Structural Safety 73 (2018) 99–113

Contents lists available at ScienceDirect

Structural Safety
journal homepage: www.elsevier.com/locate/strusafe

Effective Young’s modulus of a spatially variable soil mass under a footing T

a,⁎ a b
Jianye Ching , Yu-Gang Hu , Kok-Kwang Phoon
a
Dept of Civil Engineering, National Taiwan University, Taipei, Taiwan
b
Dept of Civil and Environmental Engineering, National University of Singapore, Singapore

A R T I C LE I N FO A B S T R A C T

Keywords: This study investigates the possibility of representing the effective Young’s modulus (Eeff) for a footing problem
Spatial variability supported on a spatially variable medium - the Young’s modulus actually “felt” by the footing - using a spatial
Young’s modulus average. The Eeff is simulated by a homogenization procedure that matches the responses between a random
Homogenization finite element analysis (RFEA) and a homogeneous finite element analysis. Emphasis is placed on whether the
Random field
spatial average can well represent the numerical value of Eeff in each spatially varying realization, not just the
Footing
statistics of Eeff within an ensemble (a weaker requirement). It is found that the conventional spatial averaging
model that treats all soil regions equally important cannot satisfactorily represent Eeff. Extensive numerical
results show that the concept of “mobilization” is essential: highly mobilized soil regions close to the footing
should be given larger weights than non-mobilized remote regions. Moreover, the non-uniform weights can be
prescribed prior to RFEA, that is, they do not depend on the specific response corresponding to a specific random
field realization. The “prescribed mobilization” for the spatially variable Young’s modulus can be contrasted
with the “emergent” mobilized shear strength in a spatially variable medium that results from the emergent
nature of the critical failure path – it cannot be predicted prior to random finite element analysis. A key con-
tribution of this paper is the development of a simple method based on the “pseudo incremental energy” to
estimate the non-uniform weights for the spatial averaging using a single run of a homogeneous finite element
analysis.

1. Introduction homogeneous field is described by a suitable spatial average. Based on

The spatial variability of soil parameters has profound impact to the
behavior of a geotechnical structure. In the literature, the impact of a
spatially variable soil Young’s modulus (E) on footing settlements has
been widely studied [23,21,10,11,17,14,20,25,1,2,3]. For footings on
soils with isotropic scales of fluctuation (SOF), an important observa-
tion made in Fenton and Griffiths [10,11] is that the statistics (mean
and variance) for the “effective” Young’s modulus (Eeff) are similar to
those for the geometric average (Eg) over a prescribed domain under
the footing. It is important to emphasize that Eeff is determined from the
deformation response of a random finite element analysis (RFEA) (i.e.,
it is an output of a boundary value problem such as a footing applying
pressure on top of a semi-infinite soil domain). The various spatial
averages can be calculated from the input random field describing the
spatial distribution of the Young’s modulus over the semi-infinite soil
domain (i.e., they are inputs unrelated to the boundary value problem).
It was reported in Fenton and Griffiths [10,11] that the settlement of
a footing overlying a random field and a homogeneous field is the
“same” in terms of second-moment statistics (mean and variance) if the
RFEA, Ching et al. [5,7] obtained a stronger conclusion for a soil
square/cube subjected to displacement-controlled compression. They
found that not only the statistics of Eeff can be well represented by a
suitable spatial average but also Eeff is very strongly correlated to the
spatial average. Note that the similarity in the statistics does not imply a
very strong correlation, e.g., two random variables can have same
statistics and yet be completely uncorrelated. This is a subtle but im-
portant point – a strong correlation between two random entities im-
plies approximate agreement at the realization level, which is critical for
reliability analysis, while comparable statistics merely imply agreement
at the ensemble level. The latter is a weaker condition in the sense that
the former implies that latter but not vice versa. Moreover, Ching et al.
[7] found that this very strong correlation only exists for the soil
square/cube subjected to displacement-controlled compression, not for
the footing problem. For the footing problem, there is a significant
scatter between Eeff and the spatial average, even though the statistics
of Eeff can be well represented by the spatial average.
The purpose of this study is to propose a new spatial averaging
method for the footing problem so that not only Eeff and the spatial

⁎
Corresponding author.
E-mail address: jyching@ntu.edu.tw (J. Ching).

