0% found this document useful (0 votes)
19 views9 pages

(Wang, 2016) Bayesian Perspective

Uploaded by

robozinho324
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
19 views9 pages

(Wang, 2016) Bayesian Perspective

Uploaded by

robozinho324
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

Engineering Geology 203 (2016) 117–125

Contents lists available at ScienceDirect

Engineering Geology

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

Bayesian perspective on geotechnical variability and site characterization


Yu Wang a,⁎, Zijun Cao b, Dianqing Li b
a
Department of Architecture and Civil Engineering, City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong
b
State Key Laboratory of Water Resources and Hydropower Engineering Science, Key Laboratory of Rock Mechanics in Hydraulic Structural Engineering, Ministry of Education, Wuhan University,
8 Donghu South Road, Wuhan, PR China

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

Article history: Geotechnical characterization of a project site for engineering applications is indispensable in engineering geol-
Received 21 May 2015 ogy and geotechnical engineering, and there are many unavoidable variabilities and uncertainties during charac-
Received in revised form 7 August 2015 terization of a project site. Different variabilities and uncertainties usually are lumped together and observed as
Accepted 11 August 2015
the total variability, which includes both the actual variability of soil and rock properties and other knowledge
Available online 15 August 2015
uncertainties, such as measurement errors and statistical uncertainty. It is the actual variability, not the total var-
Keywords:
iability, which affects directly the observed performance (i.e., actual response) of geotechnical and geological sys-
Geotechnical site characterization tems and is of primary interest in site characterization. This paper aims to consolidate recent advancement in
Bayesian methods Bayesian studies in site characterization and develops a Bayesian inverse analysis framework for direct quantifi-
Inherent variability cation of the actual variability of various soil and rock properties. To facilitate development of the framework, the
Uncertainty procedure of geotechnical site characterization is revisited from a Bayesian perspective, and the occurrence and
Soil and rock properties propagation of inherent variability, statistical uncertainty, measurement errors, and transformation uncertainty
during characterization of a project site are mapped explicitly to different stages in site investigation. Based on
the mapping, a robust framework is developed that streamlines the formulation of likelihood functions for vari-
ous soil and rock properties when estimated using different field or laboratory tests, leading to a streamlined pro-
cess for applications of the proposed Bayesian framework to different site characterization problems. Application
examples are provided to illustrate the implementation and step-by-step procedures of the proposed Bayesian
framework.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction inference is made and based on some samples, rather than all samples,
of a population, statistical uncertainty occurs in the inferred information
Geotechnical characterization of a project site for engineering appli- on the whole population. Although these uncertainties have been well-
cations is indispensable in engineering geology and geotechnical recognized in the literature, the occurrence and propagation of these
engineering. It is well-recognized that there are many unavoidable var- uncertainties during site characterization have not been explored sys-
iabilities and uncertainties in site characterization, such as inherent var- tematically or mapped explicitly to different stages in site investigation.
iability of soil and rock properties, measurement errors, statistical These uncertainties are usually lumped together as total variability in
uncertainty, and transformation uncertainty when the information of site investigation and used subsequently in reliability analysis of geo-
interest is not measured directly, but estimated through a transforma- technical and geological systems.
tion model and other measured information (e.g., Whitman, 1984; Because the total variability includes the actual variability of soil and
Phoon and Kulhawy, 1999a, 1999b; Baecher and Christian, 2003; rock properties (i.e., inherent variability) and other additional uncer-
Wang et al., 2010a; Cao and Wang, 2013). The inherent variability is tainties (e.g., measurement errors and statistical uncertainty), it is al-
also known as the actual variability since soil and rock are natural mate- ways larger than the actual variability of soil and rock properties itself.
rials, and their properties vary spatially and are affected by various geo- When the total variability is used in reliability analysis for estimating
logical process that they have undergone in their geological histories. failure probability of geotechnical and geological systems, the predicted
Measurement errors include both systematic bias and random errors as- failure probability in reliability analysis will be different from the failure
sociated with the measurement process. In addition, when statistical rate observed in practice. Note that it is the actual variability of soil and
rock properties, not the total variability, which affects the observed per-
formance of geotechnical and geological systems. In contrast, the pre-
⁎ Corresponding author.
dicted failure probability is calculated using the total variability that is
E-mail addresses: [email protected] (Y. Wang), [email protected] (Z. Cao), larger than the actual variability of soil and rock properties. This leads
[email protected] (D. Li). to two important and unresolved problems in geotechnical risk and

http://dx.doi.org/10.1016/j.enggeo.2015.08.017
0013-7952/© 2015 Elsevier B.V. All rights reserved.
118 Y. Wang et al. / Engineering Geology 203 (2016) 117–125

reliability. Christian and Baecher (2011) identified some stubbornly in- 2. Geotechnical site characterization and associated variability and
tractable issues in geotechnical risk and reliability and developed a list uncertainties
of ten important and unresolved problems in this area. The top two
problems in their list are: Geotechnical site characterization aims to estimate soil and rock
properties and to delineate underground stratigraphy (including the
“1. Why are failures less frequent than our reliability studies number and boundaries/thicknesses of soil layers) for geotechnical or
predict?” geological engineering applications (e.g., Clayton et al., 1995; Mayne
et al., 2002). It is a multi-step process that can be divided into six stages:
“2. What is the actual variability of soil and rock properties?”
desk-study, site reconnaissance, in-situ investigation, laboratory testing,
This paper addresses the problems above and presents a Bayesian in- interpretation of site observation data, and inferring soil and rock prop-
verse analysis framework for directly quantifying the actual variability erties and underground stratigraphy.
of soil and rock properties. The quantification of actual variability is for- Fig. 1 shows schematically a general procedure of geotechnical site
mulated as an inverse analysis problem, in which the observation data is characterization using textboxes with solid lines. Geotechnical charac-
used as input to an inverse analysis model for inferring the actual vari- terization of a project site often starts with desk-study and site recon-
ability of soil and rock properties as the model output. The inverse anal- naissance, which provide prior knowledge about the site such as
ysis model explicitly depicts the occurrence of inherent variability, geological maps, regional guides, soil survey maps and records, pub-
measurement errors, statistical uncertainty, and transformation uncer- lished reports and studies, engineering judgment (or expertise) and/
tainty during site characterization and their propagation towards the or engineering experience (e.g., Trautmann and Kulhawy, 1983;
total variability that used to be estimated directly from observation Clayton et al., 1995; Mayne et al., 2002). After desk-study and site recon-
data in geotechnical literature and practice. In addition to the project- naissance, in-situ investigation (e.g., in-situ boring and testing) and lab-
specific observation data, the model relies on site information available oratory testing are performed to obtain project-specific test results
prior to the project (e.g., engineering judgment/engineering experi- (i.e., site observation data) of a soil or rock property XM measured by
ence), which is referred to as “prior knowledge” under the Bayesian the tests. Note that the measured soil or rock property XM is not neces-

