FURTHER CONTRIBUTIONS TO RELIABILITY-BASED

PILE-SETTLEMENT ANALYSIS
By S. T. Quek, 1 Y. K. Chow, 2 Member, ASCE, and K. K. Phoon3

ABSTRACT: TO study the statistics of the settlement of pile foundations, the un-
certainty in the Young's modulus of the soil must first be estimated. This paper
considers the practical aspects of estimating the uncertainty in the model for quan-
tifying the Young's modulus of the soil as well as the uncertainty due to the inherent
variability of the basic soil property, namely, the shear strength. The uncertainty
in the model is evaluated by treating the correlation parameter as a random variable;
the soil is modeled as a spatially random medium. Through a numerical example
it is shown how these two uncertainties can be simultaneously considered in eval-
uating the reliability of the pile foundation against an allowable settlement limit
using design charts. The design charts were formulated based on first-order com-
putations and validated in this paper by comparing with second-order results over
the range of values of practical interest.

INTRODUCTION

In an earlier paper by Phoon et al. (1990), a finite-element-based first-

order reliability analysis of the settlement of pile foundations was performed
and design charts were presented for practical usage. The soil was assumed
to be an isotropic linear elastic medium since field and experimental evi-
dences [e.g. Poulos and Davis (1980)] suggest that the load-settlement re-
sponse of pile foundatons is essentially linear at normal working loads. The
Young's modulus of the soil medium was modeled as a homogeneous ran-
dom field and incorporated into the finite-element analysis from which the
statistics of the pile settlement were computed using a first-order pertur-
bation technique. The random field was discretized by treating the property
across each element as constant, the actual value within each element being
represented by a random variable. The covariances of the Young's modulus
between pairs of elements were derived. Reliability analysis was then per-
formed to assess the serviceability state of the pile, in which it is considered
satisfactory if the settlement of the pile head does not exceed an allowable
limit. For the purpose of formulating concise design charts for general usage,
a generalized allowable settlement limit was defined.
Phoon et al. (1990), however, did not discuss specifically how the statistics
of the random input quantity—that is, the Young's modulus of the soil—
can be estimated in practice. Often, a correlation model is used, in which
case the uncertainty in the model itself, in addition to the uncertainty in
the basic quantity evaluated (e.g. shear strength) must be incorporated.
Furthermore, the use of first-order analysis in computing the response sta-
tistics is questionable in view of the large uncertainty generally associated
'Sr. Lect., Dept. of Civ. Engrg., Nat. Univ. of Singapore, Kent Ridge, Singapore
0511,
2
Republic of Singapore.
Sr. Lect., Dept. of Civ. Engrg., Nat. Univ. of Singapore, Kent Ridge, Singapore
0511,
3
Republic of Singapore.
Sr. Tutor, Dept. of Civ. Engrg., Nat. Univ. of Singapore, Kent Ridge, Singapore
0511, Republich of Singapore.
Note. Discussion open until October 1, 1992. To extend the closing date one
month, a written request must be filed with the ASCE Manager of Journals. The
manuscript for this paper was submitted for review and possible publication on
November 7, 1990. This paper is part of the Journal of Geotechnical Engineering,
Vol. 118, No. 5, May, 1992. ©ASCE, ISSN 0733-9410/92/0005-0726/$1.00 + $.15
per page. Paper No. 878.
726
with geotechnical problems. As an extension of the earlier paper, the ob-
jectives of the present paper are: (1) To present some practical consider-
ations in the estimation of the statistics of the Young's modulus of the soil
including the incorporation of model uncertainty; (2) to illustrate the validity
of the design charts obtained from first-order analysis by presenting the
results of second-order analysis; and (3) to illustrate through a numerical
example how the design charts presented in Phoon et al. (1990) can be used
to include the uncertainty in the correlation model for the Young's modulus
of the soil, in which the correlation parameter is treated as a random var-
iable, in addition to the basic soil property considered (i.e. shear strength),
which is modeled as a random field.

ESTIMATION OF RANDOM FIELD PARAMETERS

As presented in the earlier paper (Phoon et al. 1990), the point and spatial
statistics describing the variability of the soil Young's modulus required for
the probabilistic analysis of pile-foundation settlement are: (1) Point mean
Es; (2) point coefficient of variation (COV) AE ; (3) vertical and radial scales
of fluctuation 8Z and 8r; and (4) local spatial averages x. To evaluate these
statistical parameters in practice, it is necessary to be able to measure a
reliable value of the Young's modulus of the soil. The methods of deter-
mining the Young's modulus of the soil include laboratory and in-situ tests,
pile-load tests, and empirical correlations based on previous experience.
Conventional laboratory tests such as triaxial or oedometer tests are gen-
erally not suitable for direct measurements of the soil Young's modulus
because they do not follow the stress path that the soil adjacent to the pile
takes. Laboratory model pile tests may overcome this deficiency to some
extent, but may not accurately reflect the behavior of the prototype piles,
because of the presence of scale effects. The most satisfactory method of
obtaining the soil Young's modulus at present seems to be to carry out a
pile-load test in situ and to backcalculate the soil modulus from the measured
settlements. In most practical situations it is not possible to carry out such
testing, especially in the early stages of the design. Engineers therefore
frequently resort to empirical correlations between the soil modulus and
some other soil properties that may be determined more readily from lab-
oratory or field tests.

Correlation Model for the Soil Young's Modulus

For piles in clay, the Young's modulus of the soil, Es, has often been
correlated with the undrained shear strength, cM, obtained by analyzing load-
test data and backcalculating the values of Es from settlement measurements
(Poulos 1972). In practice, linear correlation models appear to be more
prevalent. A typical linear correlation model is given by
Es = xc„ (1)
where x = correlation parameter. Ottaviani and Marchetti (1979) obtained
a x value of 1,000 from the result of a single pile-load test. The analysis of
the load test was based on an axisymmetric finite-element computer program
that used a hyperbolic rather than linear stress-strain curve for the soil
medium. The value of Es used in the computation of x corresponds to the
initial tangent modulus of this nonlinear stress-strain curve. In another pile-
load test, conducted by Cooke et al. (1979), a backcalculated value of Es
= 75 MN/m2 was obtained for a soil medium whose average undrained
727
shear strength was about 60 kN/m2, giving a value of 1,250 for x- A much
lower value of x was obtained from the results of two pile-load tests reported
by Meyerhof et al. (1981). From an average undrained shear strength of
about 180 kN/m2 and a backfigured value of Es ranging from 65 MN/m2 to
80 MN/m2, the mean value of x is approximately 400. This value of x is in
general agreement with the results reported by Butterfield and Ghosh (1980),
in whose study a series of about 20 laboratory model pile tests were back-
analyzed and the mean and standard deviation of x were found to be ap-
proximately 470 and 53, respectively. The results of Aschenbrener and Olson
(1984) seem to indicate that an in-between value of x, lying between 400
and 1,000, may be more appropriate for the linear correlation model. To
obtain the value of x, the authors analyzed a large data set consisting of 68
pile-load tests. Among the 68 load tests, 56 of them were for driven piles;
the remaining 12 were for bored piles. The estimated mean and standard
deviation of x were 740 and 170, respectively.
It is apparent from the foregoing that the value of x is quite uncertain.
This may be due to differences in the methods of determining c„ and Es,
load level at which Es is determined, pile-installation techniques, methods
of pile testing and instrumentation, overconsolidation ratios of the clay, and
types of clay. By having different values of x to cater for these influences,
the uncertainty may be reduced. However, this is currently not done due
to the lack of detailed data available. Despite the uncertainty in x, the linear
correlation model is commonly adopted in practice because of its simplicity.
The effect of the uncertainty in x o n the statistics of Es are subsequently
illustrated.

Estimation of Point Statistics

To estimate the point statistics of interest, namely the mean, Es, and
COV, AE , of the soil Young's modulus, it is assumed that x is a random
variable and Es and c„ are random fields, that is
£SR = XCUR (2)
where R = position vector of a point within the soil medium. Eq. (2) may
be approximated using first-order Taylor-series expansion about the mean
and assuming that c„ is statistically homogeneous, giving

£,R = XCU + c„(x - X) + x(c„R - c„) (3)

where x = mean of x; and cu = constant point mean of the undrained
shear strength. The point mean of the soil modulus may be obtained by
taking expectation on both sides of (3)
E(ESR) = Es = Xc„ (4)
The mean values of x and cu may be estimated using
i M
= (5)
* MS*
1 N
C
Cu = T-, X A (6)
J\ ,'=1
where x, is obtained from the rth pile load test; M = total number of load
tests used in the establishment of the correlation model; c„R, = undrained
728
shear strength measured at location R,; and N = total number of sample
points in the soil medium.
Since cu is assumed to be statistically homogeneous, the variance of c„ at
any point should be constant. Furthermore, x a n d c„R are quite likely to
be statistically independent and the covariance between them may be as-
sumed as negligible. The point COV of Es can therefore be approximated
using (3) and (4) by
1_
= A? + A2 (7)
N' M
where Ax = COV of x; and Ac = the constant point COV of the undrained
shear strength. They may be estimated using
2
- i) 2& (x, - x)
A (8)
* X2(M
1
2 (cBR, - cuf (9)
cl(N - 1) ,fx
The third and fourth terms in (7) account for the uncertainty in the esti-
mation of the mean values due to limited field data used. Note that (7) is
obtained by assuming that the values of x, are statistically independent and
that the distances between sample points in the soil medium are sufficiently
far apart for the values of c„R, to be effectively uncorrelated. If the sample
points are close to one another, a modified value of N should be used in
(7) (Vanmarcke 1983).
Typical values of Ac are presented in Table 1. For illustrative purposes,
a value of 0.4 is assumed for Ac . This value represents the uncertainty due
to the inherent variability of the soil, which, unlike the other components
in (7), can be reduced by spatial averaging. In the case of Ax, a value of
0.23 is obtained based on the linear correlation model suggested by As-
chenbrener and Olson (1984). This correlation uncertainty is systematic in
nature because it affects the soil modulus at all the points in the soil medium
by the same amount and will not decrease with spatial averaging. The
statistical uncertainties, as represented by the third and fourth terms of (7),
are also systematic and are not affected by spatial averaging, although they
can be reduced by increasing the sample sizes, i.e. using larger vaues of M

TABLE 1. Typical Coefficients of Variation for Undrained Shear Strength

Source Soil type and site COV of c„
(1) (2) (3)
Ward et al. (1959) London clay, various site 0.11-0.21
Hooper et al. (1966) London clay, Barbican site 0.11-0.33
Lumb (1966) Hong Kong marine clay 0.18
Ladd et al. (1971) Soft Bangkok clay 0.23-0.41
Singh (1971) General 0.20-0.40
Schultz (1975) Fischbach clay 0.26-0.40
Matsuo et al. (1977) Silty clay, Hatsukaichi site 0.22-0.33
Vanmarcke (1977) New Liskeard, Ontario 0.32
Yamamoto (1982) Japan, various sites 0.18-0.41
Asaoka et al. (1982) West Side Highway, New York 0.18-0.30
Cragg et al. (1987) Ontario 0.06-0.30

729
and N. For the correlation model given by Aschenbrener and Olson (1984),
a fairly large value of M ( = 68) was used in the determination of x- In the
case of N, it is assumed that 20 sampling points are used to measure the
undrained shear strength. Using these figures, the total point variance of
Es may be evaluated as

A2a = 0.42 + 0.23* + —^ + 9 | p = 0.22 (10)

The contribution to the total COV of Es due to inherent variability, cor-

relation and statistical uncertainties, in this case are 72%, 24%, and 4%,
respectively. For simplicity, the systematic uncertainties [last three terms of
(7)] are first assumed negligible. Results obtained may then be modified to
accommodate systematic uncertainties and are illustrated through an ex-
ample.

Estimation of Spatial Statistics

The local spatial averages, x„ may be defined as
x =
> hLE*Rd[i <n>
where O, denotes the domain in which the soil modulus is averaged. Re-
placing (3) into (11), the mean of xt can be evaluated as
E[x,] = xcu (12)
while the covariance between any two averages, xt and xjy is given by

COV[x,,x,.] = c5(xAx)2 ( l + ^ ) + ^ X 2 ( c A „ ) 2 4- C(0„a,.) (13)

where the covariance function, C(i2j,0 ; ), is given by (28) (replacing Es and

AEs by xcu and Ac , respectively) of Phoon et al. (1990), and involves both
the vertical and radial scales of fluctuations.
To determine C(0„fl ; ) in practice, it is necessary to evaluate the vertical
and radial scales of fluctuation besides the mean and COV of cu. A few
methods discussed by Vanmarcke (1983) may be used to determine these
scales of fluctuation, which include the following.

• If a number of observations are made within a depth (or horizontal)

interval of size T, then

(l4)
^^-(C-AJ^-I)!^-^}
where Na = total number of observations within the interval T; c„(z,)
= undrained shear strength measured at a depth of z,\ and

yT= -rrlL cu(Zi) (15)

• If a number of intervals, Nb, of identical size, T, is specified and the

local spatial average across each interval, yT„ is determined from (15),
the scale of fluctuation can be obtained as
730
 Nb

2 S (Vr,- - )>r)2 (16)

(c-„Ac„) (iv6 - i ) /e,
where yT = mean of the local spatial averages.
The correlation between the undrained shear strengths measured at
different points is given as
1
PW £ [c„(z,) - c„][c„(z / + T ) - cj (17)
(c,AJ2(Nc - 1) It
where Nc = number of pairs of undrained shear strengths, i.e. c„(z,)
and c„(z,+T), whose separation distance |z,+T - z\ = T. By fitting an
appropriate analytical correlation model (e.g. quadratic exponential
model) over the data points, P(T), 8Z can be deduced.
• Since the spatially varying shear-strength profile, cu(z), will intersect
the constant mean trend cu at various points along the z-axis, the average
distance between these intersections can be used to estimate 8Z. For
example, 5Z is V2Arr times the average distance for the quadratic ex-
ponential model.

Typical scales of fluctuation for the undrained shear strength are presented
in Table 2.

SECOND-ORDER PERTURBATION ANALYSIS OF PILE SETTLEMENT

Formulation of Second-Order Equations

As discussed in Phoon et al. (1990), the settlement analysis of single piles
can be carried out most efficiently by means of an axisymmetric finite-
element method, and the rational treatment of uncertainties may be incor-
porated using perturbation techniques. The stiffness matrix K and the dis-
placement vector u may be expressed as functions of the random variables,
Xj (defined as the spatial average of the soil modulus over the rth finite
element), using Taylor-series expansion about the mean

K (18)
1 = 1 OX;
s
1 s d2U
4>/^/ (19)
IS dxdx,
where £ = number of finite elements; cj), = x, — Es; and Es denotes the
mean of the soil modulus at a point. Note that the matrices K and u and
TABLE 2. Typical Scales of Fluctuation for Undrained Shear Strength

Scale of fluctuation
Source Soil type and site Direction (m)
(1) (2) (3) (4)
Wu (1974) Chicago clay Vertical 0.79
Matsuo et al. (1977) Nagoya port Vertical 1.25
Nagoya port Vertical 2.50
Horikawa bridge Vertical 1.82
Vanmarcke (1977) New Liskeard Vertical 5.0
Horizontal 46.0

731
the partial derivatives are all evaluated at the mean value of x. Unlike the
displacement vector, the stiffness matrix K can only be expressed as a first-
order function of x since the soil medium is assumed to be linear elastic.
Substituting (18) and (19) into the equilibrium equation and applying sec-
ond-order perturbation technique yields the following equations:
Ku = Q (20)

K— = — u (21)
dXj dxt

^ d2u 9K 30 3K 30
K. — — (22)
dXjdXj dXj dXj dXj dxt

Note that the load vector Q is assumed to be deterministic, although the

formulation can be extended to treat the load vector as random.
The unknowns in (20)-(22) are the displacements u, the first derivative
du/dxh and the second derivatives d2u/dXtdx/, and the solutions are subse-
quently used to determine the second-order mean and covariance of u. From
(19), it can be shown that the second-order mean and covariance of u are

£[u] = u + j t i - | E[<M»,] (23)

•£ i = l y = l OXflXj

+ 1 y y y ii^puui + Iyyv fiL^L

2 h h & dxfaj dxk fil(W>*J +
4 h h & h dxtdXj dxkdx,
•(E[<M>;<M>,] - £[M>]£[+A1) (24)
where ^[c^^tby] = COY[xhx.], given by (13). Note that the covariance of u
involves higher moments of x, which are generally difficult to evaluate due
to limited field data. A mathematically convenient way to circumvent this
difficulty is to limit the class of distributions for xh defined as the local
spatial average of the soil Young's modulus over the ith finite element, to
that in which the third and fourth moments can be expressed in terms of
the first two moments, such as the normal, exponential, and uniform dis-
tributions. In this study, it is assumed that x, is normally distributed. This
assumption is fairly reasonable in view of the central limit theorem, which
states that the sum or average of many uncorrelated or lightly correlated
random variables (herein, the soil Young's modulus at different points in
the finite element) tends toward the normal distribution. From the prop-
erties of the normal distribution, (24) may be simplified as
dQdQT
CTA * n VV , r ^ i j. l V V V V d2032ur

• (E[WklE[$M + mME[4>M) (25)

Despite the foregoing simplification, the computation of (25) is still tedious
because of the quadruple summation term. Therefore, only limited results
are generated to assess the accuracy of the first-order solutions presented
732
by Phoon et al. (1990). The statistics obtained in this study are the mean
(w) and COV(Au)) of the settlement at the pile head, which can be extracted
from (23) and (25) at the appropriate degree of freedom.

Second-Order Estimates of Single-Pile-Settlement Statistics

The settlement statistics of single piles are dependent on five important
parameters. The two deterministic parameters are the slenderness ratio
Lid, which describes the length-to-diameter ratio of the pile, and the stiffness
factor K, which denotes the ratio of the pile modulus Ep to the mean soil
Young's modulus Es. The three stochastic parameters are the COV of soil
modulus A£ , which represents the variability of the modulus at a point, and
the vertical 1 ^) and radial (8r) scales of fluctuation. The results of a first-
order parametric study on the effect of the five parameters on w and Au
were presented in Phoon et al. (1990).
For the charts presented in the earlier paper [Figs. 2 and 3 of Phoon et
al. (1990) in particular] to be of practical use the limits of accuracy of the
first-order solutions must be assessed, and are ascertained by means of
second-order solutions. To reduce computation time, a fairly coarse "sto-
chastic" mesh consisting of layer elements is adopted, which implicitly as-
sumes that the soil Young's modulus is perfectly correlated in the lateral
direction, i.e. 8r is infinite. The effect of using a finite or infinite value of
5r on the second-order solutions is shown in Table 3. The solutions are
presented in terms of the dimensionless mean parameter 4 , defined as
Es(adlP, where P is the applied load on the pile. It can be observed that
the second-order estimates of 1^ remain almost the same regardless of whether
5r is finite or infinite. From a comparison of first- and second-order estimates
of 1^, it can further be observed that the error is slightly larger when 8r is
infinite. Hence, besides reducing computing time, the assumption of an
infinite radial scale of fluctuation is also appropriate for the purpose of
assessing the limits of accuracy of the first-order solutions since these so-
lutions are expected to be less accurate under such circumstances.
It is quite apparent from Table 4 that the second-order COV of pile-head
settlement is almost the same as the first-order solutions (at least for values
of A£ and hJL less than 0.6 and 5.0, respectively). The differences in the
solutions are minimal even when the magnitudes of the stochastic parameters
are fairly large. For example, the error in the first-order COV for values
of bJL = 5.0 and AE = 0.6 is only about 2%. It should not, however, be
construed that the second-order effects are not significant. The similarity
in COV may be attributed to the fact that the mean and standard deviation
are magnified by more or less the same factor upon inclusion of higher-
TABLE 3. Effect of Radial Scale of Fluctuation on Second-Order Estimate of Mean
Settlement
DIMENSIONLESS MEAN PARAMETER (/J

Ae Second-Order Estimates
(%) First order estimates 8r = 2.5 8, = =°
(1) (2) (3) (4)
20 0.0834 0.0838 0.0838
40 0.0834 0.0849 0.0850
60 0.0834 0.0868 0.0869
Note: u> = PIJEJ; Lid = 25; K = 500; vp = 0.2; vs = 0.499; and 8Z/L = 0.05.

733
TABLE 4. Comparison of First and Second-Order Solutions for COV of Pile-Head
Settlement
COV of Pile-Head Settlement (AJ
First-order Second-order
bz/L A£ estimates estimates
(1) (2) (3) (4)
0.005 0.2 0.92 0.92
0.005 0.4 1.84 1.84
0.005 0.6 2.76 2.76
0.05 0.2 2.90 2.89
0.05 0.4 5.80 5.75
0.05 0.6 8.70 8.55
0.5 0.2 8.62 8.61
0.5 0.4 17.24 17.14
0.5 0.6 25.86 25.46
5.0 0.2 14.78 14.89
5.0 0.4 29.56 30.13
5.0 0.6 44.34 45.34
Note: Lid = 25; K = 500; and hrIL = «>.

0.30 0.20
L/=10 K=100 1 , =10 K = 10000
0.21 0.16

0.16 0.12
a 1,-M

0.0S
0.12
0.04
0.06
0.00 0.00
0.30 0.05
L, =100 K = 100 I , =100 K = 10000
. 'd '1
0.24 0.04

0.18 0.03
b 1
J
0.12 1st order 1 ^ 0.02
A
ES =»-2
0.06 - AE = 0.4 0.01
A
, , E , = 0.6
0.00 0.00
0.01 0.10 1.00 0.01 0.10
6Z/L 6Z/L

FIG. 1, Effect of AEs and bJL on Second-Order Mean Settlement of Single Piles

order terms. Since the coefficient of variation is defined as the ratio of the
standard deviation to the mean, it is not unreasonable to observe that the
first- and second-order COVs are almost the same. This observation suggests
that the second-order COV can be approximated accurately by using Figs.
2 and 3 of Phoon et al. (1990).
Fig. 1 illustrates the effect of A£ and bJL on the dimensionless mean
parameter, 1^. It can clearly be seen that the second-order mean may not
agree with the first-order solutions. This is because the second-order mean
increases as the magnitudes of the stochastic parameters increase [(23)].
The first-order mean [i.e. first term of (23)], on the other hand, is inde-
pendent of these stochastic parameters. Although the deviations between
734
 18.0

11.0 -

0.01 0.10 1.00

FIG. 2. Effect of A£ and hJL on Second-Order Standard Deviation of Pile Settle-

ment

first- and second-order solutions become increasingly significant as the val-

ues of bJL and A£ increase, these are only of theoretical interest. Typical
values of A£ and &z in Tables 1 and 2 are generally less than 0.4 and 2.5
m, respectively. For an average pile length of 30 m, a realistic value for 8Z/
L would therefore be about 0.1. From Fig. 1, the error in the first-order
mean, for example, is only about 3% for values of 8/L = 0.1 and A£ =
0.4. In view of this observation, it is apparent that the total uncertainty in
the pile-head settlement, as represented by the standard deviation (CTU), may
be estimated quite accurately using the first-order mean and COV for most
practical purposes (see Fig. 2). Note that a first-order estimate of the mean
or standard deviation is adequate even at large values of A£ (up to 60%)
provided the value of 8Z/L is sufficiently small. Similarly, larger values of
82/L (up to 5.0) can also be considered if the COV of the soil modulus does
not exceed 30%. For cases in which both bJL and AEs are large, the use of
second-order solutions to estimate statistics of settlement may be necessary.
To use the first-order charts to estimate the response statistics resulting
from the total uncertainty, it should be obvious from the first term of (25)
that the variance resulting from the individual components of uncertainty
(spatial, model, or statistical) can be obtained separately from the charts
and subsequently added to obtain the total variance of the response.

RELIABILITY ANALYSIS OF SINGLE PILES

Although the mean and covariance provide useful insights on the settle-
ment variation of pile foundations, it may be desirable to obtain a realistic
measure of the assurance of serviceability in the form of a reliability index
or the probability of unserviceable behavior by incorporating the first-order
second-moment reliability analysis into the framework of the stochastic
finite-element method. In pile foundation design, serviceability is assured
if the settlement does not exceed an allowable limit. For the purpose of
formulating concise design charts, a general definition of the allowable
settlement was defined (Phoon et al. 1990)
735
co„ = (1 + a-A,„)co ( 26 )
where a denotes the number of standard deviations between the allowable
settlement and the mean settlement. The performance function describing
the serviceability limit state of a pile foundation may be given as follows:
i'(co) = co„ - co (27)

where the pile foundation is considered unserviceable if g(co) < 0, from

which the probability of unserviceable behavior (or the reliability index)
can be computed using first-order reliability methods. The design charts
presented in Phoon et al. (1990) enable the reliability index p to be deter-
mined without having to run a stochastic finite-element computer program.
It may be noted that the design charts can be used easily even if the allowable
settlement were specified as a constant. The important conclusions to be
noted from the study of Phoon et al. (1990) are that due to the definition
of (26), p is practically unaffected by the parameters Lid, K, and 8r. Hence,
the significant parameters are A£ and 82.
Unlike the response statistics, the estimation of the reliability index ac-
counting for the total uncertainty (containing terms both affected and un-
affected by spatial variability) is not straightforward. To use the chart,
overall equivalent values for 8Z and AE must be obtained since the scales
of fluctuation inherent in the two types'5of uncertainties are different. This
implies automatically that the vertical dimension of the element size, T,
must be known. In the case of the chart generated by Phoon et al. (1990),
the dimension is L/10. By considering the effect of element size in (7), the
equivalent parameters, A'Es and 8Z, can be estimated iteratively using

*IT*(T&) = Ayi(7\8 z ) + A* + | * + ^L (28)

where r 2 (r,8 z ) and rzCT,8z) = variance functions that can be approximated

by the function (Vanmarcke 1983)
r z (r,8 z ) = 1 for T < 8Z (29a)

rf(r,8 z ) = | for T > 8Z (296)

With the equivalent values, the reliability index can be estimated as de-
scribed in Phoon et al. (1990).

ILLUSTRATIVE EXAMPLE

To illustrate the procedure just discussed, the uncertainty in the settle-

ment of a concrete bored pile as well as its reliability index against an
allowable settlement is determined using the design charts. Details regarding
the installation, testing, and instrumentation of the bored pile are given by
Ottaviani and Marchetti (1979). The relevant design data required for the
uncertainty analysis are summarized as follows.

Design Data
The length and diameter of the pile are 23.5 m and 0.60 m, respectively,
giving a slenderness ratio of approximately 40.
The Young's modulus of the concrete is assumed to be 31,000 MN/m2.
The vertical profile of the undrained shear strength is shown in Fig. 3.
736
 UN0RAINE0 SHEAR
STRENGTH Cu(MN/m 2 )
o' ) * SCISSOMETER
0-1 0-2

IN SITU f
2 o fR AX U
/
i •---a 4-
4
/ •
6
+
8 l/
+
1
10
/ 4>
12 •A
•

14
+
16 1
r
18 i
w
20 \

22 ;]
@--^i
•*

•
4-

24

26 *i *•

FIG. 3. Undrained Shear-Strength Profile at Viale Giulio Cesare, Rome [after Ot-
taviani and Marchetti (1979)]

Using (6) and (9), the mean and COV are estimated to be 110 kN/m2 and
0.28, respectively, based on 23 field measurements, which were obtained
using a series of scissometer or vane shear tests.
The correlation between the undrained shear strengths at two different
locations is computed using (17). By varying the vertical distance between
the two locations TZ, a plot of the correlation function P(TZ) versus TZ can
be obtained as shown in Fig. 4. The following single exponential correlation
model appears to provide a good fit for the data points:

T.
P(TZ) = exp \ - ~ (30)
2T4
Although the theoretical results obtained in this study are based on a quad-
ratic exponential model, they can still be applied provided the vertical scale
of fluctuation is evaluated as (Vanmarcke 1983) 82 = 2 x 2.4 = 4.8 m.
The value of 8/L will work out to about 0.2, whereas for 8/L an assumed
737
 1.0

0.8

0.6

0.1.

0.2

0 j i L

1 2 3 4 5 6 7
Separation distance, i:z(m)
FIG. 4. Autocorrelation of Undrained Shear Strength in Vertical Direction

value of 2.5 is adopted for the analysis. This corresponds to 8,. of about 60
m, which is of the same order of magnitude as the field value reported by
Vanmarcke (1977).
The correlation between the soil modulus and the undrained shear strength
is assumed to follow the linear model suggested by Aschenbrener and Olson
(1984). The mean and COV of the correlation parameter x are given as 740
and 0.23, respectively, estimated from 68 pile-load tests. The corresponding
mean soil modulus Es and the stiffness factor K are 81.4 MN/m2 and 380,
respectively.
The working load is assumed to be 1.5 MN. This corresponds to a factor
of safety of 2.3 since the measured ultimate load is about 3.4 MN. Typical
factors of safety for working loads generally lie between 2 and 3.

Analysis
The first step of the analysis involves determining the mean settlement
from a deterministic finite-element computer program, which is approxi-
mately 2.5 mm, based on the mean values of the parameters and the working
load. This mean settlement may also be obtained from charts that are readily
available in literature [e.g. Poulos and Davis (1980)]. The COV of the soil
modulus is then evaluated by assuming that there are no systematic uncer-
tainties, i.e. A£ = 0 . 2 8 . Together with the foregoing given values of Lid,
K, 82/L, and 8,/L, the COV of settlement arising solely from the inherent
variability of the soil may be estimated from Fig. 5 as A'M = 0.28 x 0.28
= 0.078. Fig. 5 is reproduced from Fig. 2 of Phoon et al. (1990). Next, the
COV of the soil modulus is reevaluated by assuming that there is no inherent
variability in the soil. Using the last three terms in (7), ABj = 0.24. The
effect of this AE on the COV of settlement can similarly be estimated from
Fig. 5 but with infinite scales of fluctuation since the source of uncertainty
is fully systematic. The COV of settlement due to systematic uncertainties
Aj, is Asu = 0.62 x 0.24 = 0.15. Once AJ, and Asu are determined the total
COV of settlement may be evaluated as discussed earlier using (25), giving
738
 6 Z /L
FIG. 5. Effect of hz/L on COV of Pile Settlement (8 r /L = 2.5) (Phoon et al. 1990)

1 1 ,
DATA: I ^
_ 6Z/L = )5 A
6 r /L = o.
L/d = 50
K =
vm
JJ&--+""'"
0S9__
<g,**"

*** --**"*"

0 2 t S 8 »
a
FIG. 6. Reliability Index Chart for 8 Z /L = 0.5 (Phoon et al. 1990)

Ara = [(AL)2 + (AL)2]05 = 0.17 (31)

At a 95% confidence level, the settlement is expected to lie between 2.5
- 1.96CTU (or 1.7 mm) and 2.5 + 1.96o-m (or 3.3 mm). The measured
settlement of 1.8 mm is found to lie between this interval.
To determine the reliability index corresponding to, say, an allowable
settlement of 5 mm, equivalent values of AE and bz, which can account for
both systematic uncertainties and inherent variability, must first be obtained
[see (28) and (29)]. In this study, the pile shaft is divided into 10 elements,
hence T = L/10 = 2.35 m. From (29), it is apparent that 11(2.35, 4.8) =
1. As a first step, the equivalent value of AE is computed from (28) by
assuming that the equivalent variance function t"2(T,8^) is equal to 1, giving
A'Es = 0.37. Using this value of L'Es, the ratio AJA'Es is evaluated as 0.46
and the corresponding value of b'JL is then read off from Fig. 5, giving 5^

739
 = 0.7 X 23.5 = 16.6 m. Note that the initial assumption regarding the
value of n(7',6;) is correct since 8^ is greater than T. On the other hand,
if 8' were to be less than T, it would then be necessary to reevaluate &'Et
with an updated value of r*(TX), which is given by b'JT. Using this new
value of A'Ks, a corresponding new value of b'JL may be evaluated as pre-
viously discussed. The next step of the analysis is to determine the value of
a from (26) as a = (l/0.17)[(5/2.5) - 1)] = 5.9. From these values of
k'E, 8^/L, and a, the desired reliability index may readily be obtained from
the design chart of Fig. 6 [Fig. 12 of Phoon et al. (1990)] as 0 = 2.6. The
corresponding probability of unserviceable behavior is given by pm = 4>( — 2.6)
= 0.005. The exact solution obtained from a stochastic finite-element com-
puter program is 0 = 2.5 (or pm = 0.006). This relatively good agreement
indicates that the reliability index may be estimated from the design charts
using the procedure developed in this study in place of using a stochastic
finite-element program.

CONCLUSIONS

The following conclusions may be drawn from this study. The design
charts generated by Phoon et al. (1990) are accurate for the range of un-
certainties encountered in practice based on results from second-order anal-
ysis.
The uncertainties due to inherent variability of the shear strength of the
soil, the correlation model, and the limited amount of data used in the
estimation of the Young's modulus of the soil can be easily accommodated
in the usage of the design charts. A numerical example showed that the
results obtained are in agreement with those obtained using a stochastic
finite-element program.
The uncertainty due to the correlation model can be significant. This can
be reduced by refinement of the statistics through a more detailed data
classification and analysis according to the type of sampling method and
test procedure.

APPENDIX I. REFERENCES

Asaoka, A., and Grivas, D. A. (1982). "Spatial variability of the undrained strength
of clays."/. Geotech. Engrg. Div.,ASCE, 108(5), 743-756.
Aschenbrenner, T. B., and Olson, R. E. (1984). "Prediction of settlement of single
piles in clay." Analysis and design of pile foundations. J. R. Meyer, ed., ASCE,
New York, N.Y., 41-58.
Butterfield, R., and Ghosh, N. (1980). "A linear elastic interpretation of model tests
on single piles and groups of piles in clay." Int. Conf. on Numerical Methods in
Offshore Piling, Institution of Civil Engineers (ICE), 109-118.
Cooke, R. W., Price, G., and Tarr, K. (1979). "Jacked piles in London clay: A
study of load transfer and settlement under working conditions." Geotechnique,
London, England, 29(2), 113-147.
Cragg, C. B. H., and Krishnasamy, S. G. (1987). "Probabilistic analysis of trans-
mission tower foundations in clay." Proc, 5th Int. Conf. on Applications of Sta-
tistics and Probability in Soil and Struct. Engrg., Institute for Risk Research,
University of Waterloo, Vancouver, Canada, 805-812.
Hooper, J. A., and Butler, F. G. (1966). "Some numerical results concerning the
shear strength of London clay." Geotechnique, London, England, 16(4), 282-304.
Ladd, C. C, Moh, Z-C, and Gifford, D. G (1971). "Statistical analysis of undrained
strength of soft Bangkok clay." Proc, 1st Int. Conf. on Applications of Statistics
and Probability in Soil and Struct. Engrg., University of Hong Kong, Hong Kong,
313-328.
740
Lurnb, P. (1975). "Spatial variability of soil properties." Proc, 2nd Int. Conf. on
Applications of Statistics and Probability in Soil and Struct. Engrg., Deutsche
Geseflschaft, Aachen, West Germany, 397-421.
Matsuo, M., and Asaoka, A. (1977). "Probability models of undrained strength of
marine clay layer." Soils and Found., 17(3), 53-68.
Meyerhof, G. G., Brown, J. D., and Mouland, G. D. (1981). "Prediction of friction
pile capacity in a Till." Proc, 10th Int. Conf. on Soil Mech. and Found. Engrg.,
A. A. Balkemar, Stockholm, Sweden, 2, 777-780.
Ottaviani, M , and Marchetti, S. (1979). "Observed and predicted test pile behav-
iour." Int. J. for Numerical and Analytical Methods in Geomech., 3(2), 131-143.
Phoon, K. K., Quek, S. T., Chow, Y. K. and Lee, S. L. (1990). "Reliability analysis
of pile settlement." J. Geotech. Engrg., ASCE, 116(11), 1717-1735.
Poulos, H. G. (1972). "Load-settlement prediction for piles and piers." J. Soil Mech.
Found. Div., ASCE, 98(9), 879-897.
Poulos, H. G., and Davis, E. H. (1980). Pile foundation analysis and design. John
Wiley & Sons, New York, N.Y.
Schultz, E. (1975). "Some aspects concerning the application of statistics and prob-
ability to foundation structures." Proc, 2nd Int. Conf. on Applications of Statistics
and Probability in Soil and Struct. Engrg., Aachen, West Germany, 457-494.
Singh, A. (1971). "How reliable is the factor of safety in foundation engineering."
Proc, 1st Int. Conf. on Applications of Statistics and Probability in Soil and Struct.
Engrg., Hong Kong, 389-424.
Vanmarcke, E. H. (1977). "Reliability of Earth Slopes." /. Geotech. Engrg. Div.,
ASCE, 103(11), 1247-1265.
Vanmarcke, E. H. (1983). Random fields: Analysis and synthesis. The Massachusetts
Institute of Technology Press, Cambridge, Mass.
Ward, W. H., Samuels, S. G., and Butler, M. E. (1959). "Further studies of the
properties of London clay." Geotechnique, London, England, 9(2), 33-58.
Wu, T. H. (1974). "Uncertainty, safety, and decision in soil engineering." /. Geotech.
Engrg. Div., ASCE, 100(3), 329-348.
Yamamoto, M. (1982). "Reliability analysis of braced excavation," PhD thesis Uni-
versity of Illinois at Urbana-Champaign, Urbana, 111.

APPENDIX II. NOTATIONS

The following symbols are used in this paper:

COV[ , ] = covariance operator;

c„, c„ = undrained shear strength of soil, mean of cu;
d = diameter of pile;
E[ ] = expectation operator;
Ep = Young's modulus soil at point, mean of Es;
Es, Es = Young's modulus of soil at point, mean of Es;
exp() = exponential function;
g(<*>) = performance function expressed in terms of co;
4 = Z^dw/P;
K = stiffness factor;
K, K = stiffness matrix of pile-soil system, mean of K;
L = length of pile;
M = number of pile-load tests used to determine correlation be-
tween Es and c„;
N = number of sampling points used to determine cu;
Na = number of observations in a depth interval;
Nb = number of depth intervals;
Nc = number of observation pairs used in computing correlation;

741
 n = number of nodes in group of piles;
P = external applied load acting on pile head;
=
Pus probability of unserviceable behavior;
Q = vector of external applied loads;
u, u = vector of nodal displacements, mean of u;
VAR[ ] = variance operator;
x = local spatial average for stochastic finite element;
x = vector of x, (i = 1 to £);
a = number of standard deviations between allowable settlement
and mean settlement;
P = reliability index;
Fl() = variance reduction function in vertical direction;
ACu = coefficient of variation of undrained shear strength;
A£ = coefficient of variation of soil Young's modulus;
AE = equivalent value of A £ j that considers both inherent variability
and systematic uncertainties;
As = coefficient of variation of w due to systematic uncertainty;
Ax = coefficient of variation of x;
Aw = coefficient of variation of co;
Aj, = coefficient of variation of w due to inherent variability;
Br = radial scale of fluctuation;
Sz = vertical scale of fluctuation;
82 = equivalent value of 8Z that considers both inherent variability
and systematic uncertainties;
£ = number of elements used in stochastic finite-element method;
p() = correlation function;
<rw = standard deviation of w;
<J>() = standard normal cumulative distribution;
<f>,- = xt - Es;
X, X ~ correlation parameter, mean of x;
ft, = domain in which soil modulus is averaged;
co, w = settlement of pile or pile group, mean of w; and
w„ = allowable settlement.

742