Seepage beneath water retaining structures founded

on spatially random soil

D.V. Griffiths* and Gordon A. Fenton
*

Division of Engineering, Colorado School of Mines, U.S.A.

Department of Engineering Mathematics, Dalhousie University, Canada

Abstract The effect of stochastic soil permeability on confined seepage occurring

beneath water retaining structures has been studied. Random field concepts for
the generation of soil permeability properties with a fixed mean, standard deviation and spatial correlation structure, have been combined with finite element
methods to perform Monte Carlo simulations of the seepage problem. Analyses
have been performed for the case of a dam with two cut-off walls. The results of
parametric studies to gauge the effect of the standard deviation and correlation
structure of the permeability on the output statistics relating to seepage quantities, exit gradients and uplift pressures are presented. In all cases, comparison
is made with results that would be achieved on a deterministic basis. Flow rates
and other quantities of interest are shown to be significantly affected by both the
standard deviation and the correlation structure of soil permeability.

Introduction

The great majority of geotechnical analysis is deterministic in that the soil properties used
are assumed to be average values. Variations in the soil properties are then accounted
for by the use of safety factors which are often rather arbitrarily applied to the computed
result.
This average approach to the definition of soil properties has tended to be applied,
not only to classical soil mechanics calculations, but also to numerical computations
using sophisticated numerical techniques such as the finite element method. Properties
are usually assigned on the basis of a limited number of laboratory tests. In reality,
these properties vary from point to point and can only be determined deterministically
through numerous field tests. Since this is both expensive and impractical, random field
models can be used to represent the geomaterial. The parameters of these models can
be estimated from a limited number of test results.
Mean soil properties are fairly well established, and recently there has been an improvement in the availability of data on second moment statistics (standard deviation
and spatial correlation). In large part, the data gathering has been motivated by the
availability of random field simulation algorithms and their potential for producing useful
results. The increased performance of computers has also enabled more detailed discretisation of boundary value problems, and better modelling of the stastistical properties of
the input parameters.

The finite element method is an ideal vehicle for modelling materials with a spatial
variation in properties. Stochastic finite element analysis has been implemented in a
number of areas of geotechnical interest. For example Beacher and Ingra (1981) and
Righetti and Harrop-Williams (1988) for stress analysis and settlements of foundations,
Ishii and Suzuki (1987) for slope stability and Smith and Freeze (1979a and 1979b) for
confined seepage.
Stochastic finite elements can be interpreted in different ways. On the one hand
statistical properties can be built into the finite element equations themselves (see e.g.
Vanmarcke and Grigoriu 1983), or multiple analyses (Monte Carlo) can be performed,
with each analysis stemming from a realisation of the soil properties treated as a multidimensional random field. In the present work the latter approach has been used to examine confined seepage, with particular reference to flow under a water retaining structure
founded on stochastic soil. While the Monte Carlo approach tends to be computationally intensive, it has the distinct advantage of being able to model highly variable input
properties.

Confined seepage

In the study of seepage through soils beneath water retaining structures, three important
quantities need to be assessed by the designers as shown in Figure 1): seepage quantity,
exit gradients and uplift forces.

Upstream
=10

Exit gradient
ie
Downstream

=0
1m

4m

No flow

No flow

3m

Uplift force U

No flow

No flow

No flow

No flow

6m

4m

4m

Figure 1. The boundary value problem

The classical approach used by Civil Engineers for estimating these quantities involves the use of carefully drawn flow nets (Casagrande 1940, Cedergren 1967, Verruijt
1970). Various alternatives to flow nets are available for solving the seepage problem,
however in order to perform quick parametric studies, for example relating to the effect

of cut-off wall length, powerful approximate techniques such as the Method of Fragments
(Pavlovsky 1933, Harr 1962, Griffiths 1984) are increasingly employed. The conventional methods are deterministic, in that the soil permeability is assumed to be constant
and homogeneous, although anisotropic properties and stratification can be taken into
account.
In this paper, a more rational approach to the modelling of soil properties is adopted,
whereby the permeability of the soil underlying a structure such as that shown in Figure 1
is assumed to be stochastic, i.e. the soil property in question is assumed to be a random
field (e.g. Vanmarcke 1984) defined statistically . The best known statistics are the mean
and standard deviation, however it is well known that spatial dependencies also exist
the soil properties at two points separated by 1 cm are likely to be more similar than
those at two points separated by a metre or a kilometre. This spatial dependence is often
characterized by a measure called the scale of fluctuation, which, loosely speaking, is
the distance over which properties show appreciable correlation. In general for a site of
fixed size, as the scale of fluctuation increases, the soil properties become more uniform
over the site. These statistics will be discussed further in a later section.
The analyses in this paper use a technique called Local Average Subdivision (LAS)
to generate realisations of the random permeability fields with given mean, standard
deviation and correlation structure. This technique is fully described by Fenton (1990)
and Fenton and Vanmarcke (1990).
The resulting field of permeabilities is mapped
onto a finite element mesh, and potential and stream function boundary conditions are
specified. The governing elliptic equation for steady flow (Laplace) leads to a system of
linear equilibrium equations which are solved for the nodal potential values throughout
the mesh using conventional Gaussian elimination.
Only deterministic boundary conditions are considered in this paper, the primary goal
being to investigate the effects of randomly varying soil properties on the engineering
quantities noted above. The method is nevertheless easily extended to random boundary
conditions corresponding to uncertainties in upstream and downstream water levels.
The next two sections give a brief description of the finite element technique and the
method by which permeability values are assigned to the mesh. This is followed by a
results section in which the statistics of the output quantities relating to flow rate, exit
gradients and uplift are discussed.

Finite Element Analyses

The steady flow problem is governed in 2-d by Laplaces equation, in which the dependent
variable is the piezometric head or potential at any point in the Cartesian field x-y.
kx

2
2
+
k
=0
y
x2
y 2

(1)

where kx and ky are the permeabilities in the x and ydirections. In the present work
and at the element level, the permeability field is assumed to be isotropic (kx = ky = k).
While the method discussed herein is simply extended to the anisotropic case (through
the generation of a pair of correlated random fields) it was felt that such an extension

is best restricted to a particular site of interest, the complexity introduced to a general

discussion being unwarranted.
Note that equation (3.1) is strictly only valid for constant k. In this analysis the
permeability is taken to be constant within each element, its value being given by the
local geometric average of the permeability field within the domain of the element. From
element to element, the value of k will vary, however, reflecting the random nature of
the permeability. This approximation of the permeability field is consistent with the
approximations made in the finite element method and is superior to most traditional
approaches in which the permeability of an element is taken to be simply the permeability
at some point within the element.
A typical finite element mesh used in this study is shown in Figure 2. It contains
1400 elements, and represents a model of 2-d flow beneath a dam which includes two
cut-off walls. The upstream and downstream potential values are fixed at 10 and zero
metres respectively. The cut-off walls are assumed to have zero thickness, and the nodes
along those walls have two potential values corresponding to the right and left sides of
the wall.
20 elements

30 elements

20 elements

20
elements

Figure 2. The FE mesh: all elements are 0.2m x 0.2m squares

The finite element code for the solution of Laplaces equation is broadly similar to
that published by Smith and Griffiths (2004) . The element conductivity matrices are
assembled into a global matrix in the usual way, resulting in a system of linear equations in
the unknown nodal potential values. The global conductivity relationship after assembly
becomes
K = Q

(2)

Once the global conductivity equations are solved leading to nodal potential values
held in , the output quantities relating to flow rates, uplift pressures and exit gradients
are easily deduced. More detail on how these values are obtained will be described in
later sections.
For each boundary value problem considered, multiple solutions were obtained using
successive realisations of the permeability field. The random permeability field is characterized by three parameters defining its first two moments, namely the mean k , the
standard deviation k and the scale of fluctuation k .

In order to obtain reasonably stable out statistics, it was decided that each parametric
combination would be analysed using 1000 realisations.

Generation of permeability values

Field measurements of permeability have indicated an approximately log-normal distribution (see e.g. Hoeksma and Kitanidis 1985, Sudicky 1986) . The same distribution
has therefore been adopted for the simulations presented in this Paper. Essentially, the
permeability field is obtained through the transformation
ki = exp{ln k + ln k gi }

(3)

in which ki is the permeability assigned to the ith element, gi is the local average of a
standard Gaussian random field, g, over the domain of the ith element, and ln k and ln k
are the mean and standard deviation of the logarithm of ki (obtained from the target
mean and standard deviation k and k ).
The LAS technique renders realisations of the local averages gi which are derived from
the random field g having zero mean, unit variance, and a spatial correlation controlled
by the scale of fluctuation. As the scale of fluctuation goes to infinity, gi becomes equal
to gj for all elements i and j that is the field of permeabilities tends to become uniform
on each realisation. At the other extreme, as the scale of fluctuation goes to zero, gi and
gj become independent for all i 6= j the soil permeability changes rapidly from point
to point.
In the two dimensional analyses presented in this paper, the scales of fluctuation in
the vertical and horizontal directions are taken to be equal (isotropic) for simplicity.
Although beyond the scope of this paper, it should be noted that for a layered soil
mass the horizontal scale of fluctuation is generally larger than the vertical scale due
to the natural stratification of many soil deposits. The 2-d model used herein implies
that the out-of-plane scale of fluctuation is infinite soil properties are constant in this
direction which is equivalent to specifying that the streamlines remain in the plane of
the analysis. This is clearly a deficiency of the present model, however it is believed that
useful information regarding the variability of flow quantities is still to be gained from
the 2-d model.

Deterministic Solution

With regard to the seepage problem shown in Figure 2, a deterministic analysis was
performed in which the permeability of all elements was assumed to be constant and
equal to 105 m/s. This value was chosen as it was to be the mean value of subsequent
stochastic analyses. Both the potential and the inverse streamline problems were solved,
leading to the flow net shown in Figure 3.
All output quantities were computed in non-dimensional form. In the case of the flow
rate, the global flow vector Q was computed by forming the product of the potