to as the smear zone.
Accordingly, the permeability of the soil in the smear zone is
reduced. In the present study, the ratio of the permeability (hori-
zontal direction in the 3D model) in the undisturbed soil and in
the smear zone is assumed to be approximately 5. The radius of
the smear zone is assumed to be five times the radius of the wick
drain [5].
Since there are more than 1500 wick drains installed beneath
the embankment, a full-scale 3D simulation including discretely
modeled drains would require a large computational effort. An
alternative approach is to apply the equivalent permeability con-
cept first proposed by Chai et al. [6]. This approach is a simple
approximation method for performing consolidation analysis of
PVD-improved soil without modeling the discrete drain elements.
Within this framework, the permeability (vertical direction) of the
soil is increased to an equivalent permeability, which accounts for
90 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103
Unit: m
30
Fig. 4. Geometry and FE discretization of the 3D model.
the soil and drain permeability together. Using this equivalent per-
meability, the variation in the degree of consolidation with time is
comparable to the behavior using discrete drainage elements. The
equivalent permeability kv e in the vertical direction is determined
using the following equations proposed in [6]:
2
!
2:5l kh
kv e ¼ 1þ kv ð1Þ
lD2e kv
n kh
2
2l kh
l ¼ ln þ lnðsÞ  0:75 þ p ð2Þ
s kv 3qw
where l denotes the drainage length; De is the diameter of the unit
cell; kh and kv are the initial horizontal and vertical permeability for
the undisturbed soil, respectively; n ¼ De =dw (dw = diameter of
drain); s ¼ ds =dw (ds = diameter of smear zone); and qw is the dis-
charge capacity of the PVD. It should be noted that when applying
this approach to two-way drainage conditions, the drain length
parameter l is half the thickness of the PVD-improved zone.
Under plane strain conditions, the drainage is usually modeled
as a plane sink. Since this does not capture the axisymmetric nat-
ure of horizontal flow toward drains, the soil permeability and
space between the drains under the plane strain conditions should
be modified. Indraratna and Redana [7] proposed three methods to
match the permeability under axisymmetric and plane strain con-
ditions: (1) Geometry matching, where the spacing between the
plane strain drains is varied while keeping the permeability con-
stant; (2) Permeability matching, where the permeability under
the plane strain condition is varied while keeping the spacing
between the drain constant; and (3) A combination of permeability
and geometry matching. In the present study, the permeability
matching method is adopted in the 2D modeling. The horizontal
permeability of the undisturbed soil under plane strain conditions
kh;ps is calculated according to [8]:
kh;ps 0:67ðn  1Þ2
¼ 2 ð3Þ
kh;ax n ðlnðnÞ  0:75Þ
where kh;ax denotes the horizontal permeability of the undisturbed
soil under axisymmetric conditions.
As mentioned before, the soil permeability in the smear zone
under the plane strain condition (ks;ps ) can be determined based
2D cross section
on that of the undisturbed soil (kh;ps ). Here, Indraratna and Redana
[8] proposed the following equations:
ks;ps b
¼ ð4Þ
kh;ps kh;ps ðlnðnÞ þ kh;ax lnðsÞ  0:75Þ  a
k h;ax s ks;ax
0:67ðn  sÞ3
a¼ ð5Þ
n2 ðn  1Þ
2ðn  1Þðnðn  s  1Þ þ 0:33ðs2 þ s þ 1ÞÞ
b¼ ð6Þ
n2 ðn  1Þ
Additionally, the permeability varied with the void ratio based
on the following equation:
 
k De
log ¼ ð7Þ
k0 ck
with k and k0 being the current and initial permeability, respec-
tively; De being the variation in the void ratio; and ck being the per-
meability change index, which is assumed to be 1.125 for the silt
and clay layers [5]. It should be noted that large deformations were
not considered in this study.
2.5. 2D and 3D unit cell analyses
To validate the aforementioned 2D and 3D modeling
approaches for the drainage, unit cell analyses were conducted
and are illustrated in the following section.
The 3D unit cell model is shown in Fig. 5(a). The length and
width of this model are identical and equal 1.2 m. The thickness
of each layer equals the thickness in the full-scale 3D model. The
drainage element is located in the center of the unit cell model.
The 2D unit cell model has the same soil layer distribution and
position of the drain as in the 3D model. The permeability values
for all layers and model types are given in Table 1. The numerical
staged construction steps and ground water level in the unit cell
analysis are the same as in the full-scale embankment analysis
(see Sections 2.1 and 2.3). The model bottom is fixed, and in-
plane deformations at the vertical boundaries of the unit cell
model are not allowed. Moreover, the excess pore pressure cannot
be dissipated through the vertical or bottom boundaries.
 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103 91

Fig. 5. Unit cell analyses: (a) model geometry (b) variation in the vertical settlements at point A (c) variation in the pore water pressure at point B (d) comparison of the pore
water pressure at points B and C in the 3D unit cell analysis.

Table 1
Values of permeability for different scenarios (Unit: m/day).

Model type Permeability Silt Clay 1 Clay 2 Clay 3 Sand

3D model kh;ax ¼ kv ;ax 6.40E3 9.68E3 2.40E3 2.28E3 4.32E1
ks;ax 1.28E3 1.94E3 4.80E4 4.57E4 8.64E2
2D model kh;pl ¼ kv ;pl 1.59E3 2.41E4 5.96E4 5.68E4 1.07E1
ks;pl 2.66E4 4.02E5 9.94E5 9.50E5 1.80E2
3D model equivalent case kh 6.40E3 9.68E3 2.40E3 2.28E3 4.32E1
kv e 4.50E2 6.86E3 1.70E2 1.62E2 1.95

Three numerical scenarios are considered in both the 2D and 3D
unit cell analyses: In scenario 1, only drainage is modeled, while
the disturbance due to the drain installation and its effect on the
soil permeability is not considered. In scenario 2, the impact of
the smear zone is modeled using a reduced horizontal permeability
in the corresponding zone. Finally, the equivalent permeability
concept is applied in scenario 3.
In the unit cell analyses, three observation points are defined.
Point A is placed at the center of the model surface, while point
B is located 8 m beneath point A. The horizontal distance between
points A and B is 0.34 m. Point C is defined at the corner of the unit
cell, and it has the same depth as point B.
The variations in the vertical displacements and pore water
pressure with respect to time for points A, B and C are shown in
Fig. 5(b)–(d). It can be observed that without modeling the smear
zone, less pore water pressure is generated during the construction
of the embankment, and the dissipation of the pore water pressure
is faster. Consequently, the soil domain in scenario 1 settles faster
than that in scenario 2, while the final settlement is identical. Fur-
thermore, it is found that the pore pressure variations at points B
and C are identical in the 3D unit cell where the equivalent perme-
ability concept is applied (see Fig. 5(d)). This is because the equiv-
alent permeability method approximates the average pore
pressure in the cell. When the equivalent permeability concept is
compared with the explicit method, where the drain and smear
zone are modeled, a slight disagreement of the pore pressures at
points B and C can be observed due to the different distances
between the measurement points and the drain. Generally, it can
be concluded that both methods, the equivalent permeability 3D
model and the 2D model, provide results that compare well with
the refined 3D model, where both the drain and the smear zone
were explicitly simulated.
92 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103
2.6. Constitutive modeling
Fig. 2 illustrates the subsoil beneath the constructed test
embankment as modeled in the FE model. Underneath a 1-m-
thick near-surface silt layer, a stratum of soft clay with an overall
thickness of approximately 10 m is encountered. The material
forming this layer is natural, structured Ballina clay. As a typical
soft soil, Ballina clay is characterized by its destructuration and
creep behavior as well as its high compressibility [1]. Thus, it can
be assumed that this clay layer dominates the compression behav-
ior of the overall subsoil. Consequently, in the present study, par-
ticular focus is set on the adequate constitutive modeling of this
material. To capture the change in the material behavior over
depth more precisely for this layer, additionally, it is subdivided
into three sub-layers, Clay1-3, in the following.
To achieve an adequate modeling of the compression behavior,
the choice of the constitutive law is essential. To accurately repro-
duce the behavior of a soft, natural soil, a constitutive law able to
model anisotropy, destructuration and creep is required. However,
a constitutive model incorporating all three features is very com-
plex and would involve many constitutive parameters to be deter-
mined by various complex stress path tests.
A comparison of different constitutive models suitable to model
soft clay is shown in [2]. There, S-CLAY1 [9], S-CLAY1S [10], Soft
Soil (SS) and Soft Soil Creep [11] model are compared regarding
their suitability to model the material behavior of Ballina clay in
the case of the embankment loading.
Accordingly, SSC is chosen and applied to reproduce the soil
behavior of the silt and soft clay layers beneath the embankment
fill in the present study. The effects of bonding and destructuration
in the present study are captured by an adequate determination of
the material parameters described in detail in the subsequent
section.
The sand layer and the embankment fill material are modeled
using the Hardening Soil (HS) model [12,13]. The horizontal drai-
nage layer is assumed to have identical material properties as
the embankment layer. The Mohr-Coulomb model is adopted to
model the behavior of the bottom deep clay layer.
2.7. Constitutive parameters from field and laboratory test data
To identify the material parameters for the subsoil, a multitude
of laboratory and field test data were provided by CGSE. Moreover,
Pineda et al. [1] provided a good compendium for the important
material properties of Ballina clay.
Because the settlement prediction in this study is conducted
using a finite element consolidation analysis, special focus within
the parameter determination was set on the compression parame-
ters (namely, OCR, k and j ), creep parameter (l ) as well as the
permeability (k). As described in the previous section, the silt
and clay layers assumed to dominate the compression behavior
of the subsoil beneath the embankment are modeled using the
SSC model. Since this constitutive model does not account for
destructuration and anisotropy, a careful assessment of the param-
eters considering these restrictions has to be accomplished.
A description of the complete parameter identification would
be beyond the scope of this paper; in the following, exemplarily,
the determination of the compression parameters for the clay layer
is illustrated. For this purpose, constant rate of strain (CRS) test
data provided by CGSE and additional data from the literature [1]
were employed. Fig. 6 shows the test results from a CRS test on
an undisturbed, natural clay sample from a depth of approximately
9.80 m.
Ballina clay behaves similarly to a typical structured clay: when
subjected to an increasing loading, the bonding in the soil structure
3.2
σ0 σy σf in
3
2.8 λ∗0 ∗
≈κ
2.6
Void ratio, e [-]
2.4 λ∗1
2.2
2
1.8
λ∗2
1.6
1.4
1.2
10 100 1,000
Effective vertical stress, σv [kPa]
Fig. 6. CRS test data provided by CGSE.
initially exhibits a relatively stiff behavior of the soil matrix (k0 ).
Only when the yield stress ry is exceeded does the amount of
bonding decrease, and the soil experiences destructuration. Conse-
quently, the yield stress marks the stress at the transition point,
where the compression of the undisturbed, natural soil changes
from elastic to elasto-plastic. The destructuration is reflected in
the stress-strain behavior by a sudden drop in the soil stiffness
(k1 ) and thus a significant compression caused by a rearrangement
of the soil particles [14–16]. In higher stress ranges, the soil matrix
is observed to act slightly stiffer again (k2 ).
To model this compression behavior, a constitutive law account-
ing for destructuration would be needed. However, considering the
stress range covered in the embankment analysis, the following
simplification can be applied: The mean in situ stress r0 (see
Fig. 6) of the clay layer according to [1] is lower than the yield stress.
During the construction of the embankment, an additional loading
is applied. As the approximation, the effective vertical stress rfin
after construction continues to lie within the range where k1 is rel-
evant, and the destructuration process can be modeled as a transi-
tion from over- to normally consolidated soil. This means that the
soil is modeled as an over-consolidated soil deforming under j until
reaching the yield stress ry . The compression index after this can be
described by k1 . Following these considerations, one of the most
important parameters is the yield stress ratio (YSR), which governs
the change in the compression index. For the calculation here, the
YSR is set equal to the over consolidation ratio (OCR). Mean values
for the OCR as well as the creep coefficients are taken from [1]. A
correction of the YSR/OCR due to rate effects in the CRS test was
not considered here. The relevant constitutive parameters for all
layers are given in Table 2. Further details on the parameters can
also be found in [2]. It should be noted that the coefficient of the lat-
eral earth pressure of normally consolidated soil (K NC
0 ) depends on
0 ¼ 1  sin u .
the friction angle as K NC 0
2.8. Validation of constitutive approaches
As mentioned before, the Soft Soil Creep (SSC) model is
employed in this research to reproduce the clay behavior, although
it does not consider the effect of destructuration. Therefore, it is
necessary to validate the applicability of the constitutive model
modified in this manner for the embankment prediction.
Fig. 7 shows the stress path at the three different clay layers
from numerical simulation. The initial stress is obtained by K 0
analysis (point 0). When the earth fill of the embankment starts
 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103 93

Table 2
Values of model parameters for blind prediction.

Soil strata u0 [°] c0 [kPa] Eref

ur [MPa]
a k j l e0 c [kN/m3] OCR K0 mur
Embankment 30 1 75 – – – 0.5 20.6 1.00 0.5 0.2
Silt 34 15 – 0.158 0.053 6.2 E3 1.1 17.8/21.3b 2.86 0.93 0.15
Clay 1 41 0 – 0.177 0.010 4.3 E3 2.0 15.0 2.71 0.63 0.15
Clay 2 36 0 – 0.277 0.010 4.3 E3 3.0 15.0 1.73 0.58 0.15
Clay 3 32 0 – 0.239 0.012 4.9 E3 3.0 15.0 1.81 0.71 0.15
Sand 38 0 75 – – – 0.77 17.0 1.30 0.42 0.2
Deep clay su ¼ 40 kPa E ¼ 9 MPa – – – 0.5 18.5 – 0.60 0.33
a
In Hardening soil model: Eref ref ref
ur ¼ 3E50 ¼ 3Eoed .
b
Unit weight of soil: 17.8 (above ground water level) and 21.3 (below ground water level).

to be constructed (point 1), the pore water pressure generated dur-
ing the preparation of the working platform has been fully dissi-
pated. During the embankment construction, the stress paths at
all three depths of the clay layers cross the initial yield surface
(from points 1–2). This is consistent with the aforementioned
assumption that k is determined after reaching the yield stress
and justifies the assumption of modeling the initial compression
behavior with j while setting k ¼ k1 from the testing data. After
1000 days (point 3), the excess pore water pressures have fully dis-
sipated. Additionally, it should be noted that in the upper two clay
layers, localized failure occurs at the 47th day (the end of embank-
ment construction is the 49th day). After that, the stress path fol-
lows the failure line. In this process, the settlement and excess
pore pressure continue increasing until the end of embankment
construction. With the development of the consolidation process,
the effective stresses increase due to the excess pore pressure dis-
sipation and bring the stress path away from the failure line.
3. Prediction validation
Fig. 8 shows the model responses when applying the initially
obtained parameters (see Tables 1 and 2) from the 2D model using
discrete drainage with smear zone and from the 3D model using
the equivalent permeability concept. The locations of the measure-
ment points are given in Fig. 2. Three measurement positions over
the depth were selected for comparison of the time-dependent set-
tlements (SP2, Magnet 1 and Magnet 2) and pore water dissipation
(VWP1a, VWP1b, and VWP1c). From a comparison of the 2D/3D
finite element analysis with the in situ measurement data, the fol-
lowing conclusions can be drawn:
The vertical displacements and pore water pressures at the dif-
ferent measurement points are almost identical in the 2D and 3D
finite element models. This is consistent with the findings from
the unit cell analysis in Section 2.5 and confirms the model valida-
tion regarding dimensionality.

