Você está na página 1de 16

Wang, Y. et al. Gotechnique [http://dx.doi.org/10.1680/jgeot.15.P.

219]

Probabilistic assessment of liquefiable soil thickness considering


spatial variability and model and parameter uncertainties

FS
Y. WANG , C. FU  and K. HUANG 

This paper develops a probabilistic approach for identifying liquefiable soil layers and assessing
the liquefiable soil thickness using cone penetration tests (CPTs). The inherent spatial variability of
liquefaction resistance (i.e. cyclic resistance ratio (CRR)), the model uncertainty associated with the
CPT-based Robertson and Wride model, and input parameter uncertainty in cyclic stress ratio (CSR)
are taken into consideration explicitly in the proposed approach. The probability distribution of the
CRR and the thicknesses of soil layers with statistically homogeneous CRR are first identified. Then,
for a given nominal seismic loading, the thicknesses of liquefiable soil layers are estimated. A ratio of
the liquefiable soil thickness over the total thickness of all soil layers considered is quantified for
assessment of liquefaction consequences, and it is determined probabilistically using Monte Carlo
simulations. The proposed approach is illustrated using a set of CPT data collected from New Zealand.
Case histories of the 20102011 Canterbury earthquake sequence in New Zealand are used to
systematically validate the proposed approach. The proposed approach is shown to properly identify the
liquefiable soil layers, quantify their liquefiable soil thicknesses and associated uncertainty, and provide
OO results that are consistent with earthquake field observations. It is also shown that deterministic analysis
not only cannot quantify the uncertainty in the assessment results, but also fails to provide consistent
results for the assessment of liquefaction consequences. It is therefore necessary to perform probabilistic
analysis using the proposed approach. In addition, a sensitivity study is performed to explore the effect
of spatial variability on soil liquefaction.

Q2 KEYWORDS:

INTRODUCTION (e.g. Robertson & Wride, 1998; Juang et al., 2003; Lai et al.,
Soil liquefaction refers to a phenomenon a granular material 2004).
transforming from a solid to a liquefied state due to increased Consider, for example, the Robertson & Wride (1998)
pore pressure and reduced effective stress (e.g. Youd et al., method, referred to as the RW method in the literature (e.g.
2001; Cetin et al., 2002). In situ tests, such as the cone pene- Youd et al., 2001; Ku et al., 2012). The CRR is estimated
tration test (CPT), standard penetration test (SPT) and shear from CPT test results, followed by calculations of the CSR
wave velocity test, are widely used in engineering practice to and factor of safety, FS = CRR/CSR. No liquefaction is
Q3 assess the earthquake-induced liquefaction potential of soils expected to occur when FS . 1, and liquefaction is predicted
(e.g. Stark & Olson, 1995; Andrus & Stokoe, 2000; Moss if FS  1. As the calculation models for estimating CSR
et al., 2006; Baker & Faber, 2008; Juang et al., 2013). As CPT and CRR were mainly developed from past field obser-
tests provide a repeatable and nearly continuous data profile vations and engineering experience (see later section entitled
Estimation of CRR and CSR), they inevitably involve
PR

with depth, they are popular in engineering practice (e.g.


Robertson & Campanella, 1985; Lai et al., 2006). Many various uncertainties (e.g. measurement error, uncertainty
research studies have been performed on the assessment of in the model). For example, a value of FS . 1 does not
liquefaction potential of soils using CPT (e.g. Seed & Idriss, guarantee that no liquefaction is observed in field obser-
1971; Robertson & Wride, 1998; Juang et al., 2003; Jefferies vations, and vice versa. To deal properly with these
& Been, 2006). The simplified procedure developed by uncertainties, several probabilistic methods have been devel-
Seed & Idriss (1971) has been widely used in practice for oped for liquefaction potential assessment in the literature (e.
evaluating the liquefaction potential of soils around the g. Toprak et al., 1999; Juang et al., 2002; Toprak & Holzer,
world (Juang et al., 2004). The procedure has been updated 2003; Lai et al., 2006; Moss et al., 2006; Ku et al., 2012).
periodically (e.g. Seed, 1979; Seed & Idriss, 1982; Youd et al., However, these probabilistic methods mainly focused on
2001), and the seismic loading and resistance in the current whether or not liquefaction occurs, based on data at a
procedure are expressed as cyclic stress ratio (CSR) and cyclic particular point (i.e. at a specific depth during CPT). The
resistance ratio (CRR), respectively. Although the CSR equa- spatial correlation of soil properties at different depths is
tion developed by Seed & Idriss (1971) generally remains ignored in the assessment.
unchanged in the literature, the methods of calculating the In addition, the severity of liquefaction has attracted
CRR and various factors in the CSR equation are different increasing research attention (e.g. Iwasaki et al., 1982; Youd
& Perkins, 1987; Toprak & Holzer, 2003; Maurer et al., 2014,
2015). These previous studies have shown that liquefaction
severity is mainly affected by the thicknesses of the liquefi-
Manuscript received 29 September 2015; revised manuscript able soil layers, thickness of the overlying non-liquefiable
accepted 25 August 2016. layer, depth of the liquefiable layer beneath the ground
Discussion on this paper is welcomed by the editor. surface, FS value and so on. Therefore, an important issue
 Department of Architecture and Civil Engineering, City in the liquefaction severity assessment is how to determine
University of Hong Kong, Kowloon, Hong Kong. properly the locations and thicknesses of liquefiable

1
2 WANG, FU AND HUANG
qc: MPa fs: kPa CRR, CSR FS
0 15 30 0 40 80 120 160 0 01 02 03 04 05 0 1 2 3
0
Mw = 60 Mw = 60
amax = 0315g amax = 0315g
CRR

FS
2 CSR

Depth: m 4 FS = 1

6 Sand

10

12

OO Fig. 1. A set of real CPT results from Canterbury region, New Zealand (CGD, 2016) Q27

soil layers, particularly when using the scattered test PROBABILISTIC FRAMEWORK FOR LIQUEFIABLE
results from CPTs. For example, Fig. 1 shows a set of SOIL THICKNESS ASSESSMENT
CPT data obtained from a site at Canterbury region, A percentage, P, of the thickness of liquefiable soil layers
Q4 New Zealand (CGD, 2016). The first and second columns over the total thickness of all soil layers at a site is defined in
of Fig. 1 show variation of tip resistance, qc, and sleeve the probabilistic assessment framework, and it is quantified
friction, fs, with depth. Their values are obviously scattered, probabilistically using MCS. The MCS relies on repeatedly
even within a relatively short distance. The third column calculating a mathematical or empirical operator in which
of Fig. 1 shows the distribution of the CRR calculated the variables within the operator are random with prescribed
using the RW method. For a given earthquake magnitude, probability distributions (e.g. Ang & Tang, 2007; Wang,
Mw, and peak ground acceleration, amax, distribution of 2011). For the problem considered herein, the mathematical
the CSR can be estimated using the equation developed by operator involves the determination of the load (i.e. CSR)
Seed & Idriss (1971, see section entitled Estimation of and resistance (i.e. CRR) in soil layers and judgement of
CRR and CSR). The third column of Fig. 1 shows whether or not liquefaction occurs. As illustrated in Fig. 2,
distributions of the CSR when amax = 0315g and Mw = 60 the assessment starts with characterisation of the uncertain-
by a dashed line. The fourth column of Fig. 1 shows the ties in the random variables (e.g. CRR, various input
variation of FS with depth. Since the data are scattered, parameters in the CSR equation, and model uncertainty in
engineering judgement is frequently employed to consider the RW model). Bayesian approaches are developed to
subjectively the effect of various uncertainties (e.g. spatial characterise the spatial variability and probability distri-
variability or correlation of soil properties at different depths) bution of the CRR in each layer. After the characterisation,
PR

and to estimate the thicknesses of liquefiable soil layers n0 sets of random samples of the uncertain variables are
accordingly. generated from the uncertainty quantification results, fol-
A probabilistic approach is developed in this study to lowed by calculation of FS and judgement of whether
objectively identify liquefiable soil layers and assess the liquefaction occurs for all the data points. Then, the per-
liquefiable soil thickness using CPT. The spatial variability of centage, P, of the liquefiable soil thickness in the nth layer is
liquefaction resistance (i.e. CRR), the model uncertainty estimated using the following equation
associated with the CPT-based RW method, and input X
nL
parameter uncertainty in CSR calculation are taken into hi
P 1
consideration explicitly in the proposed approach. It starts i1
Hn
with a framework for probabilistic assessment of liquefiable
soil thickness, followed by a brief description of the RW where nL is the total number of liquefiable CPT data points in
method and simplified procedure for estimating the CRR the nth layer; hi is the average of the vertical intervals
and CSR and their associated uncertainties. Probability between the ith liquefiable CPT data point and its two
distribution of the CRR and the thicknesses of soil layers adjacent points along the vertical direction, respectively; and Q5
with statistically homogeneous CRR are then identified. Hn is the thickness for the nth layer. For the CPT data points
Subsequently, the location and thickness of liquefiable soil collected at an equal vertical interval, h, the percentage,
layers are assessed probabilistically using Monte Carlo P, of the liquefiable soil thickness in the nth layer is
simulations (MCS). The proposed approach is illustrated approximated as
using a set of real CPT data collected from New Zealand. nL  h nL  h nL
Case histories of the 20102011 Canterbury earthquake P 2
Hn h  nt nt
sequence in New Zealand are used to systematically evaluate
the proposed approach. In addition, a sensitivity study is where nt is the total number of CPT data points in the nth
performed to explore the effect of spatial variability on soil layer. Finally, statistical analysis of the P values obtained
liquefaction. from the n0 sets of output for the nth layer is performed to
PROBABILISTIC ASSESSMENT OF LIQUEFIABLE SOIL THICKNESS 3 Q1
Obtain CPT and other necessary data If Ic , 26, the exponent to calculate Q should be changed to
n 05 and Ic is recalculated. If the recalculated Ic remains
less than 26, then the recalculated Q is taken as qclN. On the
Calculate CRR of all data pointsusing Robertson and Wride model
other hand, if Ic iterates above and below the value of 26,
then Q, qclN and Ic are calculated based on n 075.

FS
Quantify relevant uncertainties Subsequently, a correction factor, Kc, is estimated by the
(a) Spatial variability of CRR following equation
(b) Model uncertainty in the Robertson and Wride model 8
(c) Parameter uncertainty in CSR
< 10 for Ic  164
Kc 0403Ic4 5581Ic3  2163Ic2
:
3375Ic  1788 for Ic . 164
Generate n0 sets of CRR and CSR random samples according 4
to the results from the uncertainty quantification
The equivalent clean sand normalised CPT penetration
resistance, (qclN)cs, is calculated as
Calculate FS = CRR/CSR and judge whether liquefaction (FS 1) qclN cs Kc qclN 5
occurs for all the data points and calculate the percentage of
liquefiable soil thickness in each layer using one set of random
Finally, using the (qclN)cs, CRR is estimated using the
samples as input
following simplified equation
8  
>
> qclN cs 3
MCS > 93
< 008 for 50  qclN cs,160
No 1000
Complete n0 times of repetitive CRR  
>
> q
executions of the previous step
>
OO using a different set of random
samples as input every time

Yes

Perform statistical analysis of the resulting n0 sets of output


: 0833 clN cs 005 for qclN cs , 50
1000

The CSR value for a CPT data point (i.e. at a specific depth
during CPT) is calculated using the procedure recommended
6

by Youd et al. (2001) Q6

065 v0 =v0 amax =grd


Fig. 2. Probabilistic framework for assessment of liquefiable soil
CSR 7
MSF=K
thickness
where amax is the peak horizontal ground surface accelera-
tion; g is gravitational acceleration; rd is the shear stress
provide relevant statistics and probability distribution of P. reduction factor; MSF is the magnitude scaling factor; K is
The MCS requires information on spatial variability of the , Pa). The
the overburden correction factor (K = 1, for v0
CRR, model uncertainty in the RW model, and parametric rd, MSF and K are calculated using the following equations,
uncertainty in CSR calculation. The RW model is first used respectively (Youd et al., 2001) Q7
to calculate the CRR values for all CPT data points. Then 
spatial variability of the CRR is characterised, and the mean, 10  000765z for z  915 m
rd 8
standard deviation and correlation length of the CRR are 1174  00267z for 915 m , z  23 m
estimated using Bayesian approaches. The calculation pro-
cedure recommended by Youd et al. (2001) is used to deter- 10224
mine the CSR. The RW model and simplified procedure for MSF 9
PR

estimating the CRR and CSR, their associated model Mw256


uncertainty, parametric uncertainty in CSR calculation,
and the Bayesian characterisation of the CRR are discussed   f 1
v0
in the following four sections. K 10
Pa
where z is the depth of the CPT data point; Mw is the
ESTIMATION OF CRR AND CSR earthquake moment magnitude; f is an exponent that is
In the RW model (Robertson & Wride, 1998), a soil a function of site conditions, including relative density, stress
behaviour type index, Ic, is defined as follows history, aging and overconsolidation ratio. For routine prac-
tice and non-critical projects, f may be taken as 07 (Juang
Ic 347  log Q2 log F 1222 05 3 et al., 2008).
where Q (qc  v0/Pa2)(Pa/v0 ) is the normalised CPT
n

penetration resistance; F 100 [ fs/(qc  v0)] is the normal-


ised friction ratio in percentage; fs is the sleeve friction; v0 PARAMETER UNCERTAINTY IN CSR
are the total and effective overburden stresses, respec-
and v0 CALCULATION
tively; Pa is the atmospheric pressure in the same units as v0 Equations (7)(10) show that CSR is a function of five
(e.g. Pa 100 kPa if v0 is in kPa); Pa2 is the atmospheric input parameters (i.e. amax, Mw, v0 , v0 and rd) and other
pressure in the same units as v0 (e.g. Pa2 01 MPaif v0 is in constants (e.g. f = 07). These five parameters are treated as
MPa). The stress exponent n varies from around 05 for sands random variables. rd is taken to follow a triangular dis-
to 10 for clays. The normalisation for Q is somewhat tribution with its mode (i.e. the rd value with the maximum
complex and requires iteration. First, n 10 is used to probability density function) calculated from equation (8).
calculate Q and Ic. If Ic . 26, then the normalised cone tip The lower and upper bound of the triangular distribution are
resistance corrected for overburden stress, qclN, is set as Q. determined in accordance with the rd ranges suggested by
4 WANG, FU AND HUANG
Table 1. Typical ranges of coefficient of variation (COV) for input Cyclic resistance ratio
parameters in CSR calculation

Random Typical range References Layer 1: 1 = [1 1 1] h1


variable of COV

FS
amax 010020 Haldar & Tang (1979), Espinosa Layer 2: 2 = [2 2 2] h2
(1982), Comartin et al. (1995),
Juang et al. (1999, 2000)
Mw 005010 Juang et al. (1999)

v0 005020 Juang et al. (1999, 2000)
v0 005020 Juang et al. (1999, 2000)

Layer n: n = [n n n] hn
Table 2. Coefficients of correlation among input parameters

Random variable amax Mw


v0 v0

amax 1 09 0 0
Mw 09 1 0 0

v0 0 0 1 09
v0 0 0 09 1

Note: this table is compiled from previous studies reported in the


Layer N: N = [ N N N ]
literature, such as Campbell (1981); Idriss (1991); Juang et al. (1999, hN
2000, 2008).
OO
Seed & Idriss (1971) and recommended by Youd et al. (2001).
The other four parameters (i.e. amax, Mw, v0 and v0) are
taken to follow a lognormal distribution because of their
non-negative physical meaning, as commonly modelled in
the literature (e.g. Juang et al., 2004). Since the mean value of
Fig. 3. An illustration of the soil profile divided into N statistically
homogeneous soil layers based on cyclic resistance ratio (CRR)

uncertainty. Then, the predicted CRR in a log scale is


a parameter is frequently adopted as its nominal value in expressed as
geotechnical practice, for a lognormally distributed random  
CRR
variable with a given mean or nominal value, the only statis- ln CRR ln ln CRR  ln c
c
tical parameter needed to characterise the random variable is
its standard deviation or coefficient of variation (COV). ln CRR T 12
Table 1 summarises the typical ranges of COV for amax, Mw,
and v0 reported in the literature (e.g. Haldar & Tang,
v0 where T represents the model uncertainty. Juang et al. (2004)
1979; Espinosa, 1982; Comartin et al., 1995; Juang et al., calibrated the model uncertainty using a large field obser-
1999, 2000, 2004). In addition, owing to the physical mean- vation database and a Bayesian mapping approach (Juang
ings of amax, Mw, v0 and v0, there are obvious correlations et al., 1999). In their approach, a Bayesian mapping function
between amax and Mw and between v0 and v0, while the PL  fL()/[ fL() fNL ()] was first developed that relates
correlations for the other pairs of random variables are weak the reliability index to the probability of liquefaction PL,
(e.g. Campbell, 1981; Idriss, 1991; Juang et al., 1999, 2000, where fL(), fNL ()) are probability density functions of the
2008). Table 2 shows a matrix of correlation coefficients calculated of the group of liquefied cases and the group of
non-liquefied cases, respectively. The effect of model uncer-
PR

reported and used in the literature (e.g. Juang et al., 2008).