https://doi.org/10.1016/j.strusafe.2018.03.004
Received 26 September 2016; Received in revised form 16 March 2018; Accepted 23 March 2018
Available online 30 March 2018
0167-4730/ © 2018 Elsevier Ltd. All rights reserved.
J. Ching et al.
average have similar statistics but they are also very strongly corre-
lated, i.e., a new spatial averaging method that can effectively predict
the “numerical value” of Eeff in each realization, not only its statistics at
the ensemble level. It will be clear that the resulting spatial average is
not a uniform “mobilization” but a non-uniform mobilization. The soil
elements significantly influenced by the footing load are highly mobi-
lized, whereas those remote to the footing have negligible mobilization.
More importantly, it is found that the degree of mobilization can be
well quantified by a certain quantity that is derived from the stress/
strain change due to the footing load, and the spatial distribution of
such a quantity can be obtained by a single run of a deterministic finite
element analysis (FEA). The latter point is of critical practical im-
portance. With the new spatial averaging method, it will be possible to
simplify a RFEA involving a random field to a random variable problem
which is less costly and perhaps more importantly, make probabilistic
design more accessible to engineers. A single deterministic FEA is
needed as a pre-processing step to compose the new spatial average
(which is a random variable), but the cost of a single run is negligible
compared to thousands of runs in a RFEA.
An important conclusion of this study is that the effective Young’s
modulus can be well represented by the spatial averaging with a pre-
scribed non-uniform mobilization, which means that the non-uniform
mobilization can be determined apart from the random field realiza-
tions and prior to RFEA. This is in contrast with the observations for
shear strength: the authors [4,18,6,8] have shown that the mobilized
shear strength is not the average along a prescribed curve but the
average along the critical slip curve, which is an emergent curve that
changes from realization to realization and cannot be predicted prior to
RFEA. The possible reason for why this is true will be discussed in this
paper.
2. Footing problem under investigation
2.1. Random field model
Consider a footing on a two-dimensional (2D) spatially variable soil
mass, modeled by finite elements (FE) as shown in Fig. 1. The size of the
footing is B = 2 m, and the soil mass has horizontal dimen-
sion = L = 20 m and depth = D = 10 m. The spatially variable Young’s
modulus, denoted by E(x,z), is modeled as a stationary lognormal
random field with inherent mean = μ and inherent coefficient of var-
iation (COV) = V. To define the correlation structure between two lo-
cations with horizontal interval distance = Δx and vertical interval
distance = Δz, the single exponential auto-correlation model is con-
sidered [27–28]:
ρ(Δx,Δz) = exp(−2|Δx|/ δ x−2|Δz|/ δz) (1)
where δx and δz are the SOFs in the (x,z) directions, respectively, for the
Fig. 1. Realization of the E random field for the 2D footing problem with
δx = δz = 1 m.
100
Structural Safety 73 (2018) 99–113
ln[E(x,z)] random field. The ln[E(x,z)] random field is a 2D stationary
normal random field with mean = λ = ln[μ/(1 + V2)0.5] and var-
iance = ξ2 = ln(1 + V2). In this study, the local average for ln(E) of
each FE is taken to be the arithmetic average for ln[E(x,z)] over the
element, and this local average is simulated using the Fourier series
method [19,9]. The E of each FE is simulated as the exponential of its
local average for ln(E). The E of each FE is in fact the geometric average
for E(x,z) over the element. The mean value of this E is not exactly μ
2
[13]. The correction factor (1 + V2)0.5×(1-Γ ) is multiplied to the si-
mulated E to ensure the mean value is μ, where Γ2 is the variance re-
duction factor for the local averaging effect (see Eq. 22 in [13] for the
correction factor and Eq. 15 for the definition of Γ2). In the case that the
sizes of the element are significantly smaller than the SOFs, a geometric
average is roughly the same as an arithmetic average (correction factor
≈ 1). Fig. 1 shows a realization of the E random field with
δx = δz = 1 m. The E value for the light region is low, while that for the
dark region is high. The Poisson’s ratio (ν) is assumed to be constant
(ν = 0.3), because the impact of the spatial variability of the Poisson’s
ratio is insignificant [10,11,5].
2.2. Finite element model
The 20 m × 10 m plane strain rectangular domain is modeled by the
FE mesh shown in Fig. 1. Each FE is a 4-noded element of
size = 0.2 m × 0.2 m. In total, there are 100 × 50 = 5000 elements
with reduced integration (CPE4R). Each FE follows an isotropic elasti-
city model with E = its local geometric average multiplied by the cor-
2
rection factor (1 + V2)0.5×(1-Γ ), ν = 0.3, and unit weight γ = 20 kN/
3
m . The nodes along the vertical boundary are constrained against
horizontal displacement (roller, see Fig. 1), whereas the nodes on the
bottom boundary are fixed (hinge). The top boundary is free, except the
B = 2 m line segment under the footing. The footing is assumed to be
rigid and the soil-footing interface is assumed to be rough. The Young’s
modulus of the soil mass is modeled as a stationary lognormal random
field with inherent mean = μ = 20,000 kN/m2 and inherent coefficient
of variation V = 1.0. Cases with δx = δz = δ will be first considered.
Five SOFs are considered: δ = 1 m, 2 m, 5 m, 10 m, 100 m, and 1000 m
(δ/B = 0.5, 1, 2.5, 5, 50, and 500). For each δ, one thousand realiza-
tions of E random fields are simulated.
2.3. Simulation of effective Young’s moduli
For each random field realization, a geostatic step is adopted to
build up the in-situ stress field over the entire soil mass. Then, the
footing is loaded with a vertical downward uniform displacement
uz = 0.1 m in the FE simulation, not allowing any rotations. This is an
important practical case – footings cannot rotate because they are
constrained by ground beams. Footings can rotate in other cases, e.g. a
monopile supporting a wind turbine. This (a footing that can rotate)
will be addressed in our future work. The resulting total contact force
between the footing and the soil mass is recorded. Another FE simu-
lation with homogeneous E is conducted, following the same geostatic
step and the same displacement-controlled loading. The homogeneous
E value is adjusted until the total contact force matches that for the
random field realization. The adjusted E value is called the effective
Young’s modulus, Eeff, for the random field realization. This process is
sometimes called “homogenization” [16,15,22]. It is worth empha-
sizing Eeff is a response (or an output) from a RFEA.
2.4. Results for cases with δx = δz
Fig. 2 shows the pairwise plot for the simulated Eeff/μ versus (spatial
average)/μ for cases with δx = δz = δ, where the spatial average is the
geometric average (Eg) over the 1B × 5B domain under the footing.
This averaging domain was considered in Fenton and Griffiths [10].
This spatial averaging model will be denoted by the Eg model. Although
J. Ching et al.
Fig. 2. Eeff/μ versus Eg/μ relationships for cases with δx = δz = δ.
the statistics (e.g., mean and COV) of Eeff and Eg are similar: the cor-
relation coefficient between them is not very strong when δ/B is rela-
tively small, e.g., the Pearson correlation coefficient ρ = 0.68 and 0.80
for δ/B = 0.5 and 1, respectively. We also considered other averaging
domains, e.g., the 1B × 2B or 2B × 5B averaging domain under the
footing, but the conclusion remains: they are not very strongly corre-
lated when δ/B is relatively small. The correlation is very strong when
δ/B is relatively large (δ/B = 50 & 500) because the random field ap-
proaches a constant field, and the problem approaches a homogeneous
problem. Any averaging domain will perform well in this circumstance.
Besides the correlation coefficient (ρ) between Eeff and Eg, the model
factor, defined by M = Eeff/Eg, is also adopted to quantify how well Eeff
is represented by Eg. The M for a geotechnical calculation model is
defined as the ratio of the measured response to the calculated re-
sponse. Examples are pile capacity and settlement of a footing. In this
paper, the ratio M = Eeff/Eg is defined in a similar way. The effective
Young’s modulus from a RFEA, Eeff, is treated as a “measured” response.
The spatial average, Eg, is treated as a calculated response. Each rea-
lization of M is obtained from a pair of (Eeff, Eg) that is computed using
the same random field realization. The mean value of M, denoted by b,
quantifies the model bias, whereas the COV of M, denoted by Δ,
quantifies the model variability. Eeff can be well represented by Eg if (a)
ρ is close to unity and (b) Δ is close to zero. It is desirable to have b close
to unity as well, but the issue of b ≠ 1 can be resolved by multiplying a
bias factor to Eg. The solid square markers (■) in Fig. 3 show how ρ, b,
and Δ vary with respect to δ/B. It is clear that Eeff can be well re-
presented by Eg only when δ/B is relatively large (δ/B = 50 & 500)
where ρ is close to unity and Δ is close to zero. However, in general Eeff
cannot be well represented by Eg because ρ can deviate significantly
from unity when δ/B is relatively small and Δ can deviate significantly
from zero at intermediate δ/B.
3. New spatial averaging method with non-uniform mobilization
One possible explanation for why Eeff cannot be well represented by
Eg is that elements are not mobilized uniformly. Uniform mobilization
101
Structural Safety 73 (2018) 99–113
over the 1B × 5B domain is assumed for the Eg model. However, the
governing volume primarily lies below the footing to a depth of several
times the footing diameter. The soil volume remote from the footing is
not mobilized. A key contribution of this study is to propose a new
spatial averaging method with non-uniform mobilization.
The non-uniform mobilization will be quantified by unequal
weights that are to be calibrated by RFEA results. Recall that there are
six SOFs and one thousand realizations of RFEA for each SOF, produ-
cing n = 6000 “calibration cases” in total. Also recall that there are
m = 5000 elements in the FE mesh. For the k-th calibration case, there
is a realization of Eeff,k and a realization of (E1,k, E2,k, …, E5000,k), where
Ei,k is the Young’s modulus assigned to the i-th finite element for the k-
th calibration case. The latter (E1,k, E2,k, …, E5000,k) is denoted by Eall,k,
the set the E values for all FEs. It is clear that Eeff,k depends on Eall,k,
namely Eeff,k = g(Eall,k). The key question is: what is the functional
form for g(.)? This question can be addressed by conducting regression
analysis over the n = 6000 sets of (Eeff, Eall) data points. To keep the
problem simple, linear regression in the form of “weighted geometric
average” is adopted in this study:
m
ln(Eeff,k) ≈ ln(E wg,k) = ∑ wi × ln(Ei,k) = (lnEall,k)T × w
i= 1 (2)
where Ewg denotes the weighted geometric average; wi is the weight for
the i-th element; m = 5000 is the total number of elements; ln(Eall,k)
denotes the column vector [ln(E1,k) ln(E2,k) …ln(E5000,k)]T, where the
superscript ‘T’ means the matrix transpose; and w = (w1, w2, …,
w5000)T. The weight wi quantifies the degree of mobilization for the i-th
element. The i-th element is more mobilized than the j-th element if
|wi| > |wj|. The Eg model is a special case of Eq. (2) with all weights
being 1/(# of elements inside the domain) inside the 1B × 5B domain
and zero outside the domain.
The unknown weights w may be calibrated using the ordinary least-
square (OLS) method based on the n = 6000 calibration cases to
minimize the sum of squared errors. The errors for the calibration cases
are defined as
J. Ching et al. Structural Safety 73 (2018) 99–113

Fig. 3. Variations of ρ, b, and Δ with respect to δ/B for various spatial averaging models. (ρ = correlation between Eeff and spatial average; b = mean of Eeff/spatial
average; Δ = COV of Eeff/spatial average).

e ⎡ ln(Eeff,1) ⎤ ⎡ ln(E1,1) ln(E2,1) ⋯ ln(Em,1) ⎤ w1 λ = 0. With the penalty term, the solution for wk will be close to zero
⎡ 1⎤ ⎢ ⎥ ⎢ ln(E1,2) ln(E2,2) ⋯ ln(Em,2) ⎥ ⎡ w2 ⎤ unless there is sufficient evidence indicating otherwise. It can be shown
e
⎢ ⎥ 2 ln(E )
⎥×⎢ ⋮ ⎥
eff,2
e = ⎢⋮⎥ = ⎢ ⎥−⎢ that the solution for the RLS method is
⎢ ⋮ ⎥ ⎢ ⋮ ⋮ ⋮ ⎥ ⎢ ⎥
⎢ en ⎥ ⎢ ln(E ) ⎥ ⎢ ln(E ) ln(E ) ⋯ ln(E ) ⎥ ⎢ wm ⎥
⎣ ⎦ eff,n 1,n 2,n m,n ⎣ ⎦ * = (ATA + λ × I)−1AT × y
w RLS (5)
⎣ ⎦ ⎣ ⎦
= y−A × w (3) where I is the identity matrix. The Tikhonov factor λ can be fine-tuned
using cross-validation [24]: divide the n = 6000 calibration data into N
where ek is the error for the k-th calibration case; e = [e1 e2 … en] ; T
folds and compute w∗RLS based on (N-1) folds of data and then calculate
y = [ln(Eeff,1) ln(Eeff,2) … ln(Eeff,n)]T; A = the (n × m) matrix con- the sum of the squared prediction errors with respect to the remaining
taining all ln(Ei,k). The OLS solution for w minimizing eTe (the sum of one fold of data. Repeat this process for N times by switching the N
square errors) is denoted by w∗OLS = (A × AT)-1×AT×y, where the folds, and the average sum of the squared prediction errors is called the
superscript “-1” denotes matrix inverse. Note that the regression is high N-fold cross-validation error. One popular special case is the leave-one-
dimensional because there are 5000 unknown weights. It is challenging out cross validation. For our case, it is the 6000-fold cross validation.
to conduct the high dimensional regression analysis because of the Golub et al. [12] showed that the optimal λ that minimizes the leave-
potential over-fit problem. The resulting weights w∗OLS seem to be very one-out cross-validation error is also the minimizer of the following
noisy with large amplitudes ranging from −1500 to 1500. In particular, generalized cross-validation (GCV) error:
elements remote to the footing can still have large-amplitude weights.
This is deemed unreasonable. It turns out that the noisy weights can fit * )T (y−A × w RLS
(y−A × w RLS * )
GCV =
the calibration cases very well (an over-fit) but provide extremely poor Tr[I−A (ATA + λ × I)−1AT]2 (6)
fit to independent validation cases. The over-fitted w∗OLS is not very
where Tr(.) denotes sum of the diagonal terms of the enclosed square
meaningful.
matrix. Given the n = 6000 calibration cases, Fig. 4 shows the re-
To suppress the over-fit, the regularized least square (RLS) method
lationship between GCV and λ. The optimal λ (λ∗) is 3845. With
(e.g., Tikhonov regularization; see [26]) is adopted to determine the
λ∗ = 3845, w∗RLS can be found by Eq. (5). Fig. 5 shows the grey scale
unknown weights w by introducing an extra penalty term to encourage
plot for the resulting w∗RLS. The sum of all weights is 0.9952, very close
a “regularized” w solution. This RLS method is also called the “ridge
to unity. It is interesting to note that the amplitudes of w∗RLS are now
regression”. While the OLS method is to minimize the sum of squared
greatly suppressed and that elements remote to the footing have neg-
errors eTe, the RLS method (ridge regression) is to minimize the fol-
ligible weights. Equation (2) with weights equal to w∗RLS will be denoted
lowing objective function:
by the w∗RLS model.
min eTe + λ × wTw It is found that the w∗RLS model not only provides excellent fit to the
w (4)
calibration cases but also provides satisfactory prediction to the in-
The penalty term λ×w w discourages large weights. The parameter λ
T
dependent validation cases. Fig. 6 shows the pairwise plot for Eeff/μ
is called the Tikhonov factor: it determines the relative importance for versus Ewg/μ for the calibration cases (Ewg calculated using Eq. (2) with
the penalty. The OLS method is equivalent to the RLS method with weights equal to w∗RLS): the fit is excellent with very strong correlation

102
J. Ching et al.
Fig. 4. Relationship between GCV and λ.
Fig. 5. Grey scale plot for w*RLS.
for all δ/B. The results for the extra validation cases are shown in Fig. 7.
The prediction is also satisfactory. The issue of over-fit is resolved using
the RLS method, and the resulting weights w∗RLS are meaningful. The
prediction results for the validation cases in Fig. 7 can be compared
with those in Fig. 2. The former (Fig. 7) is for the w∗RLS model, whereas
the latter (Fig. 2) is for the Eg model. Both are genuine predictions. It is
evident that the predictions based on the w∗RLS model are superior to
those based on the Eg model, with correlation coefficients very close to
unity for all δ/B. The superior performance of the w∗RLS model can be
seen in Fig. 3 as well: the open circle markers (o) show how the cor-
relation between Eeff and Ewg (ρ), the sample mean of M = Eeff/Ewg (b),
and the sample COV of M = Eeff/Ewg (Δ) vary with respect to δ/B.
Compared with the Eg model (the square markers in Fig. 3), the w∗RLS
model can now better represent Eeff in general because ρ is closer to
unity and Δ is closer to zero. It is noteworthy that the sample mean of M
(b) does not converge to unity when δ/B is large. This is because the
summation of the weights w∗RLS is not exactly 1. If the weights w∗RLS are
normalized so that they sum up to 1, b will converge to unity when δ/B
is large.
Despite the superior performance of the w∗RLS model, there remains
a small degree of scatter between Eeff and Ewg in Fig. 7. There are at
least three possible explanations for the scatter:
1. The actual functional relationship between Eeff and Eall may be
different from the relationship assumed in Eq. (2). Equation (2) assumes
a linear relationship between ln(Eeff) and ln(Eall). The actual relation-
ship may not be linear. The linear relationship is assumed for tract-
ability. Even if the actual relationship is linear, it may not exist between
the logarithms of Eeff and Eall. It may exist between the Eeff and Eall
themselves (weighted arithmetic averaging). It may also exist between
the reciprocals of Eeff and Eall (weighted harmonic averaging). The
103
Structural Safety 73 (2018) 99–113
effect of adopting a different type of the weighted averaging model is
investigated. Taking the weighted arithmetic averaging as an example,
the following relationship between Eeff and Eall is assumed:
m
Eeff,k ≈ E wa,k = ∑ wi × Ei,k = (Eall,k)T × w
i= 1 (7)
where Ewa denotes the weighted arithmetic average. With this model, y
in Eq. (3) is now [Eeff,1 Eeff,2 … Eeff,n]T and A is now the (n × m) matrix
containing all Ei,k. The optimal λ∗ can be solved by minimizing Eq. (6.),
and w∗RLS can be obtained from Eq. (5). Given the random field reali-
zation [E1 E2 … E5000] for each validation case, Eq. (7) is adopted to
compute Ewa using the resulting w∗RLS. This Ewa is therefore the pre-
diction result for Eeff for the weighted arithmetic averaging model. The
performance of a different type of the weighted averaging model can be
evaluated by ρ, b, and Δ. Fig. 8 shows how ρ, b, and Δ vary with δ/B for
three different weighted averaging models: the weighted geometric
averaging (Ewg), weighted arithmetic averaging (Ewa), and weighted
harmonic averaging models (Ewh). It can be seen that the Ewg model is
superior to the Ewa and Ewh models, with ρ closer to unity and Δ closer
to zero. The variations in b for the three models are quite different, but
the issue for b is relatively minor. In the following, only the weighted
geometric averaging (Ewg) model will be considered.
2. The actual weights may depend to some degree on the random
field realizations. However, w∗RLS are assumed to be invariant with re-
spect to random field realizations in Eq. (2). Note that once w∗RLS are
determined using Eq. (5), they are fixed. Namely, w∗RLS are assumed to
be prescribed rather than emergent with respect to random field reali-
zations.
3. The amount of the calibration data is insufficient to eliminate the
statistical uncertainty in w∗RLS.
Nonetheless, the small scatter in Fig. 7 suggests that the assumption
for a linear relationship between ln(Eeff) and ln(Eall) and the assump-
tion for prescribed w∗RLS are acceptable.
3.1. Prescribed versus emergent
The authors [4,18,6,8] have shown that for spatially variable shear
strength, the mobilized shear strength is not the average along a pre-
scribed curve but along the critical slip curve, an emergent curve that
depends on the response of a RFEA in each realization. However, the
numerical observation here for the effective Young’s modulus indicates
the contrary. It appears to be possible to prescribe weights w∗RLS prior to
RFEA. The prescribed mobilization for the Young’s modulus is in con-
trast with the emergent mobilization for the shear strength. This sig-
nificant contrast can be explained as follows. The shear failure is ty-
pically localized along a shear band, and the location of the shear band
depends on the random field realization because it may seek the weak
zone which varies from realization to realization. Hence, the mobili-
zation for shear strength is emergent. However, such localization is ty-
pically not clearly developed under an elastic condition that is far from
the failure condition, and it may be reasonable to prescribe the mobi-
lization for the Young’s modulus prior to RFEA.
3.2. Uniform versus non-uniform mobilization
Although the mobilization for Young’s modulus can be prescribed, it
is not the conventional spatial averaging with uniform mobilization,
e.g., the Eg model. According to our numerical experiences for the
footing problem, when δ/B is less than 10, a uniform mobilization over
any prescribed domain cannot produce Eg that is very strongly corre-
lated to Eeff. However, this does not imply that the uniform mobiliza-
tion is ineffective for all problems. We have mentioned in Introduction
that for a soil square/cube subjected to displacement-controlled com-
pression, the uniform mobilization is fairly effective: it can produce Eg
that is very strongly correlated to Eeff. Later, we will propose a non-
J. Ching et al. Structural Safety 73 (2018) 99–113

Fig. 6. Eeff/μ versus Ewg/μ relationships (calibration cases, Ewg computed by the w*RLS model).

uniform mobilization model that is effective for the footing problem.
Moreover, this model reduces to uniform mobilization for the soil
square/cube problem.
3.3. Correlation to stress/strain factors
The preceding results only demonstrate that a weighted spatial
averaging scheme is more effective than the conventional spatial
averaging scheme. The methodology is however not practical. After all,
the purpose of applying the approximation Eg ≈ Eeff is to circumvent
tedious and costly RFEA, but the weights (w∗RLS) require RFEA which
defeat the practical goal of homogenization. The purpose of this section
is to show that w∗RLS can be estimated using a single deterministic FEA,
i.e. the elaborate regularized least square (RLS) regression on RFEA
results described by Eqs. (4) to (6) is not necessary. It was undertaken in
this paper to demonstrate that non-uniform mobilization of the Young’s
modulus in different finite elements is implied by the RFEA results. The
rest of this section demonstrates how this non-uniform mobilization can
be characterized without RFEA.
The elements significantly influenced by the footing load (e.g.,
elements right below the footing) should have significant weights.
Those not influenced by the footing load (e.g., elements remote to the
footing) should have negligible weights. It is expected that w∗RLS should
be correlated to factors such as stress/strain increments due to the
footing load. Let (Δσx, Δσy, Δσz, Δτxz) and (Δεx, Δεz, Δεxz) (y is the out-
of-plane direction) be the stress and strain increments of an element,
respectively, due to the footing load. For instance, the increment in σx is
denoted by Δσx = (σx after the footing load) – (σx before the footing
load), and similar for other stress/strain increments. Three categories of
such “mobilization factors” that quantifies the degree of mobilization
are considered:
1. Five invariants for the stress increments (Δτxy = Δτyz = 0 for plane
strain):
ΔI1 = Δσx + Δσy + Δσz
ΔI2 = Δσx Δσy + Δσx Δσz + Δσy Δσz−Δτ 2xz
ΔI3 = Δσx Δσy Δσz−Δσy Δτ 2xz
ΔJ2 = ΔI12 /3−ΔI2
ΔJ3 = 2ΔI13 /27−ΔI1ΔI2 /3 + ΔI3 (8)
In general, ΔI and ΔJ are not the same as the increments in I and J,
because ΔI and ΔJ are not linear in the stress increments. ΔI and ΔJ are
adopted here because they are more strongly correlated to w∗RLS than
the increments in I and J.
2. Four invariants for the strain increments (Δεy = Δεxy = Δεyz = 0 for
plane strain):

104
J. Ching et al.
Fig. 7. Eeff/μ versus Ewg/μ relationships (validation cases, Ewg computed by the w*RLS model).
ΔI′1 = Δε x + Δεz
ΔI′2 = Δε x Δεz−Δε 2xz
ΔJ′2 = ΔI′12 /3−ΔI′2
ΔJ′3 = 2ΔI′13 /27−ΔI′1 ΔI′2 /3
The prime sign is to indicate that these are strain invariants.
3. The dot product between the stress and strain increments. For
plane strain, the stress increments are (Δσx, Δσy, Δσz, Δτxz), whereas the
strain increments are (Δεx, Δεy, Δεz, 2Δεxz) (Δεy = 0). The dot product
between them is denoted by ΔU:
Δ U= Δσx Δε x + ΔσzΔεz + 2Δτ xzΔε xz
Note that ΔU is not the same as two times of the incremental strain
energy. The incremental strain energy is defined as (σxΔεx + σzΔεz +
2τxzΔεxz)/2, not (ΔσxΔεx + ΔσzΔεz + 2ΔτxzΔεxz)/2. We termed ΔU as
“pseudo incremental energy”. The incremental strain energy is not
adopted here because it is not strongly correlated to w∗RLS.
It is possible to produce more mobilization factors by the linear
combination among various stress/strain factors. However, the linear
combination among various stress/strain factors lacks physical sig-
nificance. Also, the performance is not significantly better than that
provided by single factors. Therefore, more complicated linear combi-
nations are not pursued.
It will be later shown that the pseudo incremental energy ΔU is the
most effective among the ten mobilization factors in the sense that the
resulting Ewg is the most strongly correlated to Eeff. To implement the
pseudo incremental energy, a deterministic FEA is conducted to obtain
Structural Safety 73 (2018) 99–113
the ΔU values for all 5000 elements. These values are normalized such
that the summation over all 5000 elements is equal to unity to obtained
the normalized ΔU, denoted by ΔUn. Recall that the 5000 w∗RLS values
roughly sum up to unity as well. The mesh, boundary conditions, and
(9)
loading conditions for the deterministic FEA are identical to the RFEA,
but the Young’s moduli for all elements are fixed at the mean value
20,000 kN/m2. A deterministic FEA, rather than a RFEA, is adopted
because the purpose of ΔUn is to approximate w∗RLS. Note that w∗RLS are
derived from the ensemble of 6000 RFEA simulations. Therefore, w∗RLS
represent the average behavior of the 6000 RFEA realizations, not of a
particular RFEA realization. If the purpose of ΔUn is to approximate
(10)
w∗RLS, it is therefore more reasonable to compute ΔUn not from a par-
ticular RFEA realization but from the ensemble of RFEAs. In this paper,
a deterministic FEA with element properties fixed at the mean value is
adopted for this purpose. In fact, the performance for ΔUn that varies
with each random field realization is very similar to that for ΔUn based
on one deterministic FEA in terms of ρ, b, and Δ. From a practical
viewpoint, it is critical to estimate w∗RLS without involving RFEA.
Therefore, ΔUn based on one deterministic FEA is recommended as a
cheap replacement for w∗RLS in this study. In the conventional spatial
averaging approach, Eg can be calculated directly from the input
random field. No FEA is needed at all, much less RFEA. The proposed
weighted spatial averaging approach described below will impose an
additional cost of one deterministic FEA, but this cost is deemed
worthwhile given the significant improvements achieved over the
conventional spatial averaging approach.
105
J. Ching et al.
Fig. 8. Variations of ρ, b, and Δ with respect to δ/B for various weighted averaging methods. (ρ = correlation between Eeff and spatial average; b = mean of Eeff/
spatial average; Δ = COV of Eeff/spatial average).
3.4. Mechanical basis for ΔUn?
Ching et al. [7] numerically observed that conventional spatial
averages (arithmetic, geometric, harmonic averages) may or may not
work as an effective Young's modulus depending on whether the
random field is isotropic or anisotropic. For the latter, it matters whe-
ther the horizontal scale of fluctuation is higher than the vertical scale
(soil layers) or vice versa (soil columns). In other words, there is no
mechanical basis underlying conventional spatial averages (if there is a
mechanical basis, it should work for all random fields). The geometric
average Eg (Eq. (2) with wi = 1/m) only works for isotropic fields. This
paper builds on the existing literature by considering the case of iso-
tropic fields where the soil elements are not uniformly mobilized. This
paper conjectures that generalizing the conventional geometric average
to a non-uniformly weighted geometric average (Eq. 2) can address this
situation. Our extensive numerical results support this conjecture.
There is no mechanical basis for the weighted geometric average using
the weights w∗RLS as there is no mechanical basis for the original geo-
metric average. There is no mechanical basis for ΔUn, either, since the
whole logic chain originating from the classical geometric average for
isotropic uniformly mobilized soil elements is primarily based on nu-
merical observations rather than mechanical basis.
3.5. Simplified procedure of simulating Ewg
The following simplified procedure, called the pseudo incremental
energy model, is proposed to simulate the samples of the weighted
geometric average, Ewg. It will be shown that the resulting Ewg is very
strongly correlated to the effective Young’s modulus, Eeff.
1. Perform one deterministic FEA for the footing problem. Compute ΔU
for all elements using Eq. (10), then compute the normalized form
ΔUn so that all ΔUn sum up to unity.
2. Simulate the Young’s moduli for all elements using random field (no
106
Structural Safety 73 (2018) 99–113
need to run RFEA).
3. Compute the following weighted geometric average Ewg:
m
ln(E wg) = ∑ ΔUn,i × ln(Ei)
i= 1 (11)
where Ei is the Young’s modulus assigned to the i-th element; ΔUn,i is
the normalized weight for the i-th element.
Steps 2 and 3 can be repeated to obtain a different sample of Ewg
corresponding to a different random field realization. The set of weights
in Eq. (11) (ΔUn,i) computed from a single deterministic FEA is in-
dependent of the random field realization. The set of weights is pre-
scribed in this sense. One can treat the pseudo incremental energy
model as a homogenization procedure for the spatially variable Young’s
modulus. The practical benefit for the pseudo incremental energy
model is obvious: it will be possible to simplify a RFEA problem to a
weighted spatial averaging problem which is less costly and perhaps
more importantly, make probabilistic design more accessible to en-
gineers.
The weighted geometric average Ewg simulated by the pseudo in-
cremental energy model can satisfactorily represent Eeff for the footing
problem. Consider the same footing problem in Fig. 1. Fig. 9 shows the
pairwise plot for Eeff/μ versus Ewg/μ, where Ewg is now simulated by the
pseudo incremental energy model. Fig. 9 can be compared with Fig. 7:
the former is based on the pseudo incremental energy model, whereas
the latter is based on the w∗RLS model (weights optimized to agree with
RFEA responses from all realizations). Some degraded performance is
observed for the pseudo incremental energy model (Fig. 9) as to be
expected: the data points slightly deviate from the 1:1 line for δ/
B = 0.5 and 1.0, but the calculation of Ewg is significantly cheaper than
w∗RLS. Note that the performance for the pseudo incremental energy
model (Fig. 9) is significantly better than that for the Eg model (Fig. 2),
although the calculation of Eg is slightly cheaper than Ewg. The per-
formances for various models are further compared in Fig. 3. The cross
marks (×) show how ρ, b, and Δ for the pseudo incremental energy
J. Ching et al.
Fig. 9. Eeff/μ versus Ewg/μ relationships (Ewg simulated by the pseudo incremental energy model).
model vary with respect to δ/B. The pseudo incremental energy model
and the w∗RLS model both perform well, with ρ close to unity and Δ close
to zero. Although the use of ΔUn in place of w∗RLS leads to certain de-
gradation, the degradation is not significant (their ρ and Δ values are
similar in Fig. 3). This indicates that ΔUn is a rational replacement of
w∗RLS. These two models are both superior to the Eg model.
Instead of the pseudo incremental energy model, one can adopt
another mobilization factor. Fig. 10 shows how ρ, b, and Δ vary with
respect to δ/B for some alternative models, including the ΔI1,n, ΔI2,n,
ΔJ2,n, and ΔJ3,n models (the performances of the other models are
worse and hence not shown). These results can be compared with the
results for the pseudo incremental energy model (open circles in
Fig. 10). It is evident that the pseudo incremental energy model per-
forms better than other models, because its ρ is the closest to unity and
its Δ is the closest to zero.
The weighted geometric average Ewg simulated by the pseudo in-
cremental energy model can also satisfactorily represent Eeff for the soil
square/cube problem subjected to displacement-controlled compres-
sion. Note that the deterministic FEA for the square/cube will produce
uniform mobilization factor ΔUn, because all stress/strain increments
are uniform. This suggests that the pseudo incremental energy model
reduces to the Eg model, which is associated with uniform mobilization.
Ching et al. [5,7] have shown that the Eg model can produce Eg that is
very strongly correlated to Eeff for the soil square/cube problem.
Therefore, the pseudo incremental energy model is also effective for the
soil square/cube problem.
107
Structural Safety 73 (2018) 99–113
3.6. Robustness of Eg and Ewg
Both Eg and Ewg are “homogenized” Young’s moduli that approx-
imate Eeff. A critical question is whether they can produce a settlement
that has similar distribution as the settlement for the original spatially
variable random field. To investigate this, the case with δ/B = 2.5 is
considered. For each random field realization, RFEA is used to simulate
the actual settlement, denoted by S, of the footing subjected to a 200kN
loading. For the same random field realization, Eg is computed as the
geometric averaging over the 1B × 5B zone below the footing, whereas
Ewg is computed using Eq. (11). Eg and Ewg are inputs to a homogenous
FEA to find the settlements, denoted by Sg and Swg, respectively. One
thousand random field realizations will produce 1000 samples of S and
(Sg, Swg). It is important to verify whether (Sg, Swg) have similar dis-
tribution as S, because the probability of failure will depend on the
distribution of the settlement for the serviceability limit state. Fig. 11a
shows the empirical cumulative density functions (CDF) for S and (Sg,
Swg). It is clear that the CDFs are similar. It is expected that the statistics
for Swg are similar to those for S, because Ewg and Eeff have similar
numerical values. Fenton and Griffiths [10–11] observed that the sta-
tistics for Eg are similar to those for Eeff. Therefore, it is also expected
that the statistics for Sg are also similar to those for S.
Now consider a different scenario of conditional random field where
E values are measured at depths of 0.5B, 1.5B, 2.5B, 3.5B, and 4.5B
right below the footing during site investigation. Given these measured
E values, conditional E random field can be simulated. Note that the
five locations spread all over the 1B × 5B zone for Eg. The measured
J. Ching et al. Structural Safety 73 (2018) 99–113

Fig. 10. Variations of ρ, b, and Δ with respect to δ/B for some mobilization models. (ρ = correlation between Eeff and Ewg; b = mean of Eeff/Ewg; Δ = COV of Eeff/
Ewg).

values are, respectively, 18.84 MPa, 2.67 MPa, 7.38 MPa, 8.72 MPa,
and 2.64 MPa. These values are simulated by unconditional E random
field at the five locations. For this particular set of measured values, the
measured E values at shallow depth (0.5B) is relatively high
(18.84 MPa). The consequence is that the conditional E values at
shallow depth also tend to be large. One thousand conditional random
field realizations will produce 1000 samples of S and (Sg, Swg). Fig. 11b
shows the empirical cumulative density functions (CDF) for S and (Sg,
Swg). The empirical distribution of Sg now systematically deviates from
that of S (Sg > S), whereas the deviation for Swg is not severe. Note that
Eg assumes uniform mobilization over the 1B × 5B zone. In reality,
however, shallow depth is more mobilized than greater depth. There-
fore, shallow depth is under-weighted in Eg. Recall that the conditional
E values at shallow depth tend to be large. This explains why Sg is
systematically less than S.
The conditional random field scenario demonstrates that the Eg
model does not work for reliability analysis for all scenarios, contrary to
conventional wisdom or expectations. This supports the central thesis
advocated in this paper, namely equivalency between two random
quantities must be in the form of strong correlation rather than simi-
larity in distribution. Incidentally, the former is a stronger equivalency
than the latter, because the former automatically implies the latter, but

Fig. 11. Empirical CDFs of S, Sg, and Swg: (a) unconditional random field; (b) conditional random field.

108
J. Ching et al.
Fig. 12. Eeff/μ versus Ewg/μ relationships for the D = 2 × B scenario.
not vice versa.
3.7. Applicability to other footing problems
We have been investigating the footing problem in Fig. 1. In this
section, we will show that the pseudo incremental energy model also
performs well for scenarios with a different geometry, boundary con-
dition, spatial variability pattern, etc. Some selected scenarios are
presented:
3.7.1. Depth to rigid layer
The soil mass for the problem in Fig. 1 has a depth D = 5 × B.
Beyond this depth is the bottom boundary. Two different FE meshes
with different depths are investigated: D = 2 × B and 10 × B. Figs. 12
and 13 show the Eeff/μ versus Ewg/μ relationships. The correlation
coefficient between Eeff and Ewg (ρ) as well as the COV of M = Eeff/Ewg
(Δ) are annotated in the figures. The pseudo incremental energy model
is effective for both scenarios with ρ close to unity and Δ close to zero.
The Eeff/μ versus Eg/μ relationships for the Eg model are also plotted in
these figures (grey dots), where Eg is simulated as the geometric
average (uniform weights) over the 1B × 2B domain under the footing
for the D = 2 × B scenario and over the 1B × 10B domain for the
D = 10 × B scenario. The Eg model is not as effective as the pseudo
incremental energy model, with ρ sometimes deviating significantly
from unity and Δ sometimes deviating significantly from zero (ρ and Δ
for the Eg model are shown in the parentheses).
109
Structural Safety 73 (2018) 99–113
3.7.2. A flexible footing with a smooth base
The footing in Fig. 1 is rigid with a rough base. A flexible footing
with a smooth base is investigated here. The footing is loaded with a
vertical downward uniform pressure of 100 kN/m2. For the flexible
footing with a smooth base, Eeff is calculated in a different way. For a
random field realization, the resulting average vertical displacement of
the footing base is recorded. Another FE simulation with homogeneous
E is conducted. The homogeneous E value is adjusted until the average
displacement matches that for the random field realization. The ad-
justed E value is therefore the Eeff for the random field realization.
Fig. 14 shows the Eeff/μ versus Ewg/μ plots for the flexible footing with a
smooth base. It is evident that the pseudo incremental energy model is
effective with ρ close to unity and Δ close to zero. For comparison, the
Eeff/μ versus Eg/μ relationships for the Eg model are also plotted in the
figure (grey dots).
3.7.3. A three-dimensional footing problem
We have studied the two-dimensional (2D) strip footing problem
thus far. A three-dimensional (3D) footing problem is investigated here.
Consider a square footing on a 3D spatially variable soil mass, modeled
by FE as shown in Fig. 15. The size of the footing is B × B = 2 m × 2 m,
and the soil mass has plan dimension = L × L = 30 m × 30 m and
depth = D = 5 × B = 10 m. The size for each cubic eight-node element
is 0.5 m × 0.5 m × 0.5 m. The vertical side faces of the soil mass are
constrained against horizontal displacement (roller), whereas the nodes
on the bottom boundary are fixed (hinge). The top boundary is free,
J. Ching et al.
Fig. 13. Eeff/μ versus Ewg/μ relationships for the D = 10 × B scenario.
except the 2 m × 2 m area under the footing. The footing is rigid with a
rough base. Cases with δx = δy = δz = δ are considered. The effective
Young’s modulus Eeff is simulated using the same procedure for a 2D
problem: matching the total contact force from a homogeneous FE
analysis with that from the RFEA. The equation for the pseudo incre-
mental energy is generalized to 3D:
Δ U= Δσx Δε x + Δσy Δε y + ΔσzΔεz + 2Δτ xy Δε xy + 2Δτ yzΔε yz + 2Δτ xzΔε xz
(12)
Fig. 16 shows the Eeff versus Ewg plots for the 3D footing problem. It
is evident that the pseudo incremental energy model is effective with ρ
close to unity and Δ close to zero. For comparison, the Eeff/μ versus Eg/μ
relationship for the Eg model is also plotted in the figure, where Eg is
simulated as the geometric average (uniform weights) over the
1B × 1B×5B domain under the footing. The Eg model is not as effective
as the pseudo incremental energy model, with ρ deviating significantly
from unity and Δ deviating significantly from zero.
Other scenarios are also investigated, although the results are not
shown. They include (a) a FE model with a different mesh density; (b) a
spatial variability with a different mean value; (c) a spatial variability
with a different COV; (d) a spatial variability with different δx and δz (δx
≥ δz); (e) a footing with an embedment depth, (f) a soil mass with a
different Poisson ratio, and (g) a spatial variability with a different
auto-correlation function (squared exponential). For all scenarios, the
pseudo incremental energy model seems to be effective with ρ close to
unity and Δ close to zero.
110
Structural Safety 73 (2018) 99–113
4. Conclusions
This study investigates the possibility of representing the effective
Young’s modulus (Eeff) for a footing supported on a spatially variable
medium using a suitable spatial average. The Eeff of the spatial variable
soil mass is defined as the Young’s modulus actually “felt” by the
footing and is simulated by a homogenization procedure that match the
responses between a random finite element analysis (RFEA) and a
homogeneous finite element analysis (FEA). The purpose is to find a
suitable spatial average strongly correlated to the numerical value of
Eeff in each realization, not just the statistics of Eeff at the ensemble
level. This is a subtle point – a strong correlation between two random
entities implies approximate agreement at the realization level, which is
critical for reliability analysis, while comparable statistics merely imply
agreement at the ensemble level. The latter is a weaker condition in the
sense that the former implies that latter but not vice versa. This is not
an academic difference. The spatial average that only matches the
statistics of Eeff may not reproduce the probability of failure computed
from the original RFEA responses.
It is found that the conventional spatial averaging model that treats
all soil regions equally important cannot satisfactorily represent the Eeff
for the footing problem, regardless of the location and/or size of the
averaging domain. This is in contrast with the soil square/cube problem
previously investigated by the authors (Ching et al. [7]) where the
conventional spatial averaging model is strongly correlated to Eeff.
Further numerical evidences show that the concept of “non-uniform
J. Ching et al.
Fig. 14. Eeff/μ versus Ewg/μ relationships for the smooth footing scenario.
Fig. 15. Realization of the E random field for the 3D footing problem with
δx = δy = δz = 1 m.
mobilization” is essential: highly mobilized soil regions are more im-
portant than non-mobilized regions. For the footing problem, the soil
regions close to the footing are highly mobilized, so a spatial averaging
model with non-uniform weights should be adopted. For the soil
square/cube problem, all regions are equally mobilized, so a spatial
averaging model with uniform weights should be adopted. From a
physical sense, this conclusion is obvious. The practical challenge is
whether this “non-uniform mobilization” can be characterized without
involving RFEA. In the conventional spatial averaging approach, the
111
Structural Safety 73 (2018) 99–113
geometric average of the Young’s modulus can be calculated directly
from the input random field. No RFEA is needed.
A key contribution in this study is to propose a simple method to
estimate the weights for the spatial averaging. The stress/strain dis-
tributions of a homogeneous finite element analysis based on the
average Young’s modulus can be used to deduce the weights. Given a
random field realization for the Young’s modulus, these weights can be
used to simulate the weighted spatial average that is very highly cor-
related to Eeff. Note that the set of weights computed from the above
homogeneous FEA is independent of the random field realization. The
set of weights is prescribed in this sense. The ability to prescribe
weights that can approximately reproduce Eeff (a response from RFEA)
in each random field realization prior to RFEA is very useful for prac-
tice. Previous studies [4,18,6,8] have shown that this attractive feature
does not apply to the mobilized shear strength. The reason is that
mobilized shear strength is not the average along a prescribed curve but
the average along the critical slip curve, which is an emergent curve
that changes from realization to realization and cannot be predicted
prior to RFEA. It is reasonable to postulate that this difference exists
because shear failure is localized along a narrow band while there is no
localization at the elastic stage of loading.
The proposed method works properly for a 2D or 3D footing pro-
blem with various geometries (footing size, depth to the rigid layer,
embedment depth), various boundary condition (smooth versus rigid
footing, rough versus smooth footing base), various spatial variability
pattern (mean, COV, anisotropy, auto-correlation function), various
J. Ching et al.
Fig. 16. Eeff/μ versus Ewg/μ relationships for the 3D footing problem.
Poisson ratio, etc. However, the proposed method may not work
properly for a footing that can rotate. This is because the mobilization
factors ΔUn are produced by deterministic FEA. The resulting mobili-
zation factors ΔUn are always symmetric with respect to the center line
of the footing. However, if the footing can rotate, the mobilization
factors may not be symmetric. Also, we have been considering only a
soil domain modeled by a single random field thus far. For cases with
multiple random fields, e.g., the soil domain consists of several hor-
izontal layers and each layer is modeled by a separate random field, the
proposed method may not work properly.
Acknowledgements
We would like to acknowledge the sponsorship from the Ministry of
Science and Technology of Republic of China, under 105-2221-E-002-
042-MY3 and 105-2811-E-002-058.
References
[1] Ahmed A, Soubra A-H. Probabilistic analysis of strip footings resting on a spatially
random soil using subset simulation approach. Georisk 2012;6(3):188–201.
[2] Ahmed A, Soubra A-H. Probabilistic analysis at the serviceability limit state of two
neighboring strip footings resting on a spatially random soil. Struct Saf
2014;49:2–9.
[3] Al-Bittar T, Soubra A-H. Probabilistic analysis of strip footings resting on spatially
varying soils and subjected to vertical or inclined loads. ASCE J Geotech Geoenviron
Eng 2014;140(4):04013043.
[4] Ching J, Phoon KK. Mobilized shear strength of spatially variable soils under simple
112
Structural Safety 73 (2018) 99–113
stress states. Struct Saf 2013;41:20–8.
[5] Ching J, Tong XW, Hu YG. Effective Young’s modulus for a spatially variable ele-
mentary soil mass subjected to a simple stress state. Georisk 2016;10(1):11–26.
[6] Ching J, Hu YG, Phoon KK. On characterizing spatially variable soil shear strength
using spatial average. Probabilistic Eng Mech 2016;45:31–43.
[7] Ching J, Phoon KK, Pan YK. On characterizing spatially variable soil Young’s
modulus using spatial average. Struct Saf 2017;66:106–17.
[8] Ching J, Phoon KK, Sung SP. Worse-case scale of fluctuation in basal heave analysis
involving spatially variable clays. Struct Saf 2017;68:28–42.
[9] Ching J, Sung SP. Spatial averaging of stationary random fields along curves – si-
mulation and variance reduction. J GeoEng 2016;11(1):33–43.
[10] Fenton GA, Griffiths DV. Probabilistic foundation settlement on spatially random
soil. ASCE J Geotech Geoenviron Eng 2002;128(5):381–90.
[11] Fenton GA, Griffiths DV. Three-dimensional probabilistic foundation settlement.
ASCE J Geotech Geoenviron Eng 2005;131(2):232–9.
[12] Golub GH, Heath M, Wahba G. Generlaized cross-validation as a method for
choosing a good ridge parameter. Technometrics 1979;21(2):215–23.
[13] Griffiths DV, Fenton GA. Probabilistic slope stability analysis by finite elements.
ASCE J Geotech Geoenviron Eng 2004;130(5):507–18.
[14] Griffiths DV, Fenton GA. Probabilistic settlement analysis by stochastic and random
finite-element methods. ASCE J Geotech Geoenviron Eng 2009;135(11):1629–37.
[15] Griffiths DV, Paiboon J, Huang J, Fenton GA. Homogenization of geomaterials
containing voids by random fields and finite elements. Int J. Solids Struct
2012;49:2006–14.
[16] Gross D, Seelig T. Micromechanics and homogenization. Fracture Mechanics,
Mechanical Engineering Series; 2011. p. 229–99.
[17] Houy L, Breysse D, Denis A. Influence of soil heterogeneity on load redistribution
and settlement of a hyperstatic three-support frame. Geotechnique
2005;55(2):163–70.
[18] Hu YG, Ching J. Impact of spatial variability in soil shear strength on active lateral
forces. Struct Saf 2015;52:121–31.
[19] Jha SK, Ching J. Simulating spatial averages of stationary random field using
Fourier series method. ASCE J Eng Mech 2013;139(5):594–605.
[20] Jimenez R, Sitar N. The importance of distribution types on finite element analyses
J. Ching et al.
of foundation settlement. Comput Geotech 2009;36:474–83.
[21] Nour A, Slimani A, Laouami N. Foundation settlement statistics via finite element
analysis. Comput Geotech 2002;29:641–72.
[22] Paiboon J, Griffiths DV, Huang J, Fenton GA. Numerical analysis of effective elastic
properties of geomaterials containing voids using 3D random fields and finite ele-
ments. Int J Solids Struct 2013;50:3233–41.
[23] Paice GM, Griffiths DV, Fenton GA. Finite element modeling of settlements on
spatially random soil. ASCE J. Geotech Eng 1996;122(9):777–9.
[24] Picard R, Cook D. Cross-validation of regression models. J Am Stat Assoc
113
Structural Safety 73 (2018) 99–113
1984;79(387):575–83.
[25] Rungbanaphana P, Honjo Y, Yoshida I. Settlement prediction by spatial-temporal
random process using Asaoka’s method. Georisk 2010;4(4):174–85.
[26] Tikhonov AN, Arsenin VY. Solution of Ill-posed Problems. Washington: Winston &
Sons; 1977.
[27] Vanmarcke EH. Probabilistic modeling of soil profiles. ASCE J Geotech Eng Div
1977;103(11):1227–46.
[28] Vanmarcke EH. Random Fields: Analysis and Synthesis. Cambridge, Mass: MIT
Press; 1984.