Fig. 7. Stress path at different clay layers [point0: construction starts, point1: start of the earth fill of the embankment, point 2: end of the earth fill of the embankment, point
pffiffiffi pffiffiffi construction]; (a) depth of 2 m in clay 1 layer, (b) depth of 5 m in clay 2 layer, (c) depth of 8.9 m in clay 3 layer. Here, q represents the
3: 1000 days after embankment
deviatoric stress, m ¼ 3=ð 3 cos hl þ sin hl sin u0 Þ, with hl being the Lode angle, [17].
94 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103
Fig. 8. Comparison of numerical results (blind prediction) with in situ measurements for (a) vertical displacements, (b) total pore water pressures.
The comparison between the predicted and in situ-measured
settlements reveals that the numerical models adequately repro-
duce the deformation behavior for measurement point SP2 and
Magnet 1 up to approximately 200 days. At the deeper measure-
ment point (Magnet 2), the calculation models overestimate the
settlements. It can be concluded that the compression behavior
of the clay layers below measurement point Magnet 1 in terms
of the compression parameters should be identified more accu-
rately. After 200 days, the calculation models predict a decay in
the settlements, while the settlements in situ continue increasing.
This holds true for all measurement points. This indicates that the
modeling of the consolidation and/or creep behavior of the soft silt
and clay layer requires improvement.
The comparison between the predicted and measured pore
water pressures shows that the maximum pore water pressure
and dissipation behavior at measurement point VWP1a are ade-
quately reproduced by the computational models. The in situ pore
pressures at point VWP1a after approximately 300 days exhibit
oscillations, which may be due to seasonal groundwater table fluc-
tuations as reported in [18] or could be caused by settlement of the
VWP. Since water table fluctuations are not modeled and since no
particular large displacement analysis was adopted, these effects
cannot be reproduced by the FE model used within the present
study.
The maximum pore water pressures and consequently their dis-
sipation behavior at the other measurement points (VWP1b and
VWP1c) are significantly underestimated by the computational
models. This indicates a too fast predicted consolidation behavior,
which can be improved by a more accurate identification of the
stiffness and/or permeability of soil.
For the sake of simplicity and to reduce the computational
effort for the later parameter improvement study, the 2D finite ele-
ment model is used. Since in the 2D model the difference in the
computational effort between the discrete drain modeling and
the equivalent permeability method is of minor importance, sce-
nario 2 is given preference in this study.
4. Parametric study
4.1. Global sensitivity analysis
The blind prediction is executed with parameter values that are
not precisely known. Rather than as exact values, the parameter
values need to be understood as the mean values of ranges esti-
mated by engineering judgment. The range of the parameter vari-
ations varies according to the precision of the parameter
determination. However, it is always influenced by many factors,
from the inherent variability to measurement inaccuracy and
model precision.
To validate a model, it is of great importance to know which
parameters are most significantly influencing the output or to
which parameter variations the output is most sensitive accord-
ingly. An appropriate method to identify the influence of different
input parameters on a specific model output is the so-called sensi-
tivity analysis.
In general, two groups of approaches for sensitivity analysis
exist: local and global sensitivity analysis. In local sensitivity anal-
yses, the parameters’ influences are examined independently from
each other by a small parameter variation, comparable to the par-
tial derivatives of the model. However, this method can only be
valid in the local region of the point in the input space or for con-
tinuously linear models. Moreover, possible correlation effects can-
not be considered, as all parameter variations are performed
individually.
Since output results from complex non-linear models and large
parameter ranges are anticipated in the present study, a global sen-
sitivity analysis (GSA) is more appropriate. The employed variance-
based GSA, introduced by Sobol [19] and further developed by Sal-
telli [20], allows for an impact quantification of input variances on
the output variances. The resulting first-order index Si is defined as
follows:
ðyTA yC i Þ  nðyA Þ2
Si ¼ ð8Þ
ðyTA yA Þ  nðyA Þ2
To additionally consider correlation effects between different
parameters, the total effect index ST i is employed:
ðyB  yC i ÞT ðyB  yC i Þ
ST i ¼ ð9Þ
2yTB yB  2nðyB Þ2
where yA ; yB and yCi denote vectors containing model outputs gener-
ated from samples all over the input space of the parameters i, and
A and y B are the corresponding mean values.
y
According to [20], GSA evaluations require 500–10,000 model
runs. More complex models require more runs to obtain reliable,
stable mean values and variance. Even with a simple 2D FE model,
this is significantly time consuming. A tool to overcome this type of
computational problem is provided by meta- or surrogate model-
ing approaches. Once the metamodel is established, the computa-
tional costs are very low; thus, 10,000 runs could be performed for
the GSA in the present study.
Within metamodel approaches, computationally expensive FE
models are replaced by efficient mathematical regression functions
[21]. The regression function correlating the input and output data
 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103
set is the so-called metamodel. Each input is assumed to be an s-
dimensional vector that represents the different model parame-
ters, and there are m observation points in the outputs. For the
training of a metamodel, first, 150 input parameter samples (150
sample points with s dimensions) are generated via the Latin
Hypercube sampling method in the pre-defined ranges of the cor-
responding parameters. Then, the corresponding output values
(150 outputs for m observation points) can be obtained via FE sim-
ulation. Within this framework, the metamodel is the regression
function mathematically correlating the inputs (s dimensions) with
the output at each observation point. The type of regression func-
tion, e.g., polynomial regression or radial basis function, must be
chosen according to the complexity of the computational problem.
The metamodel should be validated before its employment by
comparison of independent test samples from the FE and meta-
model. This validation ensures the accuracy of the metamodel.
Using a metamodeling approach, the total calculation time for a
GSA can be reduced to a fraction of the time needed without this
approach. The application of GSA to geotechnical problems is
demonstrated in [22–25].
In the present study, a GSA is performed with a metamodeling
approach using proper orthogonal decomposition in combination
with radial basis functions (POD-RBF); see [26].
To reduce the dimension of the metamodel, only the clay layers
assumed to dominate the compression of the subsoil are consid-
ered in the GSA. Moreover, a coupled behavior of the three clay
sub-layers is assumed. This means that all three clay sub-layers
experience identical relative changes for each parameter. Six input
parameters are introduced for each clay sub-layer: the permeabil-
ity k, over-consolidation ratio OCR, modified compression index k ,
modified recompression index j , modified creep index l , and
friction angle u0 . The parameter space is assumed through an accu-
racy estimation of the parameter assessment from laboratory and
field test data and by engineering judgment: the permeability k
is varied by a factor of 10 in both directions, and the other param-
eters are varied by ±20%. As output, the settlements after
1000 days at the ground surface (SP2) are evaluated.
Fig. 9 shows the results of two GSAs performed for the embank-
ment model. The GSA including the OCR as an input parameter
illustrates the clear dominance of the OCR on the settlement out-
put. Since the OCR defines in which stress range j lies and in
which stress state k is applied, it has an augmented influence on
the final settlement value. When neglecting the OCR as a variable
for the GSA, a clearer picture of the settlement sensitivity toward
the other parameters can be derived. The compression index k
and permeability k exhibit the highest sensitivity, while the recom-
pression index j is of minor influence. This seems plausible, as the
settlements are determined in a consolidation analysis whereby
stiffness and permeability are expected to have a dominating influ-
ence. As anticipated in the assumptions for the constitutive model,
Fig. 9. Sensitivity indices from GSA with respect to settlements at the ground
surface (point 0 in Fig. 2) after 1000 days.
95
considering creep as represented by the creep parameter, l is also
of importance. The friction angle u0 is found to have a small influ-
ence on the settlements. This can be anticipated since volumetric
deformation is dominant for the embankment behavior and since
shear deformation is less influential in this case.
Generally, GSA can be performed using any type of output infor-
mation (e.g., deformation as well as pore water pressure) gained at
any position and time point. The choice of the selected reference
output certainly has a major influence on the calculated sensitivity
indices, as the hydraulic or mechanical model responses will have
different sensitivities toward the different constitutive parameters.
In the present study, the focus was set on the adequate modeling of
the displacement and consolidation behaviors. Thus, the vertical
displacements in the center line of the embankment at the final
stage of the measurements were selected as the most relevant
model output and were chosen as the reference output for the
GSA. Of course, if other aspects come into focus, additional GSA
should be considered.
4.2. Inverse analysis
To obtain a parameter set that gives an optimum fit between
the predicted and measured values, inverse analysis is applied
for parameter identification [27–29]. In the inverse analysis, first,