Using the information from Tables 1 and 2, random samples tainty is reflected in the calculated probability because the
of amax, Mw, v0 and v0 can be generated for the MCS shown probability was calibrated through the reliability index based
in Fig. 2. In addition, it is worthwhile to note that, if the on a large database of field observations (Juang et al., 2004).
measured amax data under a specific earthquake are available A first-order reliability method was then used to quantify
at specific sites, the COV values from the measurement data the model uncertainty explicitly using a lognormal random
should be used in MCS to reflect the uncertainty associated variable, c. According to Juang et al. (2004), the mean and
with site response. The amax COV values shown in Table 1 coefficient of variation of c were characterised as 094 and
should only be used when the measured amax data are not 015, respectively. Then, the mean and standard deviation,
available. defined as m and m, of the equivalent normal variable
-lnc were calculated as 0073 and 0149, respectively (i.e. Q8
T  N(0073, 0149)).

MODEL UNCERTAINTY IN THE ROBERTSON AND


WRIDE MODEL BAYESIAN CHARACTERISATION OF CRR
As the RW model was developed based on past obser- Bayesian approaches are developed to characterise spatial
vations and engineering judgement, uncertainties inevitably variability of the CRR. As shown in Fig. 3, the inherent
exist in the model for estimating CRR. When considering the spatial variability of the CRR within the N statistically
model uncertainty in the RW model, the true CRR is given homogeneous layers is modelled by N one-dimensional
by the following equation (Juang et al., 2004) and mutually independent lognormal random fields along
CRR CRR  c 11 the vertical direction (i.e. CRRn (d), n = 1, 2, , N ) (e.g.
Vanmarcke, 1977), where the subscript n denotes the layer
where CRR is the CRR values calculated from the RW number (i.e. the nth layer of soil) and d is the depth. The
method and c is a random variable representing model mean and standard deviation of the CRR for the nth layer are
PROBABILISTIC ASSESSMENT OF LIQUEFIABLE SOIL THICKNESS 5 Q1
n and n, respectively. An exponential correlation function is where P( n|n, N ), n = 1, 2, , N, is the likelihood function
used to model the correlation structure, and the correlation for the nth soil layer. Note that ln,n and ln,n are determined
coefficient (i, j) between the logarithms of the CRR (i.e. from the mean and standard deviation of the corresponding
ln[CRR(di)] andln[CRR(dj )]) at depths di and dj is expressed lognormal distribution of the CRR (i.e. n, n). In addition,
as (e.g. Fenton, 1999a, 1999b; Wang et al., 2010, 2011; Wang the likelihood function in equation (14) is a function of not

FS
& Cao, 2013; Cao & Wang, 2014) only the model parameters for each layer (i.e. N) but also the
  layer thickness vector hN [h1, h2,..., hN]T since the data are
2jdi  dj j divided into N soil layers and used in the equations.
i; j exp  ; n 1; 2; . . .; N 13
n Similar to the likelihood function, the prior distribution
P(N|N ) is given by (e.g. Cao et al., 2016)
where n is the correlation length along the vertical direction,
also known as the vertical scale of fluctuation, within the nth Y
N
layer. PN jN Pn jN 17
Q9 Then, a kn-by-kn correlation matrix Rn, where the (i, j)-th n1
entry is given by equation (13), is constructed to represent where P(n|N ), n = 1, 2, , N, is the prior distribution
the spatial correlation of the lnCRR (i.e. the logarithm of of model parameters n [n n n] in the nth soil layer, for
the CRR within the nth layer) at different depths (i.e. along which a uniform distribution is adopted and expressed as
8
< 1 1 1
  for n [ 0; 06; n [ 0; 0173; n [ 0; 10
Pn jN 06 0173 10 18
:
0 for others
OO
Q10
vertical direction). Let hn, n = 1,2, , N, denote the thickness
of the nth soil layer (See Fig. 3), then the N lognormal
random fields are uniquely defined by a thickness vector
hN [h1, h2, , hN]T and a model parameter matrix N
that consists of N model parameter vectors n = [n n n] and
N correlation matrices Rn.
The prior distribution above is derived based on the lower
and upper bounds of n, n and n (i.e. min, max ; min, max ;
min, max), which can be estimated in accordance with their
respective physical meaning. n and n are the mean and
standard deviation of the lognormal distribution of the CRR
for the nth layer, respectively. The CRR value must be larger
Let [1, 2, , N]T be a set of ln CRR data at different than 0 (i.e. min 0) and should be smaller than 06 (i.e.
depths calculated from the RW model in a soil profile con- max 06) based on the empirical curve recommended for
taining N layers, where n = [n(d1), n(d2), , n(dkn)]T, determining CRR (Robertson & Wride, 1998). In addition,
n = 1,2, , N, is the calculated lnCRR values at the kn n is a standard variation, and it must be larger than 0
depths d1, d2, , dkn within the nth soil layer. In the context (i.e. min 0). The upper bound of n is taken at an extreme
of random field modelling and the transformation model case when the CRR is uniformly distributed between the
given by equation (12), it can be reasoned that n measured range (0, 06), and the corresponding max is estimated as
q
within the nth soil layer is a Gaussian vector with a mean max 06  02 =12 0173 (e.g. Wang et al., 2013). The
(ln,n m)ln and covariance matrix Cn = 2ln,nRn + m 2In, n is the vertical scale of fluctuation for the nth layer,
where ln,n and ln,n are the mean and standard deviation and it must be larger than 0 and is typically smaller than
of the corresponding normal variable lnCRR for the nth 10 m (e.g. Phoon & Kulhawy, 1999).
layer, respectively; ln is a vector with kn components all equal The posterior distribution of model parameters
to one; and In is a kn-by-kn identity matrix (Wang et al., (i.e. N, hN) is then obtained from equation (14) using the
2010). The model parameters (i.e. N, hN) used to define the likelihood function and prior distribution derived. An
N lognormal random fields are unknown and need to be asymptotic technique (Bleistein & Handelsman, 1986) is
estimated. Bayesian system identification and model class used to solve equation (14) for obtaining the posterior infor-
PR

selection methods (e.g. Cao & Wang, 2013; Wang et al., 2013, mation on the model parameter. The most probable value
2014) are developed to estimate these model parameters, as (MPV) of model parameters (i.e. N *, h N*) obtained by
discussed below. minimising an objective function fobj ln[P(N|, N )]
The estimation of model parameters N for N given soil is taken as the posterior mean of the model parameters
layers is formulated as a Bayesian updating process and (e.g. Cao & Wang, 2013; Wang et al., 2013, 2014).
expressed as (e.g. Ang & Tang, 2007; Wang et al., 2016) The number of soil layers (i.e. k) is considered as a constant in
PN j; N KN P jN ; NPN jN 14 the formulation mentioned above. It is further taken as a vari-
able, and a Bayesian model class selection method (e.g. Beck
where KN is a normalising constant; P(|N, N ) is the likeli- & Yuen, 2004; Cao & Wang, 2013; Wang et al., 2013, 2014) is
hood function as given by equations (15) and (16) (Wang used to determine its most probable value (i.e. k*). Let Nmax
et al., 2010; Cao & Wang, 2013), P(N|N ) is the prior denote the maximum possible number of soil layers within the
distribution of model parameters. P(|N, N ) is further depth of a CPT sounding. The most probable model class
expressed as is determined by repeatedly calculating the evidence P(|Mk)
among all model classes Mk, k = 1, 2, , Nmax, and comparing
Y
N
P jN ; N P n jn ; N 15 them. The most probable number of soil layers is the one
n1
with the maximum value of P(|Mk) (Cao & Wang, 2013). The
P(|Mk) can be approximated as (e.g. Papadimitriou et al.,
1997; Beck & Katafygiotis, 1998; Wang et al., 2014)
P n jn ; N 2kn =2 jdet Cn j1=2   1=2
  P jMk  P  ; Mk P jMk 2 jk =2 det H 
1 T 1 k k k
 exp  n  ln;n m l n Cn n  ln;n m l n
2 k 1; 2; . . .; Nmax
16 19
6 WANG, FU AND HUANG

Obtain CPT and earthquake data and choose Nmax Table 3. Bayesian model class selection results

Model ln[P(|Mk)] Most probable thicknesses, hk*: m


Calculate CRR and the CSR for each CPT data point using class Mk
equations (3)(10) h1* h2* h3* h4* h5*

FS
M1 38 10
Compute prior information P(NN) using equations (17) and (18) M2 32 32 68
M3 53 38 12 50
M4* 58 38 12 12 38
M5 43 28 10 12 12 38
Compute the likelihood function using equations (15) and (16) and
construct the objective function fobj = ln[P(N, N)]

The proposed approaches are illustrated using a set of real


Minimise the objective function, determine the MPV of layer CPT data from a site in the Canterbury region, New Zealand
thicknesses and other model parameters, and calculate the (CGD, 2016). As shown in Fig. 1, the site contains a thick
Hessian matrix
layer of sand extending from the ground surface to a depth of
more than 12 m. The groundwater table (GWT) was at about
16 m below the ground surface. It is noted that liquefaction
Calculate the conditional probability P(Mk) for the
only occurs below the GWT, and therefore, the total
kth model class using equation (19)
thickness Ht of the soil profile taken into consideration in
the liquefaction analysis is 100 m (i.e. from a depth of 16 m
to 116 m). The distributions of qc and fs, and the calculated
CRR values, are shown in Fig. 1. The proposed Bayesian
OO Repeat Nmax times for different
model classes?