E
framework. The information provided by project-specific observation sarily the design property XD that is used directly in geotechnical or geo-
data is integrated with prior knowledge through a so-called “likelihood logical analysis. The design property XD (e.g., undrained Young's
function” to provide updated or “posterior information” of interest in modulus Eu) can be estimated from the test data of XM (e.g., the “N”
the Bayesian framework (e.g., Ang and Tang, 2007). values) obtained from in-situ/laboratory testing (e.g., standard penetra-
Bayesian framework has been used in geotechnical engineering tion test (SPT)) through a transformation model MT (e.g., empirical re- MAX
(e.g., Zhang et al., 2004; Najjar and Gilbert, 2009; Ching et al., 2010; gressions and theoretical relationships) between XD and XM. Engineers
Wang et al., 2010b; Chiu et al., 2012; Juang et al., 2013; Medina-Cetina then utilize both the results interpreted from project-specific site obser-
and Esmailzadeh, 2014), including site investigation (e.g., Wang et al., vation data and prior knowledge to estimate soil and rock properties
2010a; Cao et al., 2011; Wang and Cao, 2013; Cao and Wang, 2014a, and underground stratigraphy for engineering applications.
2014b; Wang and Aladejare, 2015). However, Bayesian method has Neither the project-specific site observation data nor prior knowl-
one key limitation: its mathematical and computational complexity edge is perfect information. They are associated with various uncer-
(e.g., Zhang et al., 2009). To bypass this complexity, early Bayesian stud- tainties, such as inherent variability, measurement errors, statistical
ies in site characterization (e.g., Tang, 1971; Grigoriu et al., 1987) have uncertainty, and transformation uncertainty, as shown in Fig. 2
used conjugate pairs of distributions among likelihood function, prior (e.g., Christian et al., 1994; Kulhawy, 1996; Phoon and Kulhawy,
and posterior distributions, which represent prior knowledge and pos- 1999a, 1999b). The inherent variability exists long before the engineer-
terior information, respectively. This nevertheless introduces artificial ing project is planned, and it is usually categorized as “natural variabil-
limitations to the choice of prior distribution and formulation of likeli- ity” (or “aleatory uncertainty”), which is independent of the state of
hood function. Ideally, the prior distribution shall only reflect the prior knowledge about the soil and rock properties and cannot be reduced
knowledge on the project site and not be affected by the limited choices as the knowledge improves (Baecher and Christian, 2003; Christian,
of conjugate prior (e.g., normal distribution) due to mathematical con- 2004).
venience. Similarly, likelihood function that reflects occurrence proba- On the other hand, measurement errors, statistical uncertainty, and
bility of the observed data shall not be simply assumed to follow a transformation uncertainty are resulted from imperfect test equipments
mathematically convenient distribution (e.g., normal distribution), but and/or procedural-operator errors, lack of test data, and insufficient
needs to reflect physical insights of how the observation data are gener- knowledge about the relationship between XD and XM, respectively
ated and how different uncertainties occur and propagate during site (e.g., Kulhawy, 1996; Baecher and Christian, 2003). They are usually cat-
characterization. Some alternatives, such as Laplace asymptotic approx- egorized as “knowledge uncertainty” (or “epistemic uncertainty”) and
imation (e.g., Wang et al., 2010b) and Markov Chain Monte Carlo simu- can be reduced gradually as the knowledge improves (Baecher and
lation (MCMCS) (e.g., Zhang et al., 2012; Wang and Cao, 2013; Peng Christian, 2003; Christian, 2004). Although the natural variability and
et al., 2014), have been recently explored to bypass the computational knowledge uncertainties in soil and rock properties have been well-
complexity, while maintaining the capability in realistically formulating recognized in the literature, their propagation during geotechnical site
the prior distribution and likelihood function in site characterization characterization has not been explored systematically or mapped ex-
without using conjugate pairs. plicitly to different stages in site investigation.
This paper aims to consolidate recent advancement in Bayesian
studies in site characterization and develops a robust Bayesian 3. Uncertainty propagation during geotechnical site
framework for direct quantification of the actual variability of vari- characterization
ous soil and rock properties when estimated using different field or
laboratory tests. The Bayesian framework contains three key compo- Fig. 1 also depicts explicitly propagation of the uncertainties, includ-
nents: prior distribution, likelihood function, and how to solve the ing natural variability and various knowledge uncertainties, in soil and
Bayesian equation for obtaining and expressing the posterior infor- rock properties during different stages of geotechnical characterization
mation in a user-friendly way. Formulation of prior distribution of a project site. The prior knowledge is obtained during desk-study and
and quantification of prior knowledge are presented in a companion site reconnaissance. It is not perfect information and is associated with
paper (Cao et al., 2016). This paper focuses on the remaining two key some uncertainties, such as inherent variability and measurement er-
components. rors incorporated in existing data, subjective uncertainties of engineers'

Pois o titanita
Y. Wang et al. / Engineering Geology 203 (2016) 117–125 119

Fig. 1. Geotechnical site characterization and uncertainty propagation.

expertise, and so forth (e.g., Baecher, 1983; Vick, 2002). A systematic The knowledge uncertainties (i.e., measurement errors, statistical
treatment of the uncertainties in prior knowledge and their quantifi- uncertainty, and transformation uncertainty) arise from insufficient
cation are presented in a companion paper (Cao et al., 2016). After knowledge or lack of data during geotechnical site characterization. Al-
the prior knowledge is accumulated, project-specific site observa- though such knowledge uncertainties inevitably affect the prediction or
tion data are obtained from in-situ investigation and laboratory results of reliability studies, they have no influence on the existing con-
testing. The site observation data fluctuate because of inherent vari- dition of the project site or the observed performance (i.e., actual re-
ability and measurement errors (see Fig. 1). When the design prop- sponse) of geotechnical or geological systems. For example, the actual
erty XD is not measured directly, a transformation model M T is settlement of a foundation does not change no matter how accurately
needed to relate the observation data (i.e., measured property XM) the soil and rock properties under the foundation are measured, esti-
to XD. The uncertainty (i.e., transformation uncertainty) associated mated, and subsequently used in the analysis. Accurate estimates of
with the transformation model is then combined with the uncer- soil and rock properties only contribute to the accuracy of predicted re-
tainties included in site observation data (i.e., inherent variability sponse of geotechnical or geological systems, e.g., the predicted settle-
and measurement errors), propagating together into the interpreta- ment of foundations. On the other hand, the inherent variability (or
tion outcomes of the site observation data, as shown in Fig. 1. Finally, natural variability) of soil and rock exists long before the project and af-
engineers commonly utilize both the interpretation outcomes of fects significantly the actual response of geotechnical or geological sys-
site-specific observation data and prior knowledge to infer soil and tems (e.g., Fenton and Griffiths, 2002; Fenton et al., 2005; Wang et al.,
rock properties and underground stratigraphy. Because the inferred 2011; Li et al., 2013, 2014; Jiang et al., 2015; Li et al., 2015; Jamshidi
soil and rock properties and underground stratigraphy are based on Chenari and Alaie, 2015). The knowledge uncertainties and inherent
information at some locations or depths (i.e., some samples of the variability, therefore, should be differentiated in reliability studies.
population), rather than information at every location and depth However, they are usually lumped together and observed as the total
(i.e., the whole population), statistical uncertainty inevitably oc- variability of soil and rock properties (see Figs. 1 and 2). Since the
curred and is added to the inferred soil and rock properties and un- total variability contains both inherent variability and knowledge un-
derground stratigraphy. certainties and is larger than the inherent variability itself, using the

Fig. 2. Variabilities and uncertainties in soil and rock properties (modified after Kulhawy, 1996; Phoon and Kulhawy, 1999a).
120 Y. Wang et al. / Engineering Geology 203 (2016) 117–125

total variability in reliability analysis and risk assessment might lead to estimated using prior knowledge and site observation data. The six
an overestimation of failure probability of geotechnical and geological stages of geotechnical site characterization shown in Fig. 1 are mapped
systems. This result, at least partially, explains the problem of “Why to the three key elements (i.e., prior distribution, likelihood function,
are failures less frequent than our reliability studies predict?”, i.e., the and posterior distribution) of the Bayesian framework, as shown in
first problem in the Christian and Baecher list. To improve the prediction Fig. 3.
from reliability studies, it is important to characterize, directly and ex- It is worthwhile to note that, the Bayesian updating may be per-
plicitly, the inherent variability of design property and to account, formed in a sequential manner when additional information on the
simultaneously, for various knowledge uncertainties in a rational soil and rock properties is available. For example, after the Bayesian
manner. updating is finished with existing site observation data, if additional
With the clear differentiation between the inherent variability and in-situ or laboratory tests are subsequently performed to obtain addi-
knowledge uncertainties and improved understanding of the uncertain- tional site observation data, a new round of Bayesian updating can be
ty propagation during geotechnical characterization of a project site, performed to incorporate these additional data. The posterior knowl-
the next section presents a Bayesian inverse analysis framework to edge obtained in the previous round of Bayesian updating can be
characterize, directly and explicitly, the inherent variability of a design taken as the prior knowledge and used together with the additional
property XD and to, simultaneously, account for various knowledge un- site observation data to further update the knowledge on soil and rock
certainties in a rational and transparent manner. properties and underground stratigraphy, as shown by the dashed line
in Fig. 3. The Bayesian framework offers a systematic way of gradually
improving the knowledge of the underground conditions for a project
4. Bayesian framework for geotechnical characterization of a project site as the site observation data are accumulated.
site As shown in Fig. 3, the inherent variability of a design property XD
(e.g., Eu) is characterized, directly and explicitly, using a probabilistic
Both prior knowledge and project-specific site observation data are model MP (e.g., random variables or random fields) within the Bayesian
utilized during geotechnical site characterization to estimate soil and framework. Let ΘP denote model parameters of the probabilistic model Op
rock properties and underground stratigraphy. From a Bayesian per- MP that uniquely define MP. The information on ΘP is therefore required
spective, such a process can be considered as a process of updating the
prior knowledge using site observation data for a project site, as
for probabilistic characterization of the inherent variability of XD. Such
information is unknown and needs to be determined according to up
shown in Fig. 3. The prior knowledge obtained during desk-study and prior knowledge and site observation data X M (e.g., SPT data) during
site reconnaissance (i.e., Stages 1 and 2 shown in Fig. 1, respectively) geotechnical site characterization.
is represented quantitatively by a prior distribution (see Fig. 3). The Under the Bayesian framework, the updated knowledge (i.e., posterior
site observation data obtained from in-situ/laboratory testing knowledge) on ΘP is reflected by the posterior distribution PðΘP jX M Þ
(i.e., Stages 3 and 4 shown in Fig. 1, respectively) is reflected through based on both prior knowledge and site observation data X M , which is
a likelihood function (see Fig. 3). written as (e.g., Wang et al., 2010a; Cao and Wang, 2013):
After the prior distribution and likelihood function are specified,

I
using the Bayes' Theorem (e.g., Sivia and Skilling, 2006) gives a posterior ! " ! "
P ΘP jX M ¼ KP X M jΘP P ðΘP Þ ð1Þ
distribution that quantitatively reflects the updated knowledge
(i.e., posterior knowledge) and combines the prior knowledge and site
observation data. This process corresponds to the last two stages in which K = 1/P(X M ) = normalizing constant; P(ΘP) = prior distribution
(i.e., Stages 5 and 6 shown in Fig. 1) of geotechnical site characterization, of model parameters that reflects the prior knowledge on ΘP in the ab-
where soil and rock properties and underground stratigraphy are sence of data; PðX M jΘP Þ = likelihood function that reflects the model fit

mando verossimilhança
contas
As a
labafeto
Ctu

Fig. 3. Bayesian framework for geotechnical characterization of a project site.


Y. Wang et al. / Engineering Geology 203 (2016) 117–125 121

with site observation data X M . Using the posterior knowledge on ΘP Alternatively, a random field X D can be used to characterize the in-
(i.e., the posterior distribution PðΘP jX M Þ), the inherent variability of herent spatial variability of XD in the soil layer (e.g., Wang et al.,
XD is characterized quantitatively, as illustrated in Sections 7 and 8. 2010a). In the context of random field theory (Vanmarcke, 1983), XD
Note that characterization of the inherent variability of XD is an in- at Nx different locations in the soil layer are represented by a set of spa-
verse analysis problem, in which various uncertainties are lumped tially correlated random variables X D ¼ ½X D1 ; X D2 ; …; X DNX %, and the spa-
and observed together and are subsequently used to infer the inher- tial correlation between the variations of XD at different locations is
ent variability of interest. From an inverse problem point of view, the defined by a spatial correlation function (Vanmarcke, 1983; Cao and
Bayesian framework uses both prior knowledge and site observation Wang, 2014b). For a given spatial correlation function, the random
data that contain various uncertainties (e.g., inherent variability and field X D is uniquely represented by the random field model parameters
knowledge uncertainties) to inversely infer the inherent variability (e.g., the mean μ, standard deviation σ, and scale of fluctuation λ) of XD,
of XD. The information on the inherent variability of XD provided by e.g., ΘP = [μ, σ, λ].
prior knowledge and site observation data is reflected by the prior For a soil profile containing NL soil layers, the inherent spatial vari-
distribution P(ΘP) and likelihood function PðX M jΘP Þ (see Fig. 3), re- ability of XD within the NL soil layers can be characterized by NL mutually
spectively. Therefore, the critical components in the proposed Bayes- independent random fields X D n , n = 1,2,…, NL (Cao and Wang, 2013).
ian framework are formulations of the prior distribution and Let θ Pn , n = 1, 2, …, NL, denote the model parameters of the n-th soil
likelihood function which is discussed in the next section. layer. Similar to the case with one soil layer described above, the θ Pn in-
The prior distribution P(ΘP) is the probability density function cludes random field model parameters θ Rn (e.g., the mean μn, standard
(PDF) of ΘP according to prior knowledge (e.g., engineering judg- deviation σn, and scale of fluctuation λn) that are used to characterize
ment and experience). It quantifies probabilistically the prior knowl- the inherent spatial variability of XD within the n-th soil layer. In
edge about ΘP in the absence of data. When there is no prevailing addition to θ Rn , the θ Pn also contains geometry parameters θ Gn
prior knowledge of ΘP, a relatively uninformative prior distribution (e.g., the thickness hn of the n-th soil layer for a one-dimensional ran-
(e.g., a joint uniform distribution) can be applied (e.g., Wang and dom field along the depth) that are used to delineate the upper
Cao, 2013; Wang et al., 2013), together with typical ranges of and lower boundaries of the n-th soil layer (i.e., θ Pn = [θ Rn , θ Gn ], n =
model parameters reported in literature (e.g., Lacasse and Nadim, 1, 2, …, NL) (Cao and Wang, 2013). The model parameters ΘP of the
1996; Phoon and Kulhawy, 1999a; Baecher and Christian, 2003). As probabilistic model MP (i.e., NL mutually independent random fields)
the prior knowledge of ΘP improves, a more sophisticated and infor- are then taken as ΘP = [θ P1 ; θ P2 ; …; θ PNL ], which include not only the
mative prior distribution (e.g., a joint Gaussian PDF or arbitrary random field model parameters θ Rn , n = 1, 2, …, NL, but also the geom-
histogram type PDF) can be estimated from prior knowledge etry parameters θ Gn , n = 1, 2, …, NL, of the NL soil layers.
(e.g., Wang and Cao, 2013). Determination of the prior distribution The level of complexity for two probabilistic models M P above
according to prior knowledge is discussed systematically in the com- (i.e., random variables and random fields) is obviously quite differ-
panion paper (Cao et al., 2016). ent. Whether a relatively simple model (e.g., just a random variable)
or a relatively sophisticated model (e.g., a multiple layer random
5. Likelihood function field model) shall be used depends on objectives of the characteriza-
tion (e.g., whether or not the characterization aims to provide scale
Formulation of the likelihood function is a key step in the Bayesian of fluctuation for the subsequent reliability analysis) and the infor-
framework. The likelihood function PðX M jΘP Þ is the PDF of site observa- mation available for estimating the model parameters, including
tion data X M for a given set of model parameters ΘP. It quantifies prob- both project-specific observation data and prior information. For ex-
ample, it is extremely difficult, if not infeasible, to estimate scale of
abilistically the ΘP information provided by X M . As shown in Fig. 3,
fluctuation using just several discrete site observation data points
formulation of the likelihood function (i.e., PðX M jΘP Þ) requires a likeli-
(e.g., SPT-N values). A sophisticated model with insufficient input
hood model ML that probabilistically describes the relationship between data does not necessarily perform better than a simple model with
the model parameters ΘP of a design property XD and project-specific necessary and sound input data.
test data. Generally speaking, the likelihood model shall reflect sound When XD is not measured directly, it can be estimated from project-
physical insights into the relationship between XD and observation specific test results of the measured property XM through a transforma-
data and propagation of various uncertainties through the relationship. tion model MT between them (e.g., Kulhawy and Mayne, 1990; Phoon
Sections 2 & 3 have explored systematically occurrence and propagation and Kulhawy, 1999b; Mayne et al., 2002). Generally speaking, the trans-
of various uncertainties during site characterization and mapped the formation model MT may be written as:
uncertainties explicitly to different stages in site investigation. The re-
sults in these previous sections pave the way to streamlining the
formulation of the likelihood models and functions for various soil X D ¼ f T ðX M ; εT Þ ð2Þ
and rock properties. As shown in Fig. 3, the likelihood model M L
generally depends on the probabilistic model MP that is used to
characterize the inherent variability of XD and the transformation in which fT(⋅; εT) represents a functional relationship between XD and XM
model MT between XD and XM, which are discussed in what follows, (i.e., XD is a function of XM), and εT is a random variable representing the
respectively. transformation uncertainty or modeling error associated with the trans-
Probabilistic models MP (e.g., random variables and random fields) formation model (Phoon and Kulhawy, 1999b; Wang et al., 2010a;
are adopted to characterize, directly and explicitly, the inherent vari- Wang and Cao, 2013).
ability of XD (e.g., Eu) in the Bayesian framework. Consider, for example, Using the probabilistic model MP of XD and the transformation
a statistically homogenous soil layer. The inherent variability of XD in the model MT (i.e., Eq. (2)), a likelihood model ML is derived to probabi-
soil layer (i.e., the actual XD value at any location within the soil layer) listically relate ΘP to X M . Then, the likelihood function is formulated
can be characterized by a random variable (e.g., Wang and Cao, 2013). according to the likelihood model. The exact equations of the likeli-
The random variable can be uniquely defined by its distribution param- hood function are problem-specific and different for different soil
eters/statistics for a given distribution type, such as the mean μ and and rock properties, because the exact function forms for Eq. (2)
standard deviation σ of a Gaussian or lognormal random variable. In are different for different soil and rock properties (i.e., different
other words, the model parameters ΘP incorporate the distribution pa- XD ) when estimated using different field or laboratory test data
rameters/statistics of the random variable, i.e., ΘP = [μ, σ]. (i.e., X M ).
122 Y. Wang et al. / Engineering Geology 203 (2016) 117–125

6. Implementation procedures characterization of undrained Young's modulus Eu (i.e., XD = E u)


of clay using a limited number of SPT N data (i.e., X M = a set of
Based on the proposed framework, developing a Bayesian inverse SPT N values). In the probabilistic model MP, a lognormal random
analysis model for probabilistic characterization of the inherent vari- variable is used to model the inherent variability of E u in a clay
ability of a design property XD using project-specific site observation layer. The random variable is uniquely defined by the mean μ E
data X M involves six steps, which are summarized below: and standard deviation σE of E u , i.e., the model parameters ΘP =
[μE, σE]. Then, Eu is related to the SPT N value through a transforma-
1) Choose a probabilistic model MP (e.g., random variables or random
tion model MT developed by Kulhawy and Mayne (1990). Using the
fields) to model, directly and explicitly, the inherent variability of XD;
probabilistic model (i.e., the lognormal random variable) of Eu and
2) Choose a transformation model MT (e.g., empirical regressions or an-
the transformation model between Eu and the SPT N value, a likeli-
alytical relationships) that relates XD to the measured property XM;
hood model ML is derived that defines the probability of observing
3) Develop a likelihood model ML using MP and MT, which probabilisti-
the SPT N value as a function of the model parameters ΘP = [μE, σE]
cally relates site observation data X M to model parameters ΘP;
(Wang and Cao, 2013). Since the derivation showed that the loga-
4) Formulate the likelihood function PðX M jΘP Þ using ML and X M ;
rithm (i.e., ln(N)) of the SPT N value follows a Gaussian variable in
5) Determine the prior distribution P(ΘP) of ΘP according to prior
the likelihood model, the likelihood function is formulated as a
knowledge (see the companion paper (Cao et al., 2016));
conditional joint Gaussian PDF of ln(N) for a given set of μ E and
6) Obtain the posterior distribution of ΘP using the likelihood function
σE. In addition, the prior distribution is simply taken as a joint uni-
and prior distribution by Eq. (1).
form distribution of μ E and σE with the minimum values of μE,min
There are different ways of solving Eq. (1) and expressing the poste- and σ E,min and the maximum values of μ E,max and σ E,max . Details
rior knowledge of ΘP and inherent variability of XD. The following two of the likelihood function and prior distribution are referred to
sections illustrate how to depict directly the inherent variability of XD Wang and Cao (2013).
using Bayesian equivalent sample method and how to identify the ΘP After the likelihood function and prior distribution are formulated,
using Bayesian system identification and model class selection methods, the posterior distribution of μE and σE is obtained using Eq. (1). It quan-
respectively. tifies the posterior knowledge of μE and σE based on the prior knowl-
edge and project-specific SPT N values. The total probability theorem
7. Bayesian equivalent sample method is then used to integrate the posterior knowledge on μE and σE to the
posterior knowledge of the Eu inherent variability (i.e., the PDF of Eu).
As discussed in Section 5, the inherent variability of XD in a soil layer Subsequently, a large number of equivalent samples of Eu are generated
can be modeled by a random variable (i.e., probabilistic model MP) and from its PDF using MCMCS. Finally, histogram plots and conventional
its model parameters ΘP. For a given set of prior knowledge and site ob-
servation data X M , the posterior knowledge of ΘP is quantified by the
posterior distribution PðΘP jX M Þ of ΘP (See Eq. (1)). Using the posterior
knowledge of ΘP, the inherent variability of XD can be further expressed
as the PDF of XD using the Total Probability Theorem, i.e., PðX D jX M Þ ¼
∫PðX D jΘP ; X M ÞPðΘP jX M ÞdΘP , (Wang and Cao, 2013). Since the prior
knowledge and likelihood function in geotechnical practice can be so-
phisticated, the PDF of XD might be complicated and is difficult to ex-
press analytically and explicitly. To overcome such difficulty, MCMCS
is used to depict the XD PDF numerically (Wang and Cao, 2013).
MCMCS is a numerical process that simulates a sequence of samples
of a random variable (e.g., XD) as a Markov Chain with the PDF of the
random variable as the Markov Chain's limiting stationary distribution
(e.g., Robert and Casella, 2004; Wang and Cao, 2013). The states of the
Markov Chain after it reaches stationary condition are then used as sam-
ples of the random variable with the target PDF. It provides a feasible
way to generate samples from an arbitrary PDF, particularly when the
PDF is complicated and is difficult to express analytically and explicitly.
A large number of XD samples, as many as needed, can be generated
from PðX D jX M Þ using MCMCS. The posterior knowledge (i.e., the com-
bined knowledge from both prior knowledge and site observation
data) of XD is transformed into a large number of XD samples. From a sta-
tistical point of view, these MCMCS samples of XD are equivalent to
those XD data that are obtained physically from laboratory and/or in-
situ tests. These equivalent samples can be used subsequently to con-
struct histograms and cumulative frequency diagrams for proper esti-
mation of the PDF and cumulative distribution function (CDF) of XD
and to estimate its statistics (e.g., the mean and standard deviation)
using conventional statistical methods. Detailed treatments of the PðX D
jX M Þ formulation and MCMCS for generating equivalent samples of XD
from PðX D jX M Þ are given by Wang and Cao (2013).
Bayesian equivalent sample method has been used to charac-
terize probability distributions of undrained Young's modulus
(Wang and Cao, 2013) and undrained shear strength (Cao and
Wang, 2014b) of clay, and uniaxial compressive strength of rock Fig. 4. Standard penetration test (SPT) N values at the clay site of the NGES at Texas A&M
(Wang and Aladejare, 2015). Consider, for example, probabilistic University (modified after Briaud (2000) and Wang and Cao (2013)).
Y. Wang et al. / Engineering Geology 203 (2016) 117–125 123

statistical equations are used to estimate the probability distributions Bayesian system identification (BSI) method gives the posterior stan-
(e.g., PDF and CDF) and statistics (e.g., μE and σE) of Eu. dard deviations of random field model parameters (Wang et al.,
Real data from a clay site of the US National Geotechnical Experi- 2010a; Cao and Wang, 2013).
mentation Sites (NGES) at Texas A&M University (Briaud, 2000) were The number NL of soil layers is treated as a predetermined value in
used to demonstrate and validate the method (Wang and Cao, 2013). the BSI. Since NL might also be an unknown variable in site investigation,
Using five SPT N values obtained from the site, as shown in Fig. 4, and it is determined using a Bayesian model class selection (BMCS) method.
a joint uniform prior distribution with μE,min = 5.0 MPa, μE,max = A model class in BMCS is referred to a family of stratification models that
15.0 MPa, σE,min = 0.5 MPa, and σE,max = 13.5 MPa, the Bayesian equiv- share the same number of soil layers but have different model parame-
alent sample method provides a CDF plot for the inherent variability of ters ΘP (Cao and Wang, 2013; Wang et al., 2013, 2014). The number of
Eu (see Fig. 5) with a mean and standard deviation of 11.6 MPa and soil layers is considered as a variable k, which is a positive integer vary-
6.0 MPa, respectively. Fig. 5 shows comparison between equivalent ing from 1 to a maximum possible number NLmax. Therefore, there are
sample results and direct measurement of Eu values from 42 NLmax candidate model classes Mk, k = 1, 2, … , NLmax, and the k-th
pressuremeter tests performed at the same site. For Eu b 10 MPa and model class Mk has k soil layers. For a given set of site-specific observa-
CDF b 0.5, the two CDFs nearly match. As Eu or CDF further increases, tion data X M , the plausibility of each model class is quantified by condi-
the equivalent sample results are slightly larger than the pressuremeter tional probability PðM k jX M Þ, k = 1, 2, … , NLmax. The most probable
test results, and the difference in CDFs is within 10%. The Bayesian model class Mk⁎ is determined by comparing the conditional probabilities
equivalent samples method effectively tackles the difficulty in estimat- PðM k jX M Þ of all NLmax candidate model classes and selecting the one with
ing statistics of soil and rock properties from a generally limited number the maximum value of PðM k jX M Þ. The number of soil layers correspond-
of site-specific observation data (e.g., 5 SPT N values in this example) in ing to Mk⁎ is taken as the most probable number k* of soil layers.
engineering practice. Calculating PðM k jX M Þ involves the likelihood function and prior dis-
tribution for the model parameters ΘP of the model class Mk with k
soil layers. The likelihood function and prior distribution developed
8. Bayesian system identification and model class selection method
in the BSI method can be readily used in the BMCS method to calculate
PðM k jX M Þ by setting NL = k for all possible k values (i.e., 1, 2, … , NLmax)
For a soil profile containing NL soil layers, NL mutually independent
random fields can be used to characterize the inherent spatial variability (Cao and Wang, 2013; Wang et al., 2013, 2014). After the PðMk jX M Þ
of XD in each of the NL soil layers (see Section 5). The NL random fields values for the NLmax model classes are obtained, the most probable
serve as the probabilistic model MP in the Bayesian framework and model class Mk⁎ and the most probable number k* of soil layers are deter-
can be considered as a system that is uniquely defined by its system pa- mined accordingly.
rameters (i.e., model parameters ΘP = [θ P1 ; θ P2 ; …; θ PNL ]). Then, charac- The BSI and BMCS methods have been formulated for characteriza-
terizing the inherent spatial variability of XD is converted into tion of the inherent spatial variability of sand effective friction angle
identifying the system parameters (or model parameters) ΘP according using CPT data (Wang et al., 2010a; Cao and Wang, 2013), CPT-based
to prior knowledge and site observation data X M . probabilistic identification of underground soil strata (Wang et al.,
2013), and identification of soil strata in London Clay based on water
For a given set of prior knowledge and site observation data, the pos-
content data (Wang et al., 2014). Because of page limit, illustrative ex-
terior knowledge on ΘP is reflected by the posterior distribution. By
ample is not provided in this paper but referred to previous studies
maximizing the posterior distribution of model parameters, the
mentioned above (Wang et al., 2013, 2014; Cao and Wang, 2013), in
most probable values (MPVs) of model parameters are obtained under
which up to 5 soil layers are identified using CPT or water content data.
the spirit of Laplace asymptotic approximation (Bleistein and
Handelsman, 1986; Wang et al., 2010a). As discussed in Section 5,
the model parameters ΘP include not only the random field model pa- 9. Summary and conclusion remarks
rameters θ Rn , n = 1, 2, …, NL, but also the geometry parameters θ Gn ,
n = 1, 2, …, NL, of the NL soil layers. The MPVs of θ Rn , n = 1, 2, …, NL This paper consolidated recent advancement in Bayesian studies in
quantify the inherent spatial variability of XD in the NL soil layers, and site characterization and developed a robust framework for direct quan-
the MPVs of θ Gn, n = 1, 2, …, NL delineate the most probable boundaries tification of the actual variability of various soil and rock properties
of the NL soil layers. In addition to MPVs of model parameters, the when estimated using different field or laboratory tests. To facilitate for-
mulation of prior distribution and likelihood function, the procedure of
geotechnical site characterization was revisited from a Bayesian per-
spective, and the occurrence and propagation of inherent variability,
statistical uncertainty, measurement errors, and transformation uncer-
tainty during characterization of a project site were mapped explicitly
to different stages in site investigation. Based on the mapping, a robust
Bayesian inverse analysis framework was developed that streamlines
the formulation of likelihood functions for various soil and rock proper-
ties when estimated using different field or laboratory tests. Major con-
cluding remarks of the paper are given below:

1) The occurrence and propagation of inherent variability and various


knowledge uncertainties (e.g., measurement errors, statistical
uncertainty, and transformation uncertainty) are mapped to dif-
ferent stages in site investigation. Since the actual variability
(i.e., inherent variability) of soil and rock properties is a component
of the total variability observed during geotechnical characterization
of a project site, the actual variability is smaller than the total vari-
ability that is generally used in reliability analyses.
Fig. 5. Cumulative distribution function (CDF) of the undrained Young's modulus (modi- 2) Among all uncertainties in soil and rock properties, the inherent var-
fied after Wang and Cao (2013)). iability (i.e., actual variability) is the only one that affects the
124 Y. Wang et al. / Engineering Geology 203 (2016) 117–125

observed performance (i.e., actual response) of geotechnical and Bleistein, N., Handelsman, R., 1986. Asymptotic Expansions of Integrals. Dover, New York.
Briaud, J.L., 2000. The national geotechnical experimentation sites at Texas A&M Univer-
geological systems. Although the knowledge uncertainties do not af- sity: clay and sand. A summary. National Geotechnical Experimentation Sites, Geo-
fect observed performance, they do influence the total variability, technical Special Publication No. 93, pp. 26–51.
and hence, the prediction in reliability studies. This result, at least Cao, Z.J., Wang, Y., 2013. Bayesian approach for probabilistic site characterization using
cone penetration tests. J. Geotech. Geoenviron. 139 (2), 267–276.
partially, explains the problem of “Why are failures less frequent Cao, Z.J., Wang, Y., 2014a. Bayesian model comparison and selection of spatial correlation
than our reliability studies predict?” To improve the prediction in re- functions for soil parameters. Struct. Saf. 49, 10–17.
liability studies, the inherent variability of soil and rock properties Cao, Z.J., Wang, Y., 2014b. Bayesian model comparison and characterization of undrained
shear strength. J. Geotech. Geoenviron. Eng. 140 (6) (04014018, 1–9).
shall be differentiated from knowledge uncertainties and character- Cao, Z.J., Wang, Y., Au, S.K., 2011. CPT-Based probabilistic characterization of effective fric-
ized directly and explicitly. An inverse analysis model was devel- tion angle of sand. Geotechnical Risk Assessment and Management, Geotechnical Special
oped in this paper that uses the site observation data as input to Publication No. 224, pp. 403–410.
Cao, Z.J., Wang, Y., Li, D.Q., 2016. Quantification of prior knowledge in geotechnical site
infer the inherent variability. Future studies are needed to explore
characterization. Eng. Geol. 203, 107–116.
and compare systematically the difference in reliability analyses of Ching, J., Phoon, K.K., Chen, Y.C., 2010. Reducing shear strength uncertainties in clays by
geotechnical and geological systems when using total variability or multivariate correlations. Can. Geotech. J. 47 (1), 16–33.
inherent variability as input. Chiu, C.F., Yan, W.M., Yuen, K.-V., 2012. Reliability analysis of soil–water characteristics
curve and its application to slope stability analysis. Eng. Geol. 135–136, 83–91.
3) A robust framework was developed that streamlines the formulation Christian, J.T., 2004. Geotechnical engineering reliability: how well do we know what we
of likelihood functions for various soil and rock properties when es- are doing? J. Geotech. Geoenviron. 130 (10), 985–1003.
timated using different field or laboratory tests. The likelihood func- Christian, J.T., Baecher, G.B., 2011. Unresolved problems in geotechnical risk and reliabil-
ity. Geotechnical Risk Assessment and Management, Geotechnical Special Publication No.
tion is formulated in according to a probabilistic model MP that 224, pp. 50–63.
models the inherent variability and a transformation model MT Christian, J.T., Ladd, C.C., Baecher, G.B., 1994. Reliability applied to slope stability analysis.
that relates the observation data to the soil and rock property of in- J. Geotech. Eng. 120 (12), 2180–2207.
Clayton, C.R.I., Matthews, M.C., Simons, N.E., 1995. Site Investigation. Blackwell Science,
terest. A streamlined process for the formulation of MP and MT was Cambridge, Mass., USA.
developed, leading to streamlined formulation of likelihood function Fenton, G.A., Griffiths, D.V., 2002. Probabilistic foundation settlement on spatially random
for various soil and rock properties. soil. J. Geotech. Geoenviron. 128 (5), 381–390.
Fenton, G.A., Griffiths, D.V., Williams, M.B., 2005. Reliability of traditional retaining wall
4) MCMCS and Laplace asymptotic approximation can be used to by- design. Geotechnique 55 (1), 55–62.
pass the computational complexity when developing Bayesian Grigoriu, M.D., Kulhawy, F.H., Spry, M.J., Filippas, O.B., 1987. Probabilistic site strategy for
methods for geotechnical site characterization. The capability in re- transmission lines. In: Briaud, J.-L. (Ed.), Foundations for Transmission Line Towers
(GSP 8). ASCE, New York, NY, pp. 1–14 (Apr).
alistically formulating the prior distribution and likelihood function
Jamshidi Chenari, R., Alaie, R., 2015. Effects of anisotropy in correlation structure on the
in site characterization problems is maintained since conjugate stability of an undrained clay slope. Georisk 9 (2), 109–123.
prior and likelihood function is not used or artificial limitations Jiang, S.H., Li, D.Q., Cao, Z.J., Zhou, C.B., Phoon, K.K., 2015. Efficient system reliability anal-
to the choice of prior distribution and likelihood function are not ysis of slope stability in spatially variable soils using Monte Carlo simulation.
J. Geotech. Geoenviron. 141 (2), (04014096).
needed. Juang, C.H., Luo, Z., Atamturktur, S., Huang, H.W., 2013. Bayesian updating of soil param-
5) When site observation data is sparse and limited, MCMCS-based eters for braced excavations using field observations. J. Geotech. Geoenviron. 139 (3),
Bayesian equivalent sample method can be used to numerically de- 395–406.
Kulhawy, F.H., 1996. From Casagrande's ‘Calculated Risk’ to reliability-based design in
pict the inherent variability through equivalent samples. The Bayes- foundation engineering. Civ. Eng. Pract. 11 (2), 43–56.
ian equivalent sample method effectively tackles the difficulty in Kulhawy, F.H., Mayne, P.W., 1990. Manual on Estimating Soil Properties for Foundation
estimating statistics of soil and rock properties from a limited num- Design. Report EL 6800. Electric Power Research Inst., Palo Alto (306 pp.).
Lacasse, S., Nadim, F., 1996. Uncertainties in characterizing soil properties. Uncertainty in
ber of site-specific observation data in engineering practice. When the Geologic Environment: From Theory to Practice, Geotechnical Special Publication
relatively extensive site observation data is available (e.g., CPT No. 58 (I), pp. 49–75.
data), Laplace asymptotic approximation based Bayesian system Li, D.Q., Jiang, S.H., Cao, Z.J., Zhou, W., Zhou, C.B., Zhang, L.M., 2015. A multiple response-
surface method for slope reliability analysis considering spatial variability of soil
identification and model class selection methods can be used to properties. Eng. Geol. 187, 60–72.
identify underground strata and to estimate statistics of soil and Li, L., Wang, Y., Cao, Z., 2013. Risk de-aggregation and system reliability analysis of slope
rock properties and their spatial correlation (e.g., scale of fluctuation stability using representative slip surfaces. Comput. Geotech. 53, 95–105.
Li, L., Wang, Y., Cao, Z.J., 2014. Probabilistic slope stability analysis by risk aggregation.
and other random field parameters).
Eng. Geol. 176, 57–65.
Mayne, P.W., Christopher, B.R., DeJong, J., 2002. Subsurface Investigations — Geotechnical
Site Characterization, No. FHWA NHI-01-031. Federal Highway Administration, U. S.
Acknowledgment Department of Transportation, Washington D. C.
Medina-Cetina, Z., Esmailzadeh, S., 2014. Joint states of information from different prob-
abilistic geo-profile reconstruction methods. Georisk 8 (3), 171–191.
The authors would like to thank late Professor Fred Kulhawy at Cor-
Najjar, S.S., Gilbert, R.B., 2009. Importance of lower-bound capacities in the design of deep
nell University for his inspiration and enlightening discussions on the foundations. J. Geotech. Geoenviron. 135 (7), 890–900.
research topic reported in this manuscript. The work described in this Peng, M., Li, X.Y., Li, D.Q., Jiang, S.H., Zhang, L.M., 2014. Slope safety evaluation by integrat-
paper was supported by grants from the Research Grants Council of ing multi-source monitoring information. Struct. Saf. 49, 65–74.
Phoon, K.K., Kulhawy, F.H., 1999a. Characterization of geotechnical variability. Can.
the Hong Kong Special Administrative Region, China (Project No. Geotech. J. 36 (4), 612–624.
9042172 (CityU 11200115) and Project No. 8730035 (CityU8/CRF/ Phoon, K.K., Kulhawy, F.H., 1999b. Evaluation of geotechnical property variability. Can.
13G)), National Natural Science Foundation of China (Project No. Geotech. J. 36 (4), 625–639.
Robert, C., Casella, G., 2004. Monte Carlo Statistical Methods. Springer.
51409196, 51579190), and National Science Fund for Distinguished Sivia, D.S., Skilling, J., 2006. Data Analysis: A Bayesian Tutorial. Oxford University Press,
Young Scholars (Project No. 51225903). The financial supports are New York.
gratefully acknowledged. Tang, W.H., 1971. A Bayesian evaluation of information for foundation engineering de-
sign. Proceedings of the 1st International Conference on Applied Statistics and Prob-
ability to Soil and Structure Engineering, pp. 174–185 (Hong Kong).
References Trautmann, C.H., Kulhawy, F.H., 1983. Data sources for engineering geologic studies. Bull.
Assoc. Eng. Geol. 20 (4), 439–454.
Ang, A.H.-S., Tang, W.H., 2007. Probability Concepts in Engineering: Emphasis on Applica- Vanmarcke, E.H., 1983. Random Fields: Analysis and Synthesis. MIT Press, Cambridge.
tions to Civil and Environmental Engineering. John Wiley and Sons, New York. Vick, S.G., 2002. Degrees of Belief: Subjective Probability and Engineering Judgment. ASCE
Baecher, G.B., 1983. Professional judgment and prior probabilities in engineering risk as- Press, Reston, Virginia.
sessment. Proceedings of Fourth International Conference on Application of Statistics Wang, Y., Aladejare, A.E., 2015. Selection of site-specific regression model for charac-
and Probability in Soil and Structural Engineering. Universita di Firenze (Italy), terization of uniaxial compressive strength of rock. Int. J. Rock Mech. Min. Sci. 75,
Pitagora Editrice, pp. 635–650. 73–81.
Baecher, G.B., Christian, J.T., 2003. Reliability and Statistics in Geotechnical Engineering. Wang, Y., Cao, Z.J., 2013. Probabilistic characterization of Young's modulus of soil using
John Wiley & Sons, Hoboken, New Jersey (605 pp.). equivalent samples. Eng. Geol. 159, 106–118.
Y. Wang et al. / Engineering Geology 203 (2016) 117–125 125

Wang, Y., Au, S.K., Cao, Z.J., 2010a. Bayesian approach for probabilistic characterization of Whitman, R.V., 1984. Evaluating calculated risk in geotechnical engineering. J. Geotech.
sand friction angles. Eng. Geol. 114 (3–4), 354–363. Eng. 110 (2), 145–188.
Wang, Y., Cao, Z.J., Au, S.K., 2010b. Efficient Monte Carlo simulation of parameter sensitiv- Zhang, L.M., Tang, W.H., Zhang, L.L., Zheng, J.G., 2004. Reducing uncertainty of prediction
ity in probabilistic slope stability analysis. Comput. Geotech. 37 (7–8), 1015–1022. from empirical correlations. J. Geotech. Geoenviron. 130 (5), 526–534.
Wang, Y., Cao, Z.J., Au, S.K., 2011. Practical analysis of slope stability by advanced Monte Zhang, J., Zhang, L.M., Tang, W.H., 2009. Bayesian framework for characterizing geotech-
Carlo simulation in spreadsheet. Can. Geotech. J. 48 (1), 162–172. nical model uncertainty. J. Geotech. Geoenviron. 135 (7), 932–940.
Wang, Y., Huang, K., Cao, Z.J., 2013. Probabilistic identification of underground soil strat- Zhang, J., Tang, W.H., Zhang, L.M., Huang, H.W., 2012. Characterising geotechnical model
ification using cone penetration tests. Can. Geotech. J. 50 (7), 766–776. uncertainty by hybrid Markov Chain Monte Carlo simulation. Comput. Geotech. 43,
Wang, Y., Huang, K., Cao, Z.J., 2014. Bayesian identification of soil strata in London clay. 26–36.
Geotechnique 64 (3), 239–246.

You might also like