an objective function is defined as the normalized sum of normal-
ized differences between calculation and measurement:
"  2 #
1X M
ycalc ðXÞ  ymeas
f ðXÞ ¼ wi i i
ð10Þ
M i¼1 ymeas
i
where X is the vector of the input parameters of the constitutive
model to be identified; ycalci ðXÞ and ymeasi are the calculated result
and field measurements, respectively; M is the number of measure-
ments; and wi is an optional weighting factor for the corresponding
measurement points, assumed to be 1 in the present study.
The aim of the inverse analysis is to identify the set of input
parameters for which a minimization of the objective function is
reached. The minimum of the objective function may serve as a
measure for comparison of the quality of different inverse analy-
ses: a smaller f ðXÞ indicates a better optimization result. This can
be found using evolutionary algorithms, e.g., Genetic Algorithms
(GA) or Particle Swarm Optimization (PSO) [30,31]. For the present
study, a genetic algorithm (GA) is chosen for solving the optimiza-
tion problem. This type of algorithm is inspired by Darwin’s evolu-
tion theory. Details of applying GA algorithms to similar problems
can be found in [25]. It should be noted that all inverse analyses in
this study were performed using the metamodel to identify the
uncertain model parameters; then, the identified parameters were
adopted in the FE model to evaluate the model responses and com-
pared with the real measurements.
As described above, the inverse analysis itself and its output are
strongly dependent on the input information provided for the
parameter optimization, namely, on the optimized input parame-
ters and the output parameters incorporated in the objective func-
tion. In particular, the chosen output employed in the objective
function, in terms of type, amount and quality of data incorporated,
may have a significant influence on the quality of the inverse anal-
ysis results. Thus, in the present study, a hybrid inverse analysis
approach considering different optimized input parameters as well
as different output parameters for the objective function is per-
formed. The input parameters were defined according to the rele-
vant parameters identified in the sensitivity analysis. Specifically,
three inverse analysis scenarios are studied:
Scenario I: Optimized input parameters: relevant parameters
for the clay layer (OCR; k ; j ; l , and k) assuming a coupled behav-
ior of the three sub-layers such that any relative change in one
96 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103
parameter is assumed to be identical for this parameter in all three
sub-layers (see Section 4.1).
output parameters for objective function: Scenario I(a) – total
settlements at 3 positions over depth (SP2, Magnet 1 and Magnet
2; see Fig. 2); Scenario I(b) – pore water pressure at 3 positions over
depth (VWP1a-c; see Fig. 2); Scenario I(c) – total settlements at 3
positions over depth (SP2, Magnet 1 and Magnet 2) and pore water
pressure at 3 positions over depth (VWP1a-c).
Scenario II: Optimized input parameters: relevant parameters
for the first and third clay sub-layers as well as the silt layer
(OCR; k ; j ; l , and k) were varied independently from each other.
output parameters for objective function: layer-wise settle-
ments calculated from 5 measurement points over depth (Points
1–5; see Fig. 2) for the considered layers (1st and 3rd clay sub-
layers and silt layer).
For the objective function in each case and for each measure-
ment response (three settlements and/or three pore water pres-
sure measurements), 6 time points equally distributed over the
logarithmic time variation were chosen.
In scenario I and II, the complete time span of the available
measurement data (up to 1000 days) is used for the inverse analy-
sis. However, inverse analysis is usually performed not to fit exist-
ing data but rather to calibrate a model with data from a limited
period of time, where measurement data are available. Then, the
calibrated model is used to predict the future behavior of the sys-
tem. Therefore, in scenario III, measurement data from shorter
time spans are used to calibrate the model. Then, the model predic-
tion for up to 1000 days is compared to the monitoring data to
evaluate the influence of the time span on the prediction quality
and to analyze the ability of the model to capture the real response.
Scenario III: Optimized input parameters: as defined in sce-
nario II.
output parameters for objective function: layer-wise settle-
ments as defined in scenario II. Scenario III(a) – for 6 time points
in the first 243 days; Scenario III(b) – for 6 time points in the first
510 days;
All inverse analyses were performed using the metamodeling
approach.
4.3. Inverse analysis scenario I – Parameter optimization of three clay
sub-layers based on total settlements and pore water pressure
Fig. 10 displays the results of the soil settlements and pore
water pressure dissipation using the optimized parameters
attained from inverse analysis scenario I.
Fig. 10(a) and (b) compares the in situ measurements with the
model responses calculated using the parameter values identified
by inverse analysis based on the measured total displacements
(scenario Ia). The numerical simulation of the settlement paths
agrees with the in situ measurements overall. In particular, the
final settlements fit much better compared to the blind prediction.
Moreover, the maximum pore water pressure and dissipation
behavior at point VWP1b are modeled properly. However, some
differences between the prediction results and measured data
remain. The consolidation settlements for Magnet 2 start too early,
which indicates a drainage that is too fast. In addition, the maxi-
mum pore water pressures and dissipation behavior at VWP1a
and VWP1c require improvement. Here, it is noteworthy that the
maximum pore water pressure is overestimated at VWP1a,
whereas it is underestimated at VWP1c. This indicates that a uni-
form change in the parameters for the three clay sub-layers within
the inverse analysis might not be adequate. Therefore, an indepen-
dent optimization is needed for these three sub-layers. However, in
general, it can be stated that using parameter values identified in
scenario Ia, a clear improvement in the model response compared
to the blind prediction (see Fig. 8) can be achieved.
Fig. 10(c) and (d) presents the results obtained when using the
optimized parameters derived based on the measured pore water
pressure (scenario Ib). Although the prediction of the maximum
pore water pressure at point VWP1c is improved, the overall pore
water pressure dissipation is modeled even less accurately com-
pared to scenario Ia. Moreover, the predicted soil deformations sig-
nificantly differ from the in situ measurement data. Fig. 10(e) and
(f) shows the results using the optimized parameters derived based
on the vertical displacements and the pore water pressure (sce-
nario Ic). Here, both the settlements as well as the pore water dis-
sipation prediction exhibit strong differences to the in situ
measurements. This may be attributed to the fact that seasonal
groundwater fluctuations as reported in [18] or possible displace-
ment of the VWP are not modeled in the FEA nor is any large dis-
placement analysis adopted. Since these effects cannot be captured
by the constitutive parameters alone, an inverse analysis using this
pore pressure measurement is less accurate than an inverse analy-
sis based on settlements.
By comparison of scenarios Ia to Ic, it can be illustrated that the
choice of data to form the objective function has a significant influ-
ence on the inverse analysis. It becomes obvious that an increase in
the amount of data considered in the evaluation of the objective
function does not automatically improve the result of the inverse
analysis. Since there are more uncertainties involved in the objec-
tive function when incorporating more data, the overall prediction
might also be of less quality compared to optimization based on a
smaller amount of measurement data. According to the above dis-
cussion, inverse analysis based on soil displacements provides a
good agreement between numerical results and in situ measure-
ments. Therefore, the optimized parameters obtained via scenario
Ia are used for the further inverse analysis study.
The parameters found by inverse analysis scenario Ia enable a
rather good reproduction of the total settlements and pore water
dissipation. However, several soil layers are involved in this case
study and contribute to the total settlement. The aim of the inverse
analysis is to not only reproduce the final total settlements but also
identify parameter sets characterizing the soil behavior of each
layer adequately. Therefore, the layer-wise displacements are a
good indicator for the validation of the inverse analysis. Fig. 11
shows the total and layer settlements using the initially obtained
parameters for the blind prediction. The layer displacements are
calculated by the difference in the adjacent measurement points
(see Fig. 2). It can be observed that there is a large disagreement
between the prediction and in situ measurements for all layers.
To study the quality of inverse analysis scenario Ia, Fig. 12
shows the same comparison between the total and layer displace-
ments calculated with the optimized parameter values. The com-
parison reveals that the layer settlement prediction for the
middle clay sub-layer and the sand layer (including a thin part of
clay sub-layer 3) are improved under inverse analysis scenario Ia.
However, the layer settlement predictions for the silt and the
upper and lower clay sub-layers still do not adequately match
the in situ measurements. To follow the aim of identifying param-
eter values that reproduce the physical behavior of the soil, more
effort needs to be spent to improve the displacement prediction
for these layers. Therefore, a second inverse analysis (scenario II)
is conducted in the following.
4.4. Inverse analysis scenario II – Parameter optimization for silt layer
and first and third clay sub-layers based on layer displacements
Inverse analysis scenario II is performed based on the parame-
ter values identified in scenario Ia. The silt layer and the first and
third clay sub-layer are selected for further parameter identifica-
 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103 97