Yes

Compare P(Mk) and determine the soil layer number, the MPV
No
system identification and model class selection methods are
first used to characterise the spatial variability of CRR,
considering a maximum of five candidate model classes (i.e.
Nmax 5). Then, the CRR characterisation results are used
for probabilistic assessment of liquefiable soil thickness by
MCS. The number of MCS samples should be at least ten
times the reciprocal of the target failure probability (e.g.
of layer thickness, and other model parameters Wang, 2011). In this study, the level of liquefaction prob-
ability of interest is taken as 1  104, and a sample number
of 1 000 000 is used. For illustration purposes, the mean
Perform Monte Carlo simulations and calculate P using equation (2)
values of amax and Mw are taken as 0315g and 60, respec-
(See Fig. 2) tively, for a Canterbury earthquake sequence case history.
Nominal values of v0 and v0 at different depths are adopted
Fig. 4. Flowchart for probabilistic assessment of liquefiable soil as their mean values. The measured amax data at this site
thickness show that the amax standard deviation is 0036g. In addition,
based on Table 1, the COVs of Mw, v0 and v0 are taken as
005, 010 and 015, respectively. Table 2 is also used in MCS
where k* is the MPV of model parameters for model class to model the correlation among amax, Mw, v0 and v0.
Mk; P(|k*, Mk) is the likelihood function of Mk evaluated
at k*; P(k*|Mk) is the prior distribution of Mk evaluated at
k*; H(k*) is the Hessian matrix of Mk evaluated at k*; Bayesian characterisation results
jk = 3k is the number of model parameters of Mk. Details of
PR

Table 3 summarises the logarithm of evidence in the second


the Bayesian model class selection method are provided by column for the five candidate model classes. The value of ln
Cao & Wang (2013).. [P(|Mk)] increases from 38 to 58 as k increases from 1 to 4,
and it then decreases from 58 to 43 as k further increases
from 4 to 5. The model class with four soil layers has the
ILLUSTRATIVE EXAMPLE maximum value of ln[P(|Mk)] among all the five model
Figure 4 shows details of each step and their associated classes. Therefore, the most probable number of soil layers is 4.
equations for the proposed probabilistic assessment of Table 3 also summarises the most probable thicknesses (i.e.
liquefiable soil thickness. Although the proposed approach hk*) of the soil layers for the five candidate model classes.
involves quite a number of equations and seems complex, Using hk*, k = 1, 2, , 5, the most probable boundaries of
it is rather straightforward to program it as a user function soil layers are delineated accordingly for the five model
or toolbox using computer software, such as Matlab classes, as shown in Fig. 5 using short, dashed lines. For the
(Mathworks, Inc., 2016). The user function or toolbox can most probable model class M4*, the most probable thick-
be treated as a black box. Geotechnical engineers only need nesses of these four layers are 38 m, 12 m, 12 m and 38 m,
to provide input, such as CPT data and earthquake para- respectively. The most probable boundaries for M4 are thus at Q11
meters (e.g. amax and Mw). Then, the user function or a depth of 54 m, 66 m and 78 m, respectively. In addition,
toolbox will return the locations and thicknesses of liquefi- the identified MPV of the model parameters for the most
able soil layers and their corresponding percentages of probable model class are summarised in Table 4.
liquefiable soil thickness. This not only allows engineers to An important feature of the proposed method is the
use the proposed approach without being compromised by delineation of the statistically homogeneous CRR layers in
the complex mathematics, but also offers engineers insights an objective and rational manner. Delineation of statistically
into the distribution of liquefiable soil layers along the homogeneous layers from CPT data is often challenging and
vertical direction and the occurrence probability or uncer- involves subjective judgement because of the nature of CPT
tainty of liquefaction along the vertical direction. data, such as significant scatteredness, spatial correlation
PROBABILISTIC ASSESSMENT OF LIQUEFIABLE SOIL THICKNESS 7 Q1
Model class CRR
M1 M2 M3 M*4 M5 0 01 02 03 04 05 06
0

FS
Depth: m 4

10

12
Boundaries identified in Most probable boundaries of
different model classes statistically homogeneous CRR layers

Fig. 5. The most probable boundaries of statistically homogeneous CRR layers


OO
Table 4. Bayesian system identification results

Layer number


Model parameters


Normalised frequency: %
100
90
80
70
60
Layer 1: 1654 m

Mean: 193%
Standard deviation: 101%
Sample number: 1 000 000

1 0226 0072 0592


2 0388 0030 5947 50 168% (deterministic approach)
3 0180 0030 4605 40
4 0255 0040 0521 30
20
100 10
Mean: 192%
90 Standard deviation: 55% 0
Sample number: 1 000 000 0 10 20 30 40 50 60 70 80 90 100
80
Percentage of liquefiable soil thickness, P : %
Normalised frequency: %

70 (a)
60 116% (deterministic approach)
50 Layer 3: 6678 m
50
45
40 Mean: 538%
40
Normalised frequency: %

30 Standard deviation: 248%


35 Sample number: 1 000 000
PR

20
30
10 433% (deterministic approach)
25
0
0 10 20 30 40 50 60 70 80 90 100 20
Percentage of liquefiable soil thickness, P : % 15
10
Fig. 6. A histogram of the percentage, P, of liquefiable soil thickness
5
0
and transition of layers. The proposed Bayesian method in 0 10 20 30 40 50 60 70 80 90 100
this study uses the evidence P(|Mk) (i.e. the probability of Percentage of liquefiable soil thickness, P : %
observing the measured dataset given a model class Mk or (b)
a soil stratification model with k layers in this study) as
the objective index to compare different model classes and Fig. 7. Histograms of the percentage, P, of liquefiable soil thickness
delineate the most probable statistically homogeneous layers in layers 1 and 3: (a) layer 1; (b) layer 3
(e.g. see Table 3 and Fig. 5). It not only effectively eliminates
the need for subjective judgement, allowing engineers with
different subjective judgement to arrive at the same results, Liquefiable soil thickness assessment results
but also it is mathematically sound and rational. In addition, Figure 6 shows a histogram of the percentage P of
note that the effect of a thin layer on CPT data is not liquefiable soil thickness for all soil layers below the GWT
considered in this study, because previous studies have shown (i.e. extending from a depth of 16 m to 116 m) from MCS.
that only liquefiable soil layers with considerable thickness The mean and standard deviation of P are 192% and 55%,
(e.g. greater than 03 m) have contributions in terms of respectively. If the RW model is used deterministically with-
liquefaction severity (e.g. Youd & Perkins, 1987). out consideration of spatial variability, model uncertainty,
8 WANG, FU AND HUANG

FS
OO
Fig. 8. Locations of all CPT data used in this study. Note: The test numbers of CPTs in severe or moderate liquefaction sites are labelled in red;
whereas those in minor or no liquefaction sites are labelled in white
PR

or parameter uncertainty, the percentage of liquefiable soil is calculated as 193% and 101%, respectively. The COV of P
thickness is calculated as 116%. There is considerable differ- value in layer 1 is 101%/193% = 523%, suggesting sub-
ence between the mean value of P from MCS (i.e. 192%) and stantial variability or uncertainty in the P value obtained. In
the P value from the deterministic approach (i.e. 116%). addition, the mean P value of 193% from MCS is larger than
In addition, the proposed approach provides a probabilistic the P value (i.e. 168%) from the deterministic approach.
distribution of P and quantifies the uncertainty in the Fig. 7(b) shows similar variability in the P value for layer 3.
liquefiable soil thickness assessment results obtained. The The deterministic RW model gives a P value of 433%. In
P values from MCS vary from 0% to about 40% with a mean contrast, the P values from MCS vary from about 0% to
value of 192%, and the COV of P is 55%/192% = 286%. 100%, with a mean and standard deviation of 538% and
The variability or uncertainty is significant. In contrast, the 248%, respectively. Significant variability or uncertainty in
deterministic approach only provides an estimated P value of the P value is also observed from MCS results. The mean
116%, and no measurement is possible on the variability or P value of 538% from MCS is larger than the P value
uncertainty in this estimation. (i.e. 433%) from deterministic approach.
Figures 7(a) and 7(b) show the normalised frequency of P
in layer 1 (i.e. at depth 1654 m) and layer 3 (i.e. at depth
6678 m), respectively. Because the MCS results show that CASE HISTORIES OF THE 20102011 CANTERBURY
layers 2 and 4 are unlikely to liquefy, detailed analysis results EARTHQUAKE SEQUENCE IN NEW ZEALAND
are only shown for layers 1 and 3. As shown in Fig. 7(a), the During 20102011, a series of earthquakes struck
deterministic RW approach gives a P value of 168% for Christchurch, New Zealand, including the Mw 71 Darfield
layer 1. However, based on the MCS results, the P value in earthquake on 4 September 2010, Mw 48 aftershock on
layer 1 varies from 0% to 50% with different occurrence 26 December 2010, Mw 62 Christchurch earthquake on
frequencies. The mean and standard deviation of P in layer 1 22 February 2011, two, Mw 60 and Mw 58, respectively,
PROBABILISTIC ASSESSMENT OF LIQUEFIABLE SOIL THICKNESS 9 Q1
qc: MPa qc: MPa qc: MPa qc: MPa qc: MPa
10 20 10 20 10 20 10 20 10 20
0 0 0 0 0
qc
1 1 1 1 1 fs
2 2 qc
2

