Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Engineering Geology 114 (2010) 354363

Contents lists available at ScienceDirect

Engineering Geology
j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / e n g g e o

Bayesian approach for probabilistic characterization of sand friction angles


Yu Wang , Siu-Kui Au 1, Zijun Cao 2
Department of Building and Construction, City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong

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

Article history: Site characterization is a unique problem of geotechnical engineering that utilizes both prior information
Received 7 January 2010 (including engineering judgment) and project-specic information from test borings, in-situ testing, and/or
Received in revised form 18 May 2010 laboratory testing. The problem is further complicated by inherent spatial variability of geo-materials and the
Accepted 21 May 2010
fact that only a small portion of geo-materials are examined during site characterization. This paper
Available online 1 June 2010
describes a Bayesian approach to integrate prior information and project-specic test results for probabilistic
characterization of soil properties from a limited number of tests. The Bayesian framework is developed in
Keywords:
Bayesian approach
conjunction with cone penetration tests to estimate the sand effective friction angle and with random eld
Effective friction angle theory to model the inherent spatial variability. Posterior distributions of uncertain parameters are derived.
Cone penetration tests An approximate method is used to by-pass multi-dimensional integration involved in obtaining the marginal
Probabilistic site characterization distributions, while removing the need of using traditional conjugate prior distributions. Using conditional
Spatial variability variance formula, it is shown analytically that the posterior variance of the friction angle arises from three
Random eld sources, namely, the spatial variability (aleatory) and its uncertainty (epistemic) as well as the uncertainty in
the mean value. This provides a means to determine whether the amount of project-specic information
(e.g., in-situ and/or laboratory tests) is sufcient in site characterization. Analytical solutions are also derived
for two asymptotic cases of posterior mean, which can be used as reference cases for checking the results
from the Bayesian approach. The Bayesian approach is illustrated through a set of real results of cone
penetration tests at a National Geotechnical Experimental Site in the USA.
2010 Elsevier B.V. All rights reserved.

1. Introduction framework (e.g., Vanmarcke, 1977; Phoon and Kulhawy, 1999; Baecher
and Christian, 2003; Honjo, 2008; Srivastava and Sivakumar Babu, 2009),
One distinctive feature of geotechnical engineering is that geo- geotechnical engineers are still faced with the challenge of integrating
materials (i.e., soils or rocks) are natural materials which are formed probabilistically the prior information (including engineering judgment)
and existed long before geotechnical projects are planned. A unique and project-specic information from tests and inferring the inherent
problem of geotechnical engineering is therefore to characterize the spatial variability from a limited number of tests. This question is
geometry and properties of underground geo-materials for designs. In particularly crucial for projects with medium or relatively small sizes, in
general, geotechnical site characterization is a multi-step process which the number of tests are usually too small to generate meaningful
which relies on engineering judgment and utilizes both available prior statistics for probabilistic characterization of the site.
information and project-specic information from test borings, in-situ A Bayesian approach is developed in this paper that provides a
testing, and/or laboratory testing (Mayne et al., 2002). Prior probabilistic framework to integrate prior information and results of
information include, but are not limited to, maps and surveys, local project-specic cone penetration tests (CPT) and to provide a probabi-
experience, visual observations, and published reports and studies listic characterization of soil properties and their spatial variability using a
(Grigoriu et al., 1987; Spry et al., 1988). limited number of tests. Bayesian approach has been applied to
The problem of site characterization is further complicated by the geotechnical engineering, such as characterization of model uncertainty
inherent spatial variability of geo-materials and the fact that only a small (e.g., Gilbert and Tang, 1995; Zhang et al., 2004, Zhang et al., 2009a, b;
portion of geo-materials are examined during site characterization. Najjar and Gilbert, 2009), calibration of estimated failure probability in
Although recent developments in reliability-based geotechnical designs liquefaction evaluation (e.g., Juang et al., 1999) and slope reliability
permit rational treatment of spatial variability using probabilistic assessment (e.g., Cheung and Tang, 2005), and updating of geotechnical
parameters (e.g., Miranda et al., 2009). However, research is rare that
incorporates the spatial variability in the Bayesian framework. In
Corresponding author. Tel.: + 852 2788 7605; fax: + 852 2788 7612. addition, computational complexity has been recognized as one key
E-mail addresses: yuwang@cityu.edu.hk (Y. Wang), siukuiau@cityu.edu.hk
(S.-K. Au), zijuncao3@student.cityu.edu.hk (Z. Cao).
limitation of the Bayesian approach (e.g., Zhang et al., 2009a).
1
Tel.: + 852 2194 2769; fax: + 852 2788 7612. Traditionally, conjugate prior distributions have been used to by-pass
2
Tel.: + 852 3442 6492; fax: + 852 2788 7612. the computational problems but this nevertheless introduces articial

0013-7952/$ see front matter 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.enggeo.2010.05.013
Y. Wang et al. / Engineering Geology 114 (2010) 354363 355

limitations to the choice of prior distributions. An asymptotic approxi-


mation is presented in this paper to by-pass the computational
complexity, while maintaining the user's exibility to use non-conjugate
prior distributions.
The paper starts with random eld modeling of sand effective
friction angle and regression between cone tip resistance and the sand
effective friction angle, followed by development of the proposed
Bayesian framework and the asymptotic approximation. Then,
procedure is described for implementing the proposed approach,
and illustration is provided by applying the approach to characterize
the effective friction angle at the sand site of the US National
Geotechnical Experimental Sites (NGES) at Texas A&M University.
Effect of amount of project-specic test data on posterior uncertainty
is explored and illustrated with examples. Analytical solutions are also
derived for two asymptotic cases of posterior mean, which can be
used as reference cases for checking the results from the Bayesian
approach.

2. Random eld modeling of friction angles

Random eld theory (Vanmarcke, 1983) is applied in this work


to model the inherent spatial variability of effective friction angle, ,
within a single homogeneous stratum of sand. Consider a one- Fig. 1. Regression between effective friction angle and normalized cone tip resistance
(after Kulhawy and Mayne, 1990).
dimensional homogeneous Gaussian random eld, (D), in which D
is the depth and is a random variable with mean, , and standard
deviation, . The Bayesian approach is applicable for general choice of developed by Kulhawy and Mayne (1990), which is a semi-log
spatial variability model which nevertheless needs to be xed in this regression equation
work for development of the method. Here, the correlation structure is
= ln q = a + b + m 5
assumed to be an exponential covariance function (Fenton, 1999a, b)
  where q = (qc / pa) / (v0 / pa)0.5 is the normalized cone tip resistance;
2 2jd1;2 j
COVD1 ; D2  = exp 1 v0 and pa are vertical effective stress and standard atmospheric

pressure (i.e., 100 kPa), respectively; a = 0.209, b = 3.684, m =
where is a correlation length and d1,2 = |D1 D2| is the distance 0.586 and is a standard Gaussian random variable. The last term m
between depths D1 and D2. Note that statistically homogeneous sand represents a modeling error associated with the regression equation.
layers can be identied using prior information and modied Bartlett It should be noted that the original equation in Kulhawy and Mayne
statistics (Phoon et al., 2003). (1990) expresses in terms of ln q; it has been inverted here to
Let = [(D1),(D2),..., (Dn)]T be a vector of effective friction facilitate development in this work.
angles at depths D1,D2,..., Dn, respectively. In the context of random Let = [1,..., n]T be a vector of ln q data at depths D1,D2,..., Dn.
eld, it has the following representation in terms of a sequence of Using Eqs. (2) and (5) gives
independent and identically distributed random variables: T
= a + b
 1 + a Z + m
L  6
T
= 
1 + 
L Z 2

When spatial variability is assumed to be independent of model
where 1 = [1,..., 1]T, Z = [Z1,..., Zn]T is a standard Gaussian vector with uncertainty, i.e., Z and are independent, it can be readily reasoned
independent components, and L is a n-by-n upper-triangular matrix that is a Gaussian vector with mean (a + b)1 and covariance matrix
obtained by Cholesky decomposition of the correlation matrix R
2 2 2
satisfying C =a 
 R + mI 7

T
L 
L =
R 3 where I is a n-by-n identity matrix.

Knowledge of the model parameters , , and are required to
and the (i, j) entry of R is given by completely dene the probability distribution of . In the next section,
we shall present a Bayesian approach for updating the knowledge of
 
2jdi;j j model parameters based on prior information and eld observation
Ri;
 j = exp 4
of .

Note that the second term in Eq. (2) represents spatial variability. 4. Bayesian framework

3. Cone tip resistance and sand friction angle regression Within a Bayesian framework, the updated knowledge about the
model parameters is reected through their joint posterior distribu-
The effective friction angle of soil can be estimated from in-situ tion given the observation data and prior information:
tests, such as the cone penetration test (CPT). CPT consists of pushing
a cylindrical steel probe into the ground and measuring the resistance p; ; j = k p j; ; p; ; 8
 
to penetration. The effective friction angle of the tested soil is
obtained by means of regression between and the resistance, qc, where k = 1 / p( ) is a normalizing constant that does not depend on ,
measured at the tip of the cone. Fig. 1 shows an empirical model , and . The term p( |,,)
is called the likelihood function that
356 Y. Wang et al. / Engineering Geology 114 (2010) 354363

reects the model t with the data. The term p(,,) is the prior Setting f / = 0 and solving for gives
distribution that reects available information about the model
parameters in the absence of data. As mentioned before, given , ,
0
+ a2 C 1
1T 
w2 
and , is a Gaussian vector with mean (a + b)1 and covariance * = 13
1
w2
+ a2 C 1
1T  1
matrix C. The likelihood function is thus given by

n = 2 1 = 2 where
p j; ; = 2 j det 
Cj 9
  
1 T 1
= b
1 = a 14
exp a + b C a + b
1  1  
2 
is the best prediction from the regression model [i.e., Eq. (5)] for given
The prior distribution is assumed to be a Gaussian distribution
. Note that the MPV * depends on and through the matrix C 1, and
with mean (0,0,0) and standard deviation (w,w,w)
1T C 11 is simply the sum of all entries of C 1. The MPV of and can
1 therefore be obtained by minimizing numerically f(*(,),,) on their
p; ; = two-dimensional space.
23 =(
2
w w w )
1 2 1 2 1 2
exp 0 0 0 5.2. Posterior standard deviation for , ,
2w2 2w2 2w2
10 Under the Laplace asymptotic approximation of integrals (Bleistein
and Handelsman, 1986), the covariance matrix G of the Gaussian PDF
The posterior PDF in Eq. (8) is a joint distribution among , , and that approximates p(,,| ) is given by the inverse of the Hessian
. To obtain the posterior marginal PDF for , integration on Eq. (8) matrix of f = ln p(,,| ) evaluated at the MPV, i.e.,
over and is needed (and similar for the other two model
2 31 15
parameters). Since the integrand is complicated, analytical integration 2 f 2 f 2 f
is often infeasible. Numerical integration may be performed using a 6 7
6 2 7
two-dimensional grid on the space of and for the posterior PDF of 6 7
6 7
G=6 f 7
2 2
, but it must be performed repeatedly for a number of values of so f
 6
6
7
as to yield information about the whole marginal distribution. Such 6 2 77
6 7
computational complexity has been recognized as one key limitation 4 2 f 5
of the Bayesian approach in literature (e.g., Zhang et al., 2009a). sym:
2
Fortunately, due to the nature of the integrand arising from Bayesian
problems, the integral can be evaluated approximately using The main diagonal terms of G give the posterior variance of ,
asymptotic techniques, which is essentially a Gaussian approximation and while the off-diagonal terms give their covariance, i.e.,
of the posterior PDF. This is presented in the next section. s2 = G(1,1), s2 = G(2,2) and s2 = G(3,3). The calculation of these
posterior variances requires the Hessian matrix of f (e.g., by nite
5. Approximation of posterior PDF difference), followed by matrix inversion.
For the problem under discussion, some insights can be gained
5.1. Posterior mean for , , about s upon detailed analysis. It can be argued that 2f / and
2f / are usually small compared to 2f / 2, as is only weakly
The complexity of repeated numerical integration can be avoided correlated with and under the posterior PDF. Then G(1,1) can be
by an approximate method that involves approximating the posterior approximated to the rst order by simply the reciprocal of 2f / 2.
PDF as a Gaussian PDF, which is in the same spirit of Laplace Consequently, the standard deviation of under the posterior PDF can
asymptotic approximation of integrals (Bleistein and Handelsman, be approximated as
1986). By this approximation, the posterior PDF for the model !1 = 2 " #1 = 2
j
parameters is a joint Gaussian PDF with mean vector equal to the 2
f * 1 2 T 1
s = +a 
1 C * 1 16
most probable value (MPV) of the posterior PDF. The MPV, denoted by 2 w2 
(*,*,*), maximizes the p(,,| ).
Under this approximation, the
determination of the posterior mean for , , reduces to nding the where the last expression is obtained by direct differentiation of
MPV by numerically minimizing the function ln p(,,| ) on the Eq. (12); an asterisk (*) denotes that the quantity is evaluated at the
three dimensional space of (,,). Further, it turns out that only two- MPV
dimensional minimization is required to nd the MPV because the
 
MPV of can be found analytically in terms of and . To see this, note 2 2 jdi; j j
C * = a *2 R* + m
2
I; R*i; j = exp 17
that     *

f ; ; = ln p; ; j 11 Eq. (16) provides an understanding on the uncertainty of after



1 1 T 1 incorporating prior information and observation data. The rst term
= ln j det C j + a + b C a + b
1  1
2 2   in the bracket suggests a positive correlation between s and w,
1 2 1 2
which agrees with intuition. It can be expected that w has immaterial
+ 0 + 0 effect on s when 1 / w2 is small compared to the second term in the
2w2 2w2
bracket whose magnitude grows with the amount of data. Note that
1 2
+ 0 + constant w also has an indirect effect on the second term in the bracket,
2w2 because it affects the MPV, e.g., see Eq. (13).
The effect of incorporating data on reducing s is not immediately
apparent, however. In particular, C* depends on the MPVs * and *
f T 1 1
= a C a + b
1  1 + 2 0 12 which maximizes the posterior PDF, but as the amount of data
 w
increases there is no guarantee that they will always change in a way
Y. Wang et al. / Engineering Geology 114 (2010) 354363 357

that reduces s. This possibility can be reasoned from the fact that the
newly incorporated data may not agree with the previous data, which
is reected in the Bayesian approach by an increase in s. This shall be
illustrated in application examples later.

6. Implementation procedure

Two key elements in the formulation described above are


minimization of f(*(,),,) = ln p(*(,),,| ) on their two-
dimensional space to obtain *, *, * and calculation of the inverse of
the Hessian matrix of f = ln p(,,| ) evaluated at *, *, * to
obtain s2 , s2, s2. The minimization can be readily implemented by
conventional optimization algorithm, e.g., the function fminsearch
in MATLAB (Mathworks Inc., 2010). The differentiation of f = ln
p(,,| ) for constructing the Hessian matrix can be evaluated
numerically in MATLAB using nite difference. The Bayesian updating
procedure comprises 5 steps and involves 8 equations, and it is
implemented by a 32-line computer code calling a 27-line objective
function in MATLAB, as shown in Appendix A. Details of each step and
its associated equations are summarized in what follows:
1. Obtain cone tip resistance qc versus depth data from CPT tests and
convert them to a vector = [1,..., n]T at depths D1, D2,..., Dn using
Eq. (5);
2. Obtain appropriate prior information including the mean 0, 0, 0
and their standard deviation w, w, w;
3. Construct the objective function f(*(,),,) = ln p(*(,),
,| ) using Eqs. (11), (13), (14), (7), and (4);
4. Minimize the objective function f(*(,),,) using the MATLAB
function fminsearch and determine its corresponding MPV (i.e.,
*,*,*) as posterior mean;
5. Calculate the inverse of the Hessian matrix of the objective function
evaluated at *, *, * to obtain posterior variance s2 , s2, s2 using
Eqs. (15) and (17). Each term in the Hessian matrix [i.e., Eq. (15)] is
calculated using nite difference techniques in MATLAB.
Fig. 2. A set of CPT results at Texas A&M University NGES Sand Site (after Mayne, 2007).
It is worthwhile to point out that the algorithm can also be
implemented in Microsoft EXCEL. The minimization can be carried out
using the Solver function in Microsoft EXCEL, and the nite scenario reects a realistic situation in the current setting where a fair
difference techniques can be easily realized in EXCEL. Examples of amount of previous site-specic data is available. Fig. 3 shows the
application of the Solver function to geotechnical engineering can be posterior PDF of when the whole set of CPT results are used in the
referred to Low and Tang (1997) and Wang and Kulhawy (2008). Bayesian approach. The solid lines in Fig. 3 represent results obtained
from Gaussian approximation, while the dots show the results from
7. Application example numerical integration. These two sets of results compare favorably,
indicating that the Gaussian approximation is reasonably accurate.
The proposed Bayesian approach and MATLAB codes developed
are applied to probabilistic characterization of effective friction angle
at the sand site of the NGES at Texas A&M University (Briaud, 1997).
The site is comprised of a top layer of sands to about 12 m and
underlain by hard clay thereafter. The groundwater table is about 6 m
deep, and total unit weight of sand is about 18.4 kN/m3. Fig. 2 shows
the tip resistance data (qc) versus depth obtained from a CPT test
performed at this site (Tumay, 1997; Mayne, 2007). As a simple
preliminary calculation, the prole of the best regression prediction
for can be calculated using Eq. (14), which yields a mean of 38.6
and standard deviation of 3.6, respectively. We shall next use this set
of data to update our knowledge of the effective friction angle at this
site. For illustration, the updating is performed considering two
different pieces of prior information separately.

7.1. Prior information derived from Akkaya and Vanmarcke (2003)

The rst set of prior information include the mean, 0 = 37, 0 = 2,


0 = 3.3 m, and standard deviation, w = 0.5, w = 0.7, w = 2.5 m.
These are consistent with the previous research work by Akkaya and
Vanmarcke (2003), who summarized twenty-two CPT soundings
performed at this sand site and established the prior information. This Fig. 3. Posterior probabilistic distribution function (PDF) of .
358 Y. Wang et al. / Engineering Geology 114 (2010) 354363

Further checking with all other cases conrmed that this degree of of cone penetration depth, respectively. For a given cone penetration
agreement is typical. depth, Bayesian updating is performed with the CPT soundings up to
Fig. 4 shows the posterior MPVs of effective friction angle, with the given penetration depth. Fig. 4 therefore depicts the evolution of
Fig. 4 (a), (b) and (c) plotting variations of *, *, and * as a function the posterior estimates as more and more project-specic data are
incorporated. When cone penetration depth is zero, no data, except
prior information, is available to characterize and therefore the
curves always start with their prior means. As the cone penetration
depth increases, more CPT test results are incorporated and they
gradually inuence the posterior MPVs. The posterior * gradually
increases from prior 0 = 37 to about 37.3 and the posterior *
gradually increases from prior 0 = 2 to about 2.8. Note that the
prior information adopted in the case are derived from CPT tests
performed at the same site, and they are quite consistent with the CPT
results used in Bayesian updating. It is therefore not surprising to see
that the difference between the prior and posterior and are
relatively small. However, as we shall see in the next subsection, when
the prior information differ signicantly from the project-specic
test results, the difference between the prior and posterior can be
signicant.

7.2. Inconsistent and uninformative prior information

We next investigate the case when the prior PDF is uninformative


and does not agree with the data. Specically, suppose the mean
values are 0 = 30, 0 = 5, 0 = 6 m, and the standard deviations are
w = 6, w = 2, w = 3 m. This set of prior information is signi-
cantly different from the prior information reported by Akkaya and
Vanmarcke (2003).
Fig. 5 shows the posterior MPVs of , with Fig. 5(a), (b) and (c)
plotting variations of *, *, and * as a function of cone penetration
depth, respectively. Similar to Fig. 4, the posterior estimates start with
their prior values. However, they deviate signicantly from their prior
as the penetration depth increases and more data is incorporated. At
cone penetration depth = 10 m, the posterior * is about 39, which is
almost identical to the values (i.e., 38.6) that are directly estimated
from the CPT results (i.e., without using any prior information). The
information provided by the CPT results dominate the estimate of
posterior *, and effect of the prior distribution is rather minor.

8. Effect of data quantity on posterior uncertainty

In the presence of project-specic data , the uncertainty about the


effective friction angle is reected by the posterior variance at a
typical location in the soil stratum. Based on conditional variance
formula shown in Appendix B, the posterior variance var[| ] can be
expressed as

2 2 2
var j  = * + s + s 18


Eq. (18) indicates that given the data, the uncertainty about the
effective friction angle arises from three sources. The rst term is the
most probable estimate of the spatial variability, which can only be
improved, but not eliminated, by collecting more data. The second and
third term arise from the uncertainty associated with the estimates of
spatial variability and the mean effective friction angle, which can be
reduced as the amount of independent data increases. Eq. (18)
provides a means to determine whether the amount of project-specic
information (e.g., in-situ and/or laboratory tests) is sufcient in site
characterization. If s2 + s2 is relatively small compared to *2, the
posterior variance of is mainly attributed to the best estimate of
inherent spatial variability, which cannot be eliminated by collecting
more data. Therefore, the benet from performing more in-situ and/or
laboratory tests is expected to be marginal. On the other hand, if s2 + s2
Fig. 4. Posterior most probable values (MPV) of sand effective friction angle with is large with respect to *2, additional in-situ or laboratory tests are
consistent and informative prior. effective to further reduce the s2 + s2 .
Y. Wang et al. / Engineering Geology 114 (2010) 354363 359

variance [see Eq. (18)] in both absolute and relative scales. As


uncertainties in the prior estimates of spatial variability and the mean
effective friction angle are relatively small (i.e., w = 0.5 and w = 0.7)
and the CPT results are consistent with the prior information, the
posterior variance regarding the estimates of spatial variability and
2
mean effective friction angle are rather small [e.g., s2 = 0.24( ) and
2 ( 2)
s = 0.18 at cone penetration depth= 10 m]. Fig. 6(b) shows that
most posterior variance (more than 95%) is attributed to the most
2
probable estimate of inherent spatial variability [e.g., *2 = 7.99( ) at
cone penetration depth = 10 m]. Fig. 6 also shows that, when the cone
penetration depth is larger than 5 m (i.e., more than 5 m CPT test data
are used in the Bayesian updating), the additional reduction in the total
posterior variance is minimal. This suggests that, when the prior
information is consistent with the data and informative with small
uncertainty in the prior PDF, the posterior information converges
quickly and remains more or less the same as the prior information.

8.2. Illustrative example with inconsistent and uninformative prior

For the CPT data shown in Fig. 2 and the prior information used in
subsection 7.2, Fig. 7 shows their respective components of posterior
variance [see Eq. (18)] in both absolute and relative scales. As shown
in Fig. 7(a), the total variance and *2 tend to decrease as the cone

Fig. 5. Posterior most probable values (MPV) with inconsistent and uninformative
prior.

8.1. Illustrative example with consistent and informative prior

For the CPT data shown in Fig. 2 and the prior information used in
subsection 7.1, Fig. 6 shows their respective components of posterior Fig. 6. Components of posterior variance with consistent and informative prior.
360 Y. Wang et al. / Engineering Geology 114 (2010) 354363

9.1. Case I: approaches to 0

When approaches to 0, correlation matrix R reduces to an identity


matrix I , and the covariance matrix C dened by Eq. (7) reduces to
(a22 + 2m) I . Thus, C 1(a22 + m
2 1
) I , and the posterior *
dened by Eq. (13) can be approximated by

1
0 + n

w2 * 2 + m
2
= a2 d
* = 1 n
19
w2
+
* 2 + m2
= a2

where

1 n
d = 20
n i=1 i

is the mean effective friction angles directly estimated from CPT data
using regression formula. Eq. (19) is a weighted average of the prior
mean 0 and the data mean d with weights proportional to 1 / w2 and
n / (*2 + m 2
/ a2), respectively. The latter weight dominates for large
n, and so when is small, the posterior * approximates to d for large
n. That is, the posterior * approaches to an average of best predictions
directly from CPT data using regression equation.

9.2. Case II: approaches to innity

At the other extreme, when is extremely large, R reduces to an


n-by-n matrix with all entries equal to 1 (i.e., [1] matrix), and C reduces
to a2 2[1] + m
2
I . In this case, using the results in the Appendix C, it can
be shown that

1
0 + 1

w2 * 2 + m2
= a2 n d
* = 1 1
21
w2
+ *2 + m 2
= a2 n

which, again, is a weighted average of the prior mean 0 and d , but


now with weights proportional to 1 / w2 and [*2 + m 2
/ a2n] 1. The
2
latter weight tends to 1 / * for large n, and so for large and n, the
posterior * converges to a weighted average of the prior mean and
the sample mean with weights proportional to 1 / w2 and 1 / *2. Note
Fig. 7. Components of posterior variance with inconsistent and uninformative prior.
that the weights reect the relative precision suggested by prior
information and by data. In this case the inuence of the prior
penetration depth increases. However, s2 tends to increase when cone information (through 0) does not diminish as the amount of data
penetration depth increases to about 2 m. It then decreases when the increases because the additional data does not yield further
penetration depth further increases to 10 m. The initial increase in s2 independent information when the correlation length is large.
is attributed to the disagreement between the prior and the data
about . As further CPT data are incorporated into the prior 9.3. Illustrative example
information, the effect of CPT data dominates that of the prior,
causing a systematic decrease in s2 . To illustrate two asymptotic cases of posterior mean, CPT test data
Fig. 7(b) shows the relative contribution of *2, s2 , and s2 to the are simulated using the random eld and regression models described
total variance. The contribution of *2 tends to increase as cone in this paper with = 32, = 1.92, and various values. Note that in
penetration depth increases. However, even at cone penetration practice the actual values of soil properties are unknown, and they are
depth = 10 m, the contribution of *2 is about 70%, with the other 30% estimated through prior information and project-specic test results.
from s2 and s2, which can be reduced by incorporating more CPT test For each given value, a separate set of CPT test data is simulated for a
data. In this case, it is desirable to carry out more CPT tests to further penetration depth up to 10 m at a depth interval of 0.1 m. For
reduce the s2 and s2. example, Fig. 8 shows the normalized tip resistance, q, for = 1 m
from the simulation. Subsequently, the CPT test data are incorporated
in the Bayesian approach to estimate the *, together with a prior
9. Asymptotic cases of posterior mean mean and standard deviation of 0 = 30, = 6, 0 = 1.5 m and
w = 6, w = 1.2, w = 0.3 m, respectively.
It is instructive to consider two asymptotic cases of posterior mean Fig. 9 shows the posterior * for varying from 0 to 10 m by a solid
when the correlation length is either very small (i.e., none of the n line. The posterior * varies slightly at about 32, which is the actual
CPT data points are correlated) or very large (i.e., all CPT data points value of soil property. The * dened by Eqs. (19) and (21) for the two
are strongly correlated). In geotechnical site characterization where asymptotic cases of = 0 or = extremely large are shown by solid
the actual value of soil property is of interest but unknown, these two lines with rectangles or triangles, respectively, in Fig. 9. Their values
asymptotic cases can be used as reference cases for checking the are very consistent with the true value of 32 and compare favorably
results from the Bayesian approach. with the * values obtained from the Bayesian approach for different
Y. Wang et al. / Engineering Geology 114 (2010) 354363 361

10. Summary and conclusions

This paper presented a Bayesian approach to integrate prior


information (including engineering judgment) and project-specic
test results and to provide probabilistic characterizations of soil
properties and their spatial variability from a limited number of tests.
The Bayesian framework has been developed in conjunction with
project-specic cone penetration tests (CPT) to estimate the sand
effective friction angle, , and with random eld theory to model the
inherent spatial variability of .
The posterior PDF for , , and has been formulated, and an
approximate method has been developed for its determination that
by-passes repeated numerical integration. The approximate method
makes uses of Laplace asymptotic approximation of integrals and
approximates the posterior PDF as a Gaussian PDF, which has been
shown to be reasonably accurate. The most probable value and
covariance matrix of the posterior PDF are directly calculated from the
peak location and Hessian matrix of the posterior PDF.
It has been shown analytically that uncertainty about stems
from three sources: the inherent spatial variability (i.e., *2), and
uncertainty (i.e., s2 + s2 ) regarding the estimates of spatial variability
and the mean effective friction angle. The spatial variability *2 can
only be improved in accuracy, but cannot be eliminated, by collecting
more data. On the other hand, s2 + s2 can be reduced as the amount of
independent data increases. This provides a means to determine
whether the amount of project-specic information (e.g., in-situ and/
or laboratory tests) is sufcient in site characterization. If s2 + s2 is
relatively small with respect to *2, further in-situ and/or laboratory
tests only provide marginal reduction in overall uncertainty. If s2 + s2
is large compared to *2, overall uncertainty can be effectively
reduced by performing additional in-situ or laboratory tests.
In addition, analytical solutions have been derived for two
asymptotic cases of posterior mean, which yields insights on the
role of different information sources in the Bayesian approach. These
Fig. 8. A set of simulated CPT test results ( = 1 m).
analytical solutions can be used to provide a theoretical estimate of
the soil property and to check the results from the Bayesian approach.
The Bayesian approach has been illustrated through a set of real
cone penetration test data at a National Geotechnical Experimental
. In geotechnical site characterization where the actual value of soil Site in USA. The application example with real data showed that the
property is of interest but unknown, Eqs. (19) and (21) can be used to Bayesian approach properly accounts for the effects of prior
provide a theoretical estimate of the soil property and to check the information and data in the updating process.
results from the Bayesian approach. Note that the * values from
Eqs. (19) and (21) rely on both prior information and test data, and
Acknowledgments
they might differ signicantly in practice, as opposed to the close
values shown in this illustrative example.
The authors would like to thank Professor Fred H. Kulhawy at
Cornell University for his instructive and valuable comments during
development of this work. We also thank the anonymous reviewers
and editor for their valuable comments. This work is supported by a
grant 9041260 (CityU 121307) from the Research Grants Council of
the Hong Kong Special Administrative Region, China. The nancial
supports are gratefully acknowledged.

Appendix A. MATLAB code

function [Posterior,VarphiD,CovM] = Bayesian_Function(CoeReg,


Prior,Data)
% MATLAB code for the Bayesian approach described in this paper.
% INPUT:
% CoeReg = [a, b, m] are regression coefcients in Eq. (5)
% = ln q =a + b + m
 
0 0
% Prior = 0 are prior information
w w w
% Data is CPT test data represented by a four by n matrix,
% where n = number of data points.
Fig. 9. Two asymptotic cases of posterior mean of sand effective friction angle *. % The rst three columns are x, y, z coordinates of measurement locations,
362 Y. Wang et al. / Engineering Geology 114 (2010) 354363

% and the fourth column is the normalized cone tip resistance q in R = ex p( 2 dd / Dc); C = ( a s) ^2 R + sm ^2 eye(nz);
Eq. (5) % = ln q =a + b + m invC = inv(C);
% OUTPUT: if length(y) == 2
 
* * * m=mp(1)/a/sp(1)^2ones(1,nz)invC(bones(nz,1)xi);
% Posterior = are posterior information
s s s
m = m / (a ones(1,nz) invC ones(nz,1) + 1 / a / sp(1)^2);
% VarphiD = [*2 s2 s2 ] are three components of posterior variance else
% dened in Eq. (18) m = y(1);
% CovM is a three by three matrix for posterior covariance of *, *, * end
% Dene necessary variables
xi0 = zeros(nz,1); xi0(:) = a * m + b; dxi = xi xi0;
DX = zeros(1,3); H = zeros(3,3); D = max(Data(:,3));
f=0.5*(log(det(C))+dxi. invC dxi +sum((([m,s,Dc].mp(:))./
y0 = Prior(1,2:3);
sp(:)).^2));
% Minimize the objective function and determine its corresponding MPV
y0 = fminsearch(@(y) objective(y,CoeReg,Prior,Data),y0);
Appendix B. Analysis of posterior variance
y0 = abs(y0);
[f0,m0,invC] = objective(y0,CoeReg,Prior,Data);
In the presence of data, the uncertainty about the friction angle is
% Calculate the Hessian matrix by nite difference
reected by the posterior variance at a typical location in the soil
DX=1e3 [m0,y0]; X = [m0,y0];
stratum. According to Eq. (2), the friction angle at a given location can
fX = feval(objective,X,CoeReg,Prior,Data);
be expressed as = + Z, where, for given and , Z is standard
[X1,X2,X3,X4] = deal(X);
Gaussian. The conditional expectation of the friction angle is given by
for i = 1:3
[X20,X30] = deal(X); X20(i) = X(i) DX(i);
E j  = E j  + EZ j  = E j  B  1
   
X30(i) = X(i) +DX(i);
for j = 1:i where the last equality has made use of the fact that E[ Z| ] = 0,
X3 = X30; X2 = X20; because
X4 = X3; X4(j) = X4(j) + DX(j); X3(j) = X3(j) DX(j);
X1 = X2; X1(j) = X1(j) DX(j); X2(j) = X2(j) + DX(j); EZ j  = EEZ j; j  = EEZ j; j  = E 0j  = 0
     
f1 = feval(objective,X1,CoeReg,Prior,Data);
B  2
if i == j
[f2,f3] = deal(fX);
else To obtain the conditional variance of the effective friction angle,
the conditional variance formula is utilized (Ross, 2007),
f2 = feval(objective,X2,CoeReg,Prior,Data);
f3 = feval(objective,X3,CoeReg,Prior,Data);

end var j  = Evar j; ; j  + varE j; ; j  B  3
    
f4 = feval(objective,X4,CoeReg,Prior,Data);
H(i,j) =(f4 f3 f2+ f1) /(4 DX(i) DX(j)); H(j,i) =H(i,j); To proceed, note that var[|,, ] = 2 and E[|,, ] = , thus
end
end 2
var j  = E j  + var j  B  4
  
invH = inv(H);
% Calculate output Using the identity E[2| ] = E[| ] 2+var[| ],
the above becomes
Posterior = [m0,y0;sqrt((diag(invH)))];
VarphiD = [y0(1)^2, invH(2,2), invH(1,1)];
2
CovM = invH; var j  = E j  + var j  + var j  B  5
   
function [f,m,invC] = objective(y,CoeReg,Prior,Data)
% The objective function dened by Eq. (11)
a = CoeReg(1); b = CoeReg(2); sm = CoeReg(3); In terms of quantities computed from the posterior PDF, where
mp = Prior(1,:); sp = Prior(2,:); xi = log(Data(:,4)); * = E[| ], s2 = var[| ] and s2 = var[| ], Eq. (B-5) can be expressed
xyz = Data(:,1:3); as

if length(y) == 2 2 2 2
var j  = * + s + s B  6
s = abs(y(1)); Dc = abs(y(2)); 
else
s = abs(y(2)); Dc = abs(y(3));
end Appendix C. Background for Eq. (21)
nz = length(xi(:)); dd = zeros(nz);
for i = 1:nz 1 This appendix gives the background results that lead to Eq. (21).
for j = i + 1:nz From elementary matrix operations, it can be shown by induction
that the inverse of an n-by-n matrix A[1] + I has diagonal terms
dd(i,j) = sqrt(sum((xyz(i,:) xyz(j,:)).^2));
dd(j,i) = dd(i,j);
end 1 n1 +
A
 i; i = ; i = 1; :::; n C  1
end n +
Y. Wang et al. / Engineering Geology 114 (2010) 354363 363

and off-diagonal terms Fenton, G., 1999a. Estimation for stochastic soil models. Journal of Geotechnical and
Geoenvironmental Engineering 125 (6), 470485.
Fenton, G., 1999b. Random eld modeling of CPT data. Journal of Geotechnical and
1
A
 i; j = ; ij C  2 Geoenvironmental Engineering 125 (6), 486498.
n + Gilbert, R.B., Tang, W.H., 1995. Model uncertainty in offshore geotechnical reliability.
Proceeding of the 27th Offshore Technology Conference, Society of Petroleum
From this it can be readily obtained that Engineers, Houston, pp. 557567.
Grigoriu, M.D., Kulhawy, F.H., Spry, M.J., Filippas, O.B., 1987. Probabilistic site strategy
n for transmission lines. In: Briaud, J.L. (Ed.), Foundations for Transmission Line
T 1
1 
 A 1=
 C  3 Towers (GSP 8). ASCE, New York, pp. 114.
n + Honjo, Y., 2008. General vs. local reliability based design in geotechnical engineering.
In: Katafygiotis, L.S., Zhang, L.M., Tang, W.H., Cheung, M.M. (Eds.), Proceedings of
and for any x=[x1,..., xn]T, the 4th Asian-Pacic Symposium on Structural Reliability and its Applications
(APSSRA'08), pp. 4152. June 1920, Hong Kong.
n Juang, C.H., Rosowsky, D.V., Tang, W.H., 1999. Reliability-based method for assessing
T 1 n
 A
1  x=
 x C  4 liquefaction potential of soils. Journal of Geotechnical and Geoenvironmental
n + i = 1 i Engineering 125 (8), 684689.
Kulhawy, F.H., Mayne, P.W., 1990. Manual on estimating soil properties for foundation
design. Report EL 6800, Electric Power Research Inst., Palo Alto, p. 306.
Applying these result to C = a2 2[1] + m
2
I , the diagonal terms of
Low, B.K., Tang, W.H., 1997. Efcient reliability evaluation using spreadsheet. Journal of
C 1 are given by Engineering Mechanics 123 (7), 749752.
Mathworks Inc., 2010. MATLAB the language of technical computing, bhttp://www.
1 n1 2 + m
2
= a2 1 mathworks.com/products/matlab/N (March 9, 2009).
C
 i; i = 2 for large n; i = 1; :::; n Mayne, P.W., 2007. Cone penetration testing: a synthesis of highway practice. National
n + m = a m
2 2 2 2
m Cooperative Highway Research Program (NCHRP) Synthesis 368, Transportation
C  5 Research Board, Washington, D. C.
Mayne, P.W., Christopher, B.R., DeJong, J., 2002. Subsurface investigations geotech-
nical site characterization. Publication No. FHWA NHI-01-031, Federal Highway
and the off-diagonal terms are given by Administration, U.S. Department of Transportation, Washington D. C.
Miranda, T., Gomes Correia, A., Ribeiro e Sousa, L., 2009. Bayesian methodology for
2
1 1 updating geomechanical parameters and uncertainty quantication. International
C
 i; j = 2 for large n; ij C  6 Journal of Rock Mechanics & Mining Sciences 46, 11441153.
n + m
2 2
= a2 m
2
nm Najjar, S.S., Gilbert, R.B., 2009. Importance of lower-bound capacities in the design
of deep foundations. Journal of Geotechnical and Geoenvironmental Engineering
also, 135 (7), 890900.
Phoon, K.K., Kulhawy, F.H., 1999. Characterization of geotechnical variability. Canadian
n 1 Geotechnical Journal 36 (4), 612624.
T 1
 C
1  1= 2 2
 2 2 for large n C  7 Phoon, K.K., Quek, S.T., An, P., 2003. Identication of statistically homogeneous soil
n + m
2
a layers using modied Bartlett statistics. Journal of Geotechnical and Geoenviorn-
mental Engineering 129 (7), 649659.
and Ross, M.S., 2007. Introduction to Probability Models. Academic Press.
Spry, M.J., Kulhawy, F.H., Grigoriu, M.D., 1988. Reliability-based foundation design for
n n transmission line structures: geotechnical site characterization strategy. Reprot EL-
T 1 1 1
 C
1  = i 2 2 i for large n C  8 5507(1), Electric Power Research Inst., Palo Alto, p. 142.
 na + m i = 1
2 2 2
na i = 1 Srivastava, A., Sivakumar Babu, G.L., 2009. Effect of soil variability on the bearing
capacity of clay and in slope stability problems. Engineering Geology 108, 142152.
Tumay, M.T., 1997. In-situ testing at the national geotechnical experimentation sites
(Phase II). Final Report, FHWA Contract No. DTFH61-97-P-00161, Louisiana
References Transportation Research Center, Baton Rouge, p. 300.
Vanmarcke, E.H., 1977. Probabilistic modeling of soil proles. Journal of Geotechnical
Akkaya, A., Vanmarcke, E.H., 2003. Estimation of spatial correlations of soil parameters Engineering 103 (11), 11271246.
based on data from the Texas A&M University NGES. In: Vanmarcke, E.H., Fenton, Vanmarcke, E.H., 1983. Random Fields: Analysis and Synthesis. MIT Press, Cambridge.
G.A. (Eds.), Probabilistic Site Characterization at the National Geotechnical Wang, Y., Kulhawy, F.H., 2008. Economic design optimization of foundations. Journal of
Experimentation Sites (GSP 121). ASCE, Reston, Virginia, pp. 2940. Geotechnical and Geoenvironmental Engineering 134 (8), 10971105.
Baecher, G.B., Christian, J.T., 2003. Reliability and Statistics in Geotechnical Engineering. Zhang, L.M., Tang, W.H., Zhang, L.L., Zheng, J.G., 2004. Reducing uncertainty of
John Wiley & Sons Inc., Hoboken, New Jersey, p. 605. prediction from empirical correlations. Journal of Geotechnical and Geoenviron-
Bleistein, N., Handelsman, R., 1986. Asymptotic Expansions of Integrals. Dover, New mental Engineering 130 (5), 526534.
York. Zhang, J., Zhang, L.M., Tang, W.H., 2009a. Bayesian framework for characterizing
Briaud, J.L., 1997. The national geotechnical experimentation sites at Texas A&M geotechnical model uncertainty. Journal of Geotechnical and Geoenvironmental
University: clay and sand. Report, NGES-TAMU-007-September 1997, Texas A&M Engineering 135 (7), 932940.
University, p. 38. Zhang, L.L., Tang, W.H., Zhang, L.M., 2009b. Bayesian model calibration using
Cheung, R.W.M., Tang, W.H., 2005. Realistic assessment of slope reliability for effective geotechnical centrifuge tests. Journal of Geotechnical and Geoenvironmental
landslide hazard management. Geotechnique 55 (1), 8594. Engineering 135 (2), 291299.

You might also like