Fig. 10. Results of inverse analysis scenario I: (a) vertical displacements using scenario Ia, (b) pore water pressure variation using scenario Ia, (c) vertical displacements using
scenario Ib, (d) pore water pressure variation using scenario Ib, (e) vertical displacements using scenario Ic, (f) pore water pressure variation using scenario Ic.

tion in terms of their relevant parameters (OCR; k ; j ; l , and k). To
form the objective function, the layer-wise settlements calculated
from 5 measurement points over depth (see Fig. 2) for the consid-
ered layers are used.
Fig. 13 shows the model response compared to in situ measure-
ments from this study. It can be stated that the calibrated numer-
ical model is able to well predict the total as well as layer-wise
settlements. Additionally, the pore water pressures displayed in
Fig. 14 are precisely reproduced in terms of the maximum total
pore water pressure and dissipation behavior.
Comparing the results of scenarios I and II, a further improve-
ment in the prediction can be found. This improvement is observed
in the improved modeling of the layer-wise settlements for the silt
and upper clay layers and consequently in the improved total set-
tlements. This can be quantified according to the minimum value
of the objective function for the two scenarios: 0.0679 for scenario
Ia and 0.0657 for scenario II. The layer-wise settlements for the
lower clay layer are improved regarding its shape with an
unchanged final value. The remaining differences between the
measurements and the model predictions could be eliminated by
further improvement of the constitutive soil model in describing
the anisotropy and destructuration behavior.
Table 3 gives the initially chosen parameters (blind prediction)
and the optimized parameters (scenario II). According to the com-
98 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103

Fig. 11. Model responses using the initial prediction parameter set; (a) total displacements, (b) layer displacements.

Fig. 12. Model responses of inverse analysis scenario Ia: (a) total displacements, (b) layer displacements.

Fig. 13. Model responses of inverse analysis scenario II: (a) total displacements, (b) layer displacements.

parison between these parameters, the optimized soil properties
are physically reasonable. This is also supported by the sortation
of the parameter magnitude, as the OCR and the permeability
become smaller with increasing depth.
Additionally, it can be concluded that all initial OCR values
for the clay were over-estimated by approx. 10–20%. This may
be attributed to the fact that rate effects in the CRS test were
not considered in the initial determination of the OCR [1]. Con-
sequently, the initial yield stress is over-estimated, which
results in a stiffer compression behavior. The transition from
over- to normally consolidated soil was set too late. The impact
of this effect was strengthened by the underestimation of the
compression and recompression indices. This particularly holds
true for j values of the clay layers. Moreover, Fig. 14 shows
the model validation regarding the pore water pressure
generation and dissipation. The accurate permeability values
enable the model to capture the in situ pore water dissipation
process.
 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103
Fig. 14. Model responses of inverse analysis scenario II: pore water pressure
dissipation.
4.5. Inverse analysis scenario III – Prediction of soil behavior based on
parameter optimization
Figs. 15 and 16 display the model response from the parameter
identification using measurement data for up to 243 days (scenario
IIIa) and 510 days (scenario IIIb).
A comparison between scenario IIIa (243 days), scenario IIIb
(500 days) and scenario II (1000 days) shows that the conformity
between the monitoring data and the model prediction increases
with increasing time span used in the process of parameter identi-
fication. Although scenario IIIb reproduces the overall settlement
behavior well and only slightly underestimates the final settlement
values in clay layers 1 and 2, scenario IIIa does not capture the
ongoing settlement increase monitored after 243 days; it shows a
decay in the deformation course. It can be concluded that measure-
ment data from more than 243 days is required to accurately iden-
tify the parameters to model the complex time-dependent
phenomenon of consolidation and destructuration in the clay layer.
To quantitatively evaluate the different numerical simulation
scenarios, Table 4 compares the objective function values calcu-
lated from Eq. (10) for settlement measurements and prediction
data at points 0–3 at times 97d, 243d, 510d, 810d and 1110d. Gen-
erally, it can be stated that with more available data, a smaller
error (value of the objective function) can be derived. Moreover,
it is interesting to see that the initial blind prediction is more accu-
rate than the prediction based on the inverse analysis using only a
restricted amount of data (scenario IIIa). Since the back calculation
for scenario IIIa is based on data up to 243 days only, the long-term
settlement behavior with considerable creep settlements is not
adequately captured. In contrast, in the blind prediction, the
parameters were determined by engineering judgment accounting
for this effect.
For a more detailed study on the evolution of the parameter val-
ues within the different inverse analysis scenarios, Fig. 17 visual-
izes the evolution of the most relevant parameters for the four
considered layers and for different model scenarios. All parameter
Table 3
Values of optimized parameters from initial and optimum parameter sets.
Soil strata OCR k
a b a b
Silt 2.86 /3.99 0.158 /0.115
Clay 1 2.71a/2.09b 0.177a/0.170b
Clay 2 1.73a/1.60b 0.277a/0.320b
Clay 3 1.81a/1.58b 0.239a/0.279b
a
Parameter values from initial parameter set (blind prediction).
b
Parameter values from optimized parameter set (scenario II).
99
values are normalized with respect to the value of the best sce-
nario: scenario II. It can be observed that the permeability in the
initial prediction and scenario IIIa significantly differs from the
final value. The same holds true for the OCR value of clay layer 2.
The deviation of these sensitive parameters may explain the higher
error in these two prediction scenarios.
5. Probabilistic evaluation
The initial as well as optimized predictions are based on param-
eter values that are not precisely known but that are mean values
with ranges estimated according to the parameter assessment
accuracy from laboratory and field tests as well as engineering
judgment. To observe the impact of the parameter uncertainty on
the model output, an uncertainty analysis is executed. This uncer-
tainty analysis provides information on how accurate the proposed
results might be. This study is provided for the initial as well as
optimized prediction model, as a comparison between the output
uncertainty from the initial and optimum parameters will help
provide a better understanding of the performed improvement
strategy.
In soil mechanics, the three generally accepted sources of
uncertainty are inherent variability, measurement inaccuracy,
and transformation or model uncertainty [32]. The inherent vari-
ability is caused by the natural origin of the encountered soils,
which are heterogeneous materials. As illustrated in Section 2.7,
laboratory and field test measurements can only be conducted to
a specific precision. This inaccuracy of measurements of course
affects the precision of the parameter determination. Finally, the
transformation uncertainty refers to the ability of computational
models to reproduce the physical processes occurring in situ.
In the following, the deterministic calculations performed pre-
viously are expanded to probabilistic analyses. To perform these
analyses, probability distributions instead of discrete values for
the parameters are assumed. In the present study, normal distribu-
tions are selected, with the exception of the permeability and fric-
tion angle, which are assumed to be log-normally distributed.
These assumptions are commonly used in geotechnical applica-
tions in case only limited information on the soil properties is
available [33].
From the parameter space using the Monte-Carlo Method
(MCM), random samples are selected corresponding to the chosen
distributions. In the present study, 10,000 samples are employed.
Analogous to the GSA, a metamodeling approach used to minimize
the computational effort is applied for the calculation of the output
distribution. The distributions are characterized by the estimate
value EðuÞ and the standard deviation r.
The probabilistic evaluation for the initial prediction is per-
formed using the parameter values derived from laboratory and
field test data as the mean values (see Table 3). The standard devi-
ations r defining the parameter ranges are assessed with reference
to the available experimental data corresponding to the parameter
range defined for the metamodel (see Section 4.1). For the proba-
bilistic evaluation of the optimized prediction, the parameter val-
ues identified in scenario II are used as the mean values,
j l k [m/day]
a b a
0.053 /0.048 6.2 E3 /5.0 E3b 2 E3a/9 E4b
0.010a/0.064b 4.3 E3a/4.4 E3b 2 E4a/7 E4b
0.010a/0.048b 4.3 E3a/4.0 E3b 6 E4a/1 E4b
0.012a/0.023b 4.9 E3a/6.3 E3b 6 E4/6 E5b
100 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103
Fig. 15. Model responses of inverse analysis scenario IIIa: (a) total displacements, (b) layer displacements.
Fig. 16. Model responses of inverse analysis scenario IIIb: (a) total displacements, (b) layer displacements.
Table 4
Quantitative evaluation of numerical simulation scenarios.
Scenario Value of objective functiona
Initial (blind prediction) 0.0582
Ia 0.0170
II 0.0045
IIIa 0.0993
IIIb 0.0182
a
Note: The objective function is given in Eq. (10), and the measurements and
predictions of the settlements at points 0–3 at times 97d, 243d, 510d, 810d and
1110d are used.
assuming the same standard deviation r. To compare and quantify
the different outputs, the standard deviations of the output distri-
butions are calculated by root extraction of the variances.
Fig. 18(a) displays the distribution of final settlements at
ground surface (measurement point SP2) from uncertainty analysis
for initial and optimized prediction. Fig. 18(b) illustrates the corre-
sponding maximum pore water pressure distributions in point
VWP 1a. For all cases, the mean value of the distributions equals
the value calculated in the deterministic analyses. According to
the parameter distribution for the optimized parameter set it can
be observed that the measured values lie within the range of stan-
dard deviation around the predicted mean value. Therefore, it can
be concluded that this model reproduces the measurement data
with sufficient accuracy.
Given that the standard deviation for the pore water pressure
output is always larger than the standard deviation for settlements,
it can be concluded that the prediction of hydraulic behavior is
more complex. This is also reflected in the fact that the final settle-
ment value is more accurately predicted than is the maximum pore
water pressure. This may be attributed to the fact that pore water
pressure measurements are subject to stronger variations than
are settlement measurements and that some effects, e.g., seasonal
water table fluctuations, are not considered in the FE model.
It can be observed that the prediction after inverse analysis
exhibits a larger standard deviation than the blind prediction. This
holds true for the final settlement as well as the maximum pore
water pressure value and is caused by the combination of two
facts: On the one hand, the mean values in the analysis are shifted
due to the better knowledge from the inverse analysis. Simultane-
ously, the parameter value ranges, mathematically. Due to the non-
linear behavior of soil under consolidation analysis, the output dis-
tributions not only are shifted to another mean value but also are
different in shape – in this case, they are larger than the initial
distributions.
The question of why after an inverse analysis the improved
knowledge of the parameter value is not considered for the param-
eter range can be answered as follows: The parameter ranges in
terms of the standard deviation r characterize how precisely the
parameter is known. Keeping the standard deviation constant for
the optimized parameter set according to the terminology of [34]
is a pessimistic approach. However, the standard deviation r not
only considers the uncertainty due to measurement inaccuracies,
which certainly improves through inverse analysis, but also
accounts for the inherent variability of the parameter values, e.g.,
 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103 101