FS
fs qc 2 2
3 3 qc
qc 3 fs
3 fs 3
4 fs 4
Depth: m

Depth: m

Depth: m

Depth: m

Depth: m
4
5 5 4 4
5
6 6 5 5
6
7 7
7 6 6
8 8
9 9 8 7 7

10 10 9 8 8
0 100 200 0 100 200 0 100 200 0 100 200 0 100 200
fs: kPa fs: kPa fs: kPa fs: kPa fs: kPa
CPT_39 CPT_40 CPT_41 CPT_107 CPT_172

qc: MPa qc: MPa qc: MPa qc: MPa qc: MPa
10 20 10 20 10 20 10 20 10 20
0 0 0 0 0
qc qc qc
1 qc
fs 1 1 1
fs fs 1 fs
OO 2
3
4
2
3
4
2
3
4
qc
fs
2
3
4
2

3
Depth: m

Depth: m

Depth: m

Depth: m

Depth: m
5
6 5 5 5 4
7 6 6 6 5
8
7 7 7
9 6
8 8 8
10
9 9 9 7
11
12 10 10 10 8
0 100 200 0 100 200 0 100 200 0 100 200 300 400 0 100 200
fs: kPa fs: kPa fs: kPa fs: kPa fs: kPa
CPT_177 CPT_282 CPT_283 CPT_311 CPT_314
(a)

Fig. 9. CPT data collected from the Canterbury Geotechnical Database: (a) severe or moderate liquefaction; (b) minor or no liquefaction. Note:
horizontal solid lines in the figures represent the most probable boundaries of statistically homogeneous soil layers identified from CPT data
(continued on next page)

Q12 earthquakes on 13 June 2011, and two, Mw 59 and Mw proposed approach in previous sections (see Fig. 1) relate to
58, respectively, earthquakes on 23 December 2011. These CPT_6. The CRR is first calculated from qc and fs data
PR

earthquakes caused widespread liquefaction, lateral spread- points. Then the proposed Bayesian system identification and
ing and ground surface subsidence. After the earthquake model class selection methods are used to characterise the
on 4 September 2010, the New Zealand Earthquake spatial variability of the CRR. The soil layers identified from
Commission (EQC) carried out extensive geotechnical site the Bayesian characterisation are also shown in Fig. 9 and
investigations in the Canterbury region affected by conse- summarised in Table 5. A MCS run with a sample number of
quential liquefaction. The site investigation data were com- 1 000 000 is subsequently performed for each set of CPT
piled by the Canterbury Earthquake Recovery Authority data. The nominal value of Mw and the mean and standard
(CERA) to establish a searchable repository of geotechnical deviation of amax corresponding to each set of CPT data are
data to the public, called the Canterbury Geotechnical collected from CGD and summarised in Table 5. Based on
Q13 Database (CGD, 2016). CGD includes extensive data on Table 1, the COVs of Mw, v0 and v0 are taken as 005, 010
CPT, soil borehole, shear wave velocity, groundwater data and 015, respectively. The correlations shown in Table 2 for
and so on. It is an online geographic information system amax, Mw, v0 and v0 are also used in MCS.
(GIS)-based database built on the Project Orbit platform, Table 5 summarises the mean and standard deviation of P
and it allows users to interact directly with the database from MCS. For the ten sets of CPT data where severe or
through a map (e.g. see Fig. 8). moderate liquefaction has been observed in the field, the
Twenty sets of CPT data are collected from CGD and used mean P value varies from more than 60% to about 90%.
together with the proposed approach, including ten sets Results from the proposed approach show that a majority of
of data where severe or moderate liquefaction has been the soil below the water level liquefied during the earthquake.
observed (see Fig. 9(a)) and ten sets of data where minor or This is consistent with the field observations of severe or
no liquefaction has been observed (see Fig. 9(b)). Locations moderate liquefaction. However, for the other ten sets of
of the 20 CPTs were shown in Fig. 8. Table 5 summarises CPT data where minor or no liquefaction has been observed
the information collected from the CGD. Figure 9 shows the in the field, the mean P value generally varies from about 0%
distributions of qc and fs. The CPT data used to illustrate the to about 20%. The results suggest that only minor, or even no,
10 WANG, FU AND HUANG
qc: MPa qc: MPa qc: MPa qc: MPa qc: MPa
10 20 10 20 10 20 30 40 10 20 30 40 10 20
0 0 0 0 0
qc qc
1 1 1 1 fs
2 qc
1 fs
2 2 qc
fs qc 2

FS
3 3 fs
3 2
4 fs 3
4
4
Depth: m

Depth: m

Depth: m

Depth: m

Depth: m
5 3
5 4
6 5
6 5 4
7 6
7
8 6
8 7 5
9
8 7
10 9
6
11 10 9 8

12 11 10 9 7
0 100 200 0 100 200 0 100 200 300 400 0 100 200 300 400 0 100 200
fs: kPa fs: kPa fs: kPa fs: kPa fs: kPa
CPT_6 CPT_20 CPT_236 CPT_231 CPT_758

qc: MPa qc: MPa qc: MPa qc: MPa qc: MPa
10 20 30 40 10 20 10 20 10 20 30 40 10 20 30 40
0 0 0 0 0
1 qc qc qc qc
1 1
fs fs 2 fs 1
2 fs
OO 3
4
5
qc
fs
2
3
4
2
3
4
4

6
2

3
Depth: m

Depth: m

Depth: m

Depth: m

Depth: m
5
6 4
6 5 8
7 5
7 6
8 10
8 6
9 7
10 9 12
8 7
11 10
14 8
12 11 9
13 12 10 16 9
0 100 200 0 100 200 0 100 200 0 100 200 300 400 0 100 200
fs: kPa fs: kPa fs: kPa fs: kPa fs: kPa
CPT_477 CPT_1383 CPT_1387 CPT_446 CPT_1484
(b)

Fig. 9. Continued

liquefaction occurred during the earthquake. The results Note that the calculation model for CRR and CSR (e.g.
agree with the field observations of minor or no liquefaction. equations (3)(10)) is not a linear calculation model, and that
PR

The proposed approach successfully identifies liquefiable soil only a linear calculation model will guarantee that the mean
layers and provides results consistent with field observations. of the probabilistic analysis output is identical to the output
In addition, the proposed approach provides the standard of a deterministic analysis with the mean values of input
deviation of the P value and quantifies the uncertainty in the parameters as input. Therefore, it is not surprising to see the
estimated liquefiable soil thickness. The standard deviation P value from the deterministic approach differ from the mean
varies from a few percent to about 14%, indicating significant P value from MCS. To provide a consistent assessment on
uncertainty in the estimated liquefiable soil thickness. liquefiable soil thickness, it is necessary to perform prob-
For comparison, Table 5 also includes the P values abilistic analysis and properly consider the spatial variability,
obtained from the deterministic approach (i.e. without con- model uncertainty and parameter uncertainty, as proposed in
sideration of spatial variability, model uncertainty or par- the present study.
ameter uncertainty). Much higher P values are also obtained
from the deterministic approach for severe or moderate
liquefaction cases than those for minor or no liquefaction EFFECT OF SPATIAL VARIABILITY ON
cases. Although the general trend from the deterministic LIQUEFIABLE SOIL THICKNESS
approach is consistent with that obtained from the proposed To further investigate the effect of spatial variability on
probabilistic approach, the deterministic P values might be liquefiable soil thickness, a sensitivity study with 12 simulated
significantly different from the mean P values. Consider, CRR cases is performed for the first two soil layers identified
for example, CPT_107, for which mean the P value is 78%. in the section entitled Illustrative example (i.e. from a depth
But its deterministic P value is just 59%, which is about of 16 m to 66 m in Fig. 5). The random field model and
20% smaller than the mean P value. In contrast, CPT_314 spatial correlation model described in the previous sections are
has a mean P value of 72%, but its deterministic P value used to simulate the CRR samples, together with the MPV
is 88%, which is about 16% larger than the mean P value. values of mean and standard deviation obtained from the
The deterministic analysis does not provide consistent Bayesian approaches (see Table 4). The mean and standard
results for quantifying the liquefiable soil thickness. deviation of the lognormally distributed CRR for the two
PR
OO
Table 5. Summary of Canterbury earthquake sequence case histories

PROBABILISTIC ASSESSMENT OF LIQUEFIABLE SOIL THICKNESS


Field observation CPT no. Earthquake moment Peak horizontal ground Thickness of statistically Percentage of liquefiable P from deterministic
magnitude, Mw surface acceleration, amax: g homogeneous soil layers: m* soil thickness, P: % approach: %

Mean Standard deviation Mean Standard deviation

Severe or moderate CPT_39 62 036 0034 83 = 12 + 60 + 11 89(90/99/37) 4 88


liquefaction CPT_40 62 0395 0034 81 = 10 + 36 + 18 + 17 85(72/97/95/53) 5 84
CPT_41 62 04 0034 66 = 28 + 14 + 24 64(82/21/69) 13 65
CPT_107 62 0478 0031 62 = 10 + 14 + 24 + 14 78(53/100/91/52) 9 59
CPT_172 62 054 0034 62 = 14 + 10 + 10 + 28 73(96/84/67/59) 14 66
CPT_177 62 057 0039 81 = 10 + 42 + 12 + 17 81(76/98/94/33) 4 77

FS
CPT_282 62 0372 0031 64 = 14 + 40 + 10 84(94/89/46) 8 80
CPT_283 62 0352 0034 79 = 38 + 20 + 10 + 11 76(96/99/37/1) 5 78
CPT_311 62 036 0031 72 = 12 + 40 + 10 + 10 66(84/93/0/11) 7 68
CPT_314 62 034 0036 59 = 10 + 38 + 11 72(43/84/59) 12 88
Minor or no liquefaction CPT_6 60 0315 0036 100 = 38 + 12 + 12 + 38 19(19/0/54/14) 6 12
CPT_20 60 0315 0034 85 = 28 + 44 + 13 19(20/23/0) 11 24
CPT_236 62 0212 0044 86 = 22 + 14 + 24 + 26 18(33/53/5/0) 6 21
CPT_231 62 0216 0044 67 = 34 + 10 + 23 9(17/1/0) 7 11
CPT_758 60 0165 0039 60 = 12 + 34 + 14 8(18/0/21) 6 27
CPT_477 59 0198 0024 105 = 20 + 18 + 42 + 25 14(0/17/16/18) 8 8
CPT_1383 60 0302 0039 86 = 20 + 42 + 24 15(23/0/34) 6 23
CPT_1387 60 0342 0039 58 = 12 + 18 + 14 + 14 21(64/0/0/33) 7 21
CPT_446 59 0193 0025 107 = 36 + 32 + 20 + 19 1(0/0/4/0) 1 0
CPT_1484 59 016 0035 57 = 18 + 12 + 14 + 13 6(0/13/8/5) 4 0

*This column shows the total soil thickness and thickness of each layer obtained from the Bayesian characterisation. For example, 83 = 12 + 60 + 11 means that the total thickness of soil below the
groundwater level is 83 m, and that the soil is divided into three layers with thickness of 12 m, 60 m and 11 m, respectively.
This column represents the mean P values for the total soil thickness and each layer defined in column 6. For example, 89(90/99/37) means that the mean P value of the total soil thickness is 89%, and that
the mean P values for all the layers defined in column 6 are 90%, 99% and 37%, respectively.
This column represents the standard deviation of P for the total soil thickness.

11 Q1
12 WANG, FU AND HUANG
100 40

90 = 100 m
=5m
80
Normalised frequency: %

= 059 m 30
70 = 001 m

FS
Mean of P: %
60
50 20
40
Layer 1
30
10 Layer 2
20
10
0 0
0 10 20 30 40 50 60 70 80 90 100 001 01 1 10 100
Percentage of liquefiable soil thickness, P: % Correlation length: m
(a) (a)

100 30

90 = 100 m
80 = 595 m
Normalised frequency: %

= 05 m

Standard deviation of P: %
70 = 001 m 20
60
OO 50
40
10 Layer 1
30
20 Layer 2

10
0 0
0 10 20 30 40 50 60 70 80 90 100 001 01 1 10 100
Percentage of liquefiable soil thickness, P: % Correlation length: m
(b) (b)

Fig. 10. Effect of spatial variability on the percentage, P, of Fig. 11. Relationship between correlation length and the statistics of
liquefiable soil thickness: (a) layer 1; (b) layer 2 the percentage, P, of liquefiable soil thickness: (a) mean of P;
(b) standard deviation of P

layers are taken as 0226, 0072 and 0388, 0030, respectively.


To explore the effect of spatial variability, the correlation standard deviation of P is significantly influenced by the
length in these 12 different cases varies over a large range, spatial variability or correlation. In this case, ignoring spatial
including: 001 m (i.e. approximated as 0 for a mutually correlation (i.e. 0) leads to underestimation of the stan-
independent case), 003 m, 01 m, 05 m, 10 m, 15 m, 20 m, dard deviation of P and the occurrence probability of extreme
25 m, 5 m, 10 m, 30 m, 100 m (i.e. approximated as for a cases (e.g. P [ [0%, 10%] for the almost no liquefaction case
fully correlated case), and their respective MPV values of or P [ (90%, 100%] for a case where almost all soils liquefy).
obtained from Bayesian approaches (i.e. 0592 m for layer 1 Such an effect is similar to the effect of spatial variability on
PR

and 5947 m for layer 2). A number of MCS runs are per- the failure probability of geotechnical structures, as reported
formed to assess the thickness of liquefiable soil in each case in the literature (e.g. Wang et al., 2011; Jiang et al., 2015).
with a MCS sample number of 1 000 000. Fig. 10 shows the It should be noted that, in layer 1, the mean values of the
normalised frequency of P in these two layers for the four CRR and CSR profiles are close to each other (i.e. both at
different correlation lengths. about 0102; see Fig. 1 and Table 4). In contrast, Fig. 10(b)
Figure 10(a) shows the results for layer 1. The P values shows the results for layer 2, where the mean values of CRR
spread over a wide range from 0% to 100%, indicating large and the CSR are quite different. In layer 2, the mean CSR
variability in the assessment results. When (i.e. the value of about 0102 (see Fig. 1) is significantly smaller
fully correlated case), the largest normalised frequency is than the mean CRR value of 0388 (see Table 4) and lies in
more than 50% for P [ [0%, 10%] (i.e. almost no liquefaction the probability distribution tail of the CRR. The probability
occurs in layer 1). As the value decreases or the correlation that CRR is less than CSR is therefore relatively small,
within the soil layer diminishes, the normalised frequency despite the spatial variability. Fig. 11 also includes the vari-
for P [ [0%, 10%] decreases and the normalised frequencies ation of mean and standard deviation of P with correlation
for P [ [10%, 40%] gradually increase. When 0 (i.e. the length for layer 2. It is obvious that the effect of spatial vari-
mutually independent case), the largest normalised fre- ability is minimal for layer 2. The effect of spatial variability
quency is about 50% for P [ (10%, 20%]. is significant only when the mean values of CRR and CSR
Figure 11 shows the variation of the mean and standard are close to each other, such as layer 1.
deviation of P with correlation length. For layer 1, the mean
P value remains constant at about 19%; while the standard
deviation increases steadily from 7% to 23% as the correla-
tion length increases from 001 m to 100 m. Because the SUMMARY AND CONCLUSIONS
spatial variability is modelled as a stationary random field This paper has developed a probabilistic approach for
in this study, it is expected to have no effect on the mean assessment of liquefiable soil thickness using CPT data and
P value, as shown in Fig. 11(a). On the other hand, the considering spatial variability and model and parameter
PROBABILISTIC ASSESSMENT OF LIQUEFIABLE SOIL THICKNESS 13 Q1
uncertainties. The CRR was first calculated from the CPT qc cone tip resistance measured from CPT
data at different depths using the RW model. Then, Bayesian qclN normalised cone penetration resistance corrected for
system identification and model class selection methods were overburden stress in the Robertson and Wride model
(qclN)cs equivalent clean sand normalised CPT penetration
developed to characterise probabilistically the CRR and the resistance in the Robertson and Wride model
thicknesses of statistically homogeneous layers. The inherent Rn

FS
spatial variability and model and parameter uncertainties rd stress reduction coefficient in CSR calculation to account
were incorporated explicitly in the proposed approach. MCS for flexibility in soil profile
was performed to quantify the percentage P of liquefiable soil z depth below ground surface (m)
thickness for the assessment of liquefaction consequences. reliability index
The proposed approach was illustrated using a set of real hi average vertical interval between the ith liquefiable
CPT data collected from New Zealand and was further vali- CPT data point and its two adjacent points along the
dated using relevant CPT data and earthquake field observ- vertical direction
T model uncertainty
ations from the 20102011 Canterbury seismic sequence in
N model parameters for all soil layers
New Zealand. The proposed approach properly identifies the n model parameters for nth soil layer
thicknesses of the liquefiable soil layers, quantifies their n
liquefiable soil thickness and associated uncertainty, and n correlation length along the vertical direction within the
provides results that are consistent with earthquake field nth soil layer
observations. In addition, it was shown that the P value from ln,n mean of logarithms of the CRR within the nth soil layer
deterministic analysis might be either much larger or smaller m mean of model uncertainty in the Robertson and
than the mean P value from probabilistic analysis. The Wride model
deterministic approach not only cannot quantify the uncer- n mean of the CRR within the nth soil layer
tainty in the assessment results, but also fails to provide
(i, j) correlation coefficient between the logarithms of the CRR
consistent results for the assessment of liquefiable soil thick- at depths di and dj
OO ness. It is therefore necessary to perform probabilistic
analysis in engineering practice. The approach proposed in
this study offers a rational vehicle to properly and explicitly
consider the spatial variability and model and parameter
uncertainties in the assessment of liquefiable soil thickness.
A sensitivity study is also performed to explore the effect of
the spatial variability on the liquefiable soil thickness. It is
ln,n

n
v0

v0
standard deviation of the logarithms of the CRR within
the nth soil layer
standard deviation of the model uncertainty in the
Robertson and Wride model
standard deviation of the CRR within the nth soil layer
total vertical overburden stress of soil
effective vertical overburden stress of soil
found that, when the mean values of the CRR and CSR are
close to each other, spatial variability has a significant effect
on the variability or uncertainty in the assessment results (i.e. REFERENCES
the standard deviation of P); but it has a minimal effect on Andrus, R. D. & Stokoe, K. H. (2000). Liquefaction resistance
the mean value of P. In this case, ignoring spatial correlation of soils from shear wave velocity. J. Geotech. Geoenviron. Engng
126, No. 11, 10151025.
leads to underestimation of the standard deviation of P and Ang, A. H. S. & Tang, W. H. (2007). Probability concepts in
the occurrence probability of extreme cases, such as the engineering: emphasis on applications to civil and environmental
almost no liquefaction case and a case where almost all soils engineering. New York, NY, USA: John Wiley and Sons.
Q14 liquefy. Baker, J. & Faber, M. (2008). Liquefaction risk assessment using
Geostatistics to account for soil spatial variability. J. Geotech.
Geoenviron. Engng 134, No. 1, 1423.
Beck, J. L. & Katafygiotis, L. S. (1998). Updating models and their
Q15 NOTATION uncertainties. I: Bayesian statistical framework. J. Engng Mech
amax peak horizontal ground surface acceleration 124, No. 4, 455461.
Cn Beck, J. L. & Yuen, K. V. (2004). Model selection using response
c random variable representing model uncertainty measurements: Bayesian probabilistic approach. J. Engng Mech.
PR

F normalised friction ratio in the Robertson and Wride 130, No. 2, 192203.
model Bleistein, N. & Handelsman, R. (1986). Asymptotic expansions of
f exponent estimated from site conditions used in integrals. New York, NY, USA: Dover.
calculation of K in CSR calculation Campbell, K. W. (1981). Near-surface attenuation of pear ground
fs sleeve friction measured from CPT acceleration. Bull. Seismol. Soc. Am. 71, No. 6, 20392070.
g gravitational acceleration Cao, Z. & Wang, Y. (2013). Bayesian approach for probabilistic
Q16 H Hessian matrix site characterization using cone penetration tests. J. Geotech.
Hn thickness for the nth soil layer Geoenviron. Engng 139, No. 2, 267276.
hN thickness vector Cao, Z. & Wang, Y. (2014). Bayesian model comparison and
Ic soil behavior type index based on CPT data selection of spatial correlation functions for soil parameters.
In Struct. Safety 49, 1017. Q17
Kc correction factor for grain characteristics in the Cao, Z., Wang, Y. & Li, D. (2016). Quantification of prior
Robertson and Wride model knowledge in geotechnical site characterization. Engng Geol.
K correction factor for soil layers subjected to large static 203, 107116. Q18
normal stresses in CSR calculation Cetin, K. O., Der Kiureghian, A. & Seed, R. B. (2002). Probabilistic
ln models for the initiation of seismic soil liquefaction. Struct.
Mw earthquake moment magnitude Safety 24, No. 1, 6782.
Nmax maximum possible number of soil layers within the CGD (Canterbury Geotechnical Database) (2016). https://
depth of a CPT sounding canterburygeotechnicaldatabase.projectorbit.com (accessed
nL total number of liquefiable CPT data points in the nth layer 00/00/2016). Q19
nt total number of CPT data points in the nth layer Comartin, C. D., Greene, M. & Tubbesing, S. K. (1995). The
P percentage of liquefiable soil thickness Hyogo-Ken Nanbu earthquake preliminary reconnaissance report,
Pa atmospheric pressure in same units as v0 EERI report No. 95-04. Oakland, CA, USA: Earthquake
Pa2 atmospheric pressure in same units as v0 Engineering Research Institute. Q20
Q normalised CPT penetration resistance in the Robertson Espinosa, A. F. (1982). ML and MO determination from strong-
and Wride model motion accelerograms, and expected intensity distribution.
14 WANG, FU AND HUANG
In The Imperial Valley, California, earthquake of October 15, Mathworks, Inc. (2016). MATLAB the language of technical
1979, geological survey professional paper 1254, pp. 433438. computing. Mathworks, Inc.. See http://www.mathworks.
Washington, DC, USA: United States Government Printing com/products/matlab/ (accessed 08/08/2016). Q24
Q21 Office. Moss, R. E. S., Seed, R. B., Kayen, R. E., Stewart, J. P., Der
Fenton, G. (1999a). Estimation for stochastic soil models. Kiureghian, A. & Cetin, K. O. (2006). CPT-based probabilistic
J. Geotech. Geoenviron. Engng 125, No. 6, 470485. and deterministic assessment of in situ seismic soil liquefaction

FS
Fenton, G. (1999b). Random field modeling of CPT data. potential. J. Geotech. Geoenviron. Engng 132, No. 8, 10321051.
J. Geotech. Geoenviron. Engng 125, No. 6, 486498. Papadimitriou, C., Beck, J. L. & Katafygiotis, L. S. (1997).
Haldar, A. & Tang, W. H. (1979). Probabilistic evaluation Asymptotic expansions for reliability and moments of uncertain
of liquefaction potential. J. Geotech. Engng 104, No. 2, systems. J. Engng Mech. 123, No. 12, 12191229.
145162. Phoon, K. K. & Kulhawy, F. H. (1999). Characterization of
Idriss, I. M. (1991). Earthquake ground motions at soft soil sites. geotechnical variability. Can. Geotech. J. 36, No. 4, 612624.
In Proceedings of the 2nd international conference on recent Robertson, P. K. & Campanella, R. G. (1985). Liquefaction
advances in geotechnical earthquake engineering and soil potential of sands using the CPT. J. Geotech. Geoenviron.
dynamics, vol. 3, pp. 22652271. Rolla, MO, USA: University Engng 111, No. 3, 384403.
Q22 of Missouri-Rolla. Robertson, P. K. & Wride, C. E. (1998). Evaluating cyclic
Iwasaki, T., Tokida, K., Tatsuoka, F., Watanabe, S., Yasuda, S. & liquefaction potential using the cone penetration test. Can.
Sato, H. (1982). Microzonation for soil liquefaction potential Geotech. J. 35, No. 3, 442459.
using simplified methods. Proceedings of the 3rd international Seed, H. B. (1979). Soil liquefaction and cyclic mobility evaluation
earthquake microzonation conference, Seattle, WA, USA, for level ground during earthquakes. J. Geotech. Engng Div. 105,
pp. 13191330. No. 2, 201255.
Jefferies, M. G. & Been, K. (2006). Soil liquefaction a critical state Seed, H. B. & Idriss, I. M. (1971). Simplified procedure for
approach. New York, NY, USA: Taylor & Francis Group. evaluating soil liquefaction potential. J. Soil Mech. Found.
Jiang, S., Li, D., Cao, Z., Zhou, C. & Phoon, K. (2015). Efficient Div., ASCE 97, No. SM9, 12491273.
system reliability analysis of slope stability in spatially variable Seed, H. B. & Idriss, I. M. (1982). Ground motions and
OO soils using Monte Carlo simulation. J. Geotech. Geoenviron. soil liquefaction during earthquakes. Oakland, CA, USA:
Engng 141, No. 2, 04014096. Earthquake Engineering Research Institute.
Juang, C. H., Rosowsky, D. V. & Tang, W. H. (1999). Stark, T. & Olson, S. (1995). Liquefaction resistance using CPT
Reliability-based method for assessing liquefaction potential of and field case histories. J. Geotech. Geoenviron. Engng 121,
soils. J. Geotech. Geoenviron. Engng 125, No. 8, 684689. No. 12, 856869.
Juang, C. H., Chen, C. J., Rosowsky, D. V. & Tang, W. H. (2000). Toprak, S. & Holzer, T. L. (2003). Liquefaction potential index: field
CPT-based liquefaction analysis. Part 2. Reliability for design. assessment. J. Geotech. Geoenviron. Engng 129, No. 4, 315322.
Gotechnique 50, No. 5, 593599, http://dx.doi.org/10.1680/ Toprak, S., Holzer, T. L., Bennett, M. J. & Tinsley, J. C. (1999).
geot.2000.50.5.593. CPT and SPT-based probabilistic assessment of liquefaction
Juang, C. H., Jiang, T. & Andrus, R. D. (2002). Assessing potential. Proceedings of 7th USJapan workshop on earthquake
probability-based methods for liquefaction potential evaluation. resistant design of lifeline facilities and countermeasures
J. Geotech. Geoenviron. Engng 128, No. 7, 580589. against liquefaction, Seattle, WA, USA, Technical Report
Juang, C. H., Yuan, H. M., Lee, D. H. & Lin, P. S. (2003). Simplified MCEER-99-0019. Q25
cone penetration test-based method for evaluating liquefaction Vanmarcke, E. H. (1977). Probabilistic modeling of soil profiles.
resistance of soils. J. Geotech. Geoenviron. Engng 129, No. 1, J. Geotech. Engng 103, No. 11, 11271246.
6680. Wang, Y. (2011). Reliability-based design of spread foundations by
Juang, C. H., Yang, S. H., Yuan, H. & Khor, E. H. (2004). Monte Carlo simulations. Gotechnique 61, No. 8, 677685,
Characterization of the uncertainty of Robertson and Wride http://dx.doi.org/10.1680/geot.10.P.016.
model for liquefaction potential evaluation. Soil Dynamics Wang, Y. & Cao, Z. (2013). Expanded reliability-based design
Earthquake Engng 24, No. 9, 771780. of piles in spatially variable soil using efficient Monte Carlo
Juang, C. H., Fang, S. Y. & Li, D. L. (2008). Reliability analysis simulations. Soils Found. 53, No. 6, 820834.
of liquefaction potential of soils using standard penetration Wang, Y., Au, S. K. & Cao, Z. (2010). Bayesian approach for
test. In Reliability-based design in geotechnical engineering (ed. probabilistic characterization of sand friction angles. Engng
Q23 K. K. Phoon), Ch. 13, pp. 497526. Abingdon, Oxfordshire, Geol. 114, No. 34, 354363.
UK: Taylor & Francis. Wang, Y., Cao, Z. & Au, S. K. (2011). Practical reliability analysis of
PR

Juang, C. H., Ching, J. & Luo, Z. (2013). Assessing SPT-based slope probability by advanced Monte Carlo simulations in
probabilistic models for liquefaction potential evaluation: a spreadsheet. Can. Geotech. J. 48, No. 1, 162172.
10-year update. Georisk 7, No. 3, 137150. Wang, Y., Huang, K. & Cao, Z. (2013). Probabilistic identification
Ku, C. S., Juang, C. H., Chang, C. W. & Ching, J. (2012). of underground soil stratification using cone penetration tests.
Probabilistic version of the Robertson and Wride method for Can. Geotech. J. 50, No. 7, 766776.
liquefaction evaluation: development and application. Can. Wang, Y., Huang, K. & Cao, Z. (2014). Bayesian identification of
Geotech. J. 49, No. 1, 2744. soil strata in London Clay. Gotechnique 64, No. 3, 239246,
Lai, S. Y., Hsu, S. C. & Hsieh, M. J. (2004). Discriminant model http://dx.doi.org/10.1680/geot.13.T.018.
for evaluating soil liquefaction potential using cone pene- Wang, Y., Cao, Z. & Li, D. (2016). Bayesian perspective on
tration test data. J. Geotech. Geoenviron. Engng 130, No. 12, geotechnical variability and site characterization. Engng Geol.
12711282. 203, 117125. Q26
Lai, S. Y., Chang, W. J. & Lin, P. S. (2006). Logistic regression model Youd, T. L. & Perkins, D. M. (1987). Mapping of liquefaction
for evaluating soil liquefaction probability using CPT data. severity index. J. Geotech. Engng 113, No. 11, 13741392.
J. Geotech. Geoenviron. Engng 132, No. 6, 694704. Youd, T. L., Idriss, I. M., Andrus, R. D., Arango, I., Castro, G.,
Maurer, B., Green, R., Cubrinovski, M. & Bradley, B. (2014). Christian, J. T., Dobry, R., Liam Finn, W. D., Harder, L. F. Jr,
Evaluation of the liquefaction potential index for assessing Hynes, M. E., Ishihara, K., Koester, J. P., Laio, S. S. C.,
liquefaction hazard in Christchurch, New Zealand. J. Geotech. Marcuson, W. F. III, Martin, G. R., Mitchell, J. K., Moriwaki,
Geoenviron. Engng 140, No. 7, 04014032. Y., Power, M. S., Robertson, P. K., Seed, R. B. & Stokoe, K. H.
Maurer, B., Green, R., Cubrinovski, M. & Bradley, B. (2015). (2001). Liquefaction resistance of soils: summary report from
Assessment of CPT-based methods for liquefaction evaluation the 1996 NCEER and 1998 NCEER/NSF workshops on
in a liquefaction potential index framework. Gotechnique evaluation of liquefaction resistance of soils. J. Geotech.
65, No. 5, 328336, http://dx.doi.org/10.1680/geot.SIP.15.P.007. Geoenviron. Engng 127, No. 10, 817833.
QUERY FORM

Institution of Civil Engineers (ICE)

Journal Title: (JGEOT)


Article No: 15P219
AUTHOR: The following queries have arisen during the editing of your manuscript. Please answer the queries by making the
requisite corrections at the appropriate positions in the text.

Query No. Nature of Query Authors Response


Q1 Please check shortened title for running head
Q2 Please provide up to six keywords selected only from the approved
list of terms for this journal. See http://www.icevirtuallibrary.com/
pb-assets/for%20authors/GeokeywordlistJune2016.pdf
Q3 Please confirm the change of spelling from Stark & Olsen (1995)
to Stark & Olson (1995) in text citations as per the reference list.
Q4 URL (here and elsewhere) has been removed from text and added
to reference list, with citation in accordance with preferred
journal style please check.
Q5 Can respectively be deleted here? It is not clear what this is
respective to?
Q6 Please check layout/denominator of equation (7). It was not clear
from original manuscript how the division should be displayed.
Q7 Please check equation (8) m has been interpreted as m for
metres, OK?
Q8 Please clarify whether minus lnc is intended here, i.e. lnc. The
symbol used in the manuscript seems like a hyphen and its
meaning is unclear.
Q9 Kn-by-kn has been changed to kn-by-kn , OK?
Q10 Bold italic font has been used to denote vector throughout, not
underscore, please confirm this is OK and check throughout that
these are correctly identified.
Q11 Is respectively needed in this sentence and the next? Not clear,
respectively to what specifically?
Q12 Again here respectively is used twice in this sentence but it is not
clear respective to what? Can these be deleted?
Q13 Again the web address has been removed from the text here and
replaced by reference citation, please check.
Q14 Acknowledgements section has been deleted in accordance with
journal preferred style, which is not to include thanks to reviewers.
Q15 Notation has been added. Please check carefully and amend/com-
plete with definitions where necessary.
Q16 Bold upright font has been used to denote matrix and bold italic for
vector throughout, not underscore, please confirm this is OK and
check throughout that these are correctly identified.
Q17 Please provide the issue number in Cao and Wang (2014).
Q18 Please provide the issue number in Cao et al. (2016).
Q19 CGD (2016) added as reference so that URLs can be removed from
the text, as per preferred journal style. Please check/amend as
necessary, and provide last accessed date.
Q20 Please check city/town and country names added by copy editor in
Comartin et al. (1995).
Q21 Please provide editor names, if any, in Espinosa (1982).
Q22 Please provide editor names in Idriss (1991).
Q23 Please check chapter no., page range, publisher name, publisher
location details added by copy editor for Jung et al. (2008).
Q24 Please provide company location i.e. city/town, state (if USA) and
country details in Mathworks, Inc. (2016).
Q25 Please provide page range if possible in Toprak et al. (1999).
Q26 Please provide the issue number in Wang et al. (2016).
Q27 Fig. 1. A set of real CPT has been changed to A set of real CPT
results OK? Also, reference has been cited instead of supplying
web address, in accordance with preferred journal style. Please
check.

Você também pode gostar