Fig. 17. Relative parameter development for different scenarios.

Fig. 18. Histograms of the probability distribution of the settlements at point 0 (a) and the pore water pressure at point VWP1a (b).

that caused by the soil heterogeneity in the field or the variability
of conditions in different cross sections, and the model uncertainty
itself.
An alternative approach would be a Bayesian updating as
described in [35,34]. This approach would allow for a reduction
in the input uncertainty based on the measured outputs. Since
probabilistic evaluation is not the focus of this study, the applica-
tion of this approach is beyond the scope of this paper.
Summarizing from a methodical point of view, it can be stated
that only improved knowledge of the parameter value ranges, i.e., a
smaller standard deviation r mathematically speaking, reduces the
uncertainty. A shift in the mean value from improved knowledge of
this characteristic does not decrease the uncertainty of the model
response. Consequently, improved knowledge of the measurement
accuracy and that of the homogeneity of the in situ conditions are
desirable. Precise measurements derived by design of experiment
(DOE) methods are essential for an adequate numerical simulation.
6. Conclusion and outlook
In the framework of this study, a blind prediction was per-
formed for an embankment constructed on soft, structured clay.
A numerical model was implemented based on assumptions and
simplifications regarding the geometrical conditions and staged
construction of the embankment, the initial and boundary condi-
tions as well as the constitutive modeling. The calibration of the
numerical model was conducted based on in situ measurement
data to identify the corresponding model parameters. Different
methods for parameter optimization were investigated. The main
conclusions can be stated as follows:
102 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103
1. Model simplifications are a useful tool for reducing computa-
tional effort and create an efficient calculation model. However,
each simplification needs to be validated and justified regarding
its ability to reproduce the in situ situation adequately. In the
case of the present study, the following model assumptions
could be justified in the model validation:
 Vertical drains for consolidation acceleration may be mod-
eled through the discrete modeling of the drains with a
smeared zone reproducing the surrounding, disturbed zone
or through an approach using soil material of equivalent per-
meability. This holds true for 2D as well as 3D FE models.
When doing so, a unit cell validation is needed to prove
the equivalence of the methods.
 Structured soils may be modeled adequately by constitutive
models not accounting for destructuration. However, when
doing so, a validation regarding the occurring stress-strain
range of the geotechnical application has to be performed.
Additionally, a careful assessment of the relevant model
parameters is required, e.g., considering rate effects in the
determination of YSR from CRS tests.
2. Global sensitivity analysis is a powerful method to distinguish
the relative importance of parameters on the model response
and to identify the most relevant parameters. Looking at com-
pression processes, the most relevant parameters are the OCR
and the compression/recompression indices k =j . The sensitiv-
ity of the last two parameters, however, is strongly influenced
by the choice of the OCR value.
3. Using inverse analysis, optimized values for the relevant consti-
tutive parameters can be identified to adequately reproduce the
measurement data. Regarding the inverse analysis, particular
attention has to be paid to the selection of the identified param-
eters and the data incorporated in the objective function. More
data forming the objective function does not automatically
improve the result of the inverse analysis. Since there are more
uncertainties involved, the overall prediction may also be of les-
ser quality compared to optimizations based on a smaller
amount of measurement data. A step-wise inverse analysis is
to be preferred, as it guarantees a holistic optimization strategy.
4. Parametric evaluation provides information on the uncertainty
of a numerical model using a specific parameter set. The uncer-
tainty analysis can be used to study mathematically the impact
of the inherent variability and measurement inaccuracy. The
variance VARðuÞ of the mean output significantly depends on
the input parameter range, as the COV accounts for all three
sources of uncertainty (inherent variability, measurement inac-
curacy, and model uncertainty) in equal measures. Thus, precise
measurements derived by design of experiment (DOE) methods
are essential for obtaining an adequate numerical simulation.
The mathematical methods (meta-modeling, sensitivity, and
inverse and uncertainty analyses) presented and applied in the
present study are powerful tools to identify optimum computa-
tional models and to analyze the influences of the input character-
istics on the model response. However, it becomes obvious that
engineering judgment has a significant impact on the quality of
these methods.
Acknowledgment
The authors would like to thank Wolfgang Lieske, M.Sc. and
Christoph Schmüdderich, M.Sc. for their expertise and support dur-
ing the initial prediction modeling. The third author acknowledges
financial support provided by the German Science Foundation
(DFG) under the framework of subproject C2 of Collaborative
Research Center SFB 837.
References
[1] Pineda JA, Suwal LP, Kelly RB, Bates L, Sloan SW. Characterisation of ballina
clay. Gotechnique 2016;66(7):556–77. http://dx.doi.org/10.1680/jgeot.15.
P.181. http://dx.doi.org/10.1680/jgeot.15.P.181 arXiv:http://dx.doi.org/
10.1680/jgeot.15.P.181..
[2] Hölter R, Lieske W, Müthing N, Schmüdderich C, Zhao C, Schanz T.
Settlement prediction for an embankment on soft, structured clay using
smeared soil-drain replacement technique and probabilistic analysis. In:
Embankment prediction symposium; vol. 1; 1 ed. ISBN 978-0-9953750-0-0;
2016, p. 166–79.
[3] Vakili K, Barciaga T, Lavasan A, Schanz T. A practical approach to constitutive
models for the analysis of geotechnical problems. In: Proceedings of 3rd
international symposium on computational geomechanics (ComGeo III). p.
738–49.
[4] Zhao C, Lavasan A, Barciaga T, Kämper C, Mark P, Schanz T. Prediction of tunnel
lining forces and deformations using analytical and numerical solutions. Tunn
Undergr Space Technol 2017;64:164–76.
[5] Yildiz A, Karstunen M. Three-dimensional analyses of PVD-improved soft soils.
In: Geotechnics of soft soils: focus on ground improvement. p. 197–203.
[6] Chai J, Shen S, Miura N, Bergado D. Simple method of modeling pvd improved
subsoil. J Geotech Geoenviron Eng 2001;127:965–72.
[7] Indraratna B, Redana I. Laboratory determination of smear zone due to vertical
drain installation. J Geotech Geoenviron Eng 1998;124(2):180–4.
[8] Indraratna B, Redana I. Plain strain modeling of smear effects associated with
vertical drains. J Geotech Geoenviron Eng 1997;123(5):474–8.
[9] Wheeler S, Näätänen A, Karstunen M, Lojander M. An isotropic elastoplastic
model for soft clays. Can Geotech J 2003;40:403–18.
[10] Karstunen M, Krenn H, Wheeler S, Koskinen M, Zentar R. Effect of anisotropy
and destructuration on the behavior of murro test embankment. Int J Geomech
2005;5(2):87–97.
[11] Vermeer P, Neher H. A soft soil model that accounts for creep. In: Proceedings
of 1st international PLAXIS symposium on beyond 2000 in computational
geotechnics. Balkema; 1999. p. 249–61.
[12] Schanz T. Zur Modellierung des mechanischen Verhaltens von
Reibungsmaterialien. Habilitationsschrift, Mitteilung 45 des Instituts für
Geotechnik. Universität Stuttgart; 1998.
[13] Schanz T, Vermeer P, Bonnier P. The hardening soil model: formulation and
verification. In: Proceedings of 1st international PLAXIS symposium on beyond
2000 in computational geotechnics. Balkema; 1999. p. 281–96.
[14] Burland J. On the compressibility and shear strength of natural clays.
Géotechnique 1990;40(3):329–78.
[15] Cotecchia F, Chandler R. The influence of structure on the pre-failure
behaviour of a natural clay. Géotechnique 1997;47(3):523–44.
[16] Cotecchia F, Chandler RJ. A general framework for the mechanical behavior of
clays. Géotechnique 2000;50(4):431–47.
[17] Zienkiewicz OC, Humpheson C. Viscoplasticity: a generalized model for
description of soil behaviour. In: Desai CS, Christian JT, editors. Numerical
methods in geotechnical engineering. New York: McGraw-Hill Book Company;
1977.
[18] Kelly RB, Pineda JA, Bates L, Suwal LP, Fitzallen A. Site characterisation for the
ballina field testing facility. Gotechnique 2017;67(4):279–300. http://dx.doi.
org/10.1680/jgeot.15.P.211. http://dx.doi.org/10.1680/jgeot.15.P.211 arXiv:
http://dx.doi.org/10.1680/jgeot.15.P.211..
[19] Sobol’ I. Sensitivity estimates for nonlinear mathematical models. Math
Modell Comput Exp 1993;1:407–14.
[20] Saltelli A, Ratto M, Andres T, Campolongo F, Cariboni J, Gatelli D. Global
sensitivity analysis. The Primer. John Wiley and Sons; 2008.
[21] Khaledi K, Miro S, Koenig M, Schanz T. Robust and reliable metamodels for
mechanized tunnel simulations. Comput Geotech 2014;61:1–12.
[22] Miro S, Hartmann D, Schanz T. Global sensitivity analysis for subsoil parameter
estimation in mechanized tunneling. Comput Geotech 2014;56:80–8.
[23] Zhao C, Lavasan A, Schanz T. Sensitivity analysis of the model response in
mechanized tunnelling simulation – a case study assessment. In: 4th
International conference on engineering optimization. Lisbon, Portugal: Boca
Raton; 2014. p. 491–6.
[24] Zhao C, Lavasan A, Barciaga T, Hölter R, Datcheva M, Schanz T. Constitutive
parameter adjustment for mechanized tunneling with reference to sub-system
effects. In: 8th International conference on numerical methods and
applications. Borovets, Bulgaria. p. 217–25.
[25] Khaledi K, Mahmoudi E, Datcheva M, Knig D, Schanz T. Sensitivity analysis and
parameter identification of a time dependent constitutive model for rock salt. J
Comput Appl Math 2016;293:128–38.
[26] Buljak V. Inverse analysis with model reduction: proper orthogonal
decomposition in structural mechanics. Springer Verlag; 2012.
[27] Sakurai S. Back analysis for tunnel engineering as a modern observational
method. Tunn Undergr Space Technol 2003;18:185–96.
[28] Yazdani M, Sharifzadeh M, Kamrani K, Ghorbani M. Displacement-based
numerical back analysis for estimation of rock mass parameters in siah bisheh
powerhouse cavern using continuum and discontinuum approach. Tunn
Undergr Space Technol 2012;28:41–8.
[29] Zhao C, Lavasan AA, Barciaga T, Zarev V, Datcheva M, Schanz T. Model
validation and calibration via back analysis for mechanized tunnel simulations
– the Western Scheldt tunnel case. Comput Geotech 2015;69:601–14.
 N. Müthing et al. / Computers and Geotechnics 93 (2018) 87–103 103

[30] Kennedy J, Eberhart R. Particle swarm optimization. Proceedings of IEEE
international conference on neural networks, vol. 4. p. 1942–8.
[31] Shi Y, Eberhart R. A modified particle swarm optimizer. In: Proceedings of IEEE
congress on evolutionary computation. p. 69–73.
[32] Phoon KK, Kulhawy F. Characterization of geotechnical variability. Can
Geotech J 1999;36:612–24.
[33] Ching J, Phoon KK. Constructing multivariate distributions for soil parameters.
In: Ching J, Phoon KK, editors. Risk and reliability in geotechnical
engineering. Boca Raton: CRC Press; 2014. p. 3–77. chap. 1.
[34] Miro S, König M, Hartmann D, Schanz T. A probabilistic analysis of subsoil
parameters uncertainty impacts on tunnel-induced ground movements with a
back-analysis study. Comput Geotech 2015;68:38–53.
[35] Kelly R, Huang J. Bayesian updating for one-dimensional consolidation
measurements. Can Geotech J 2015;52(9):1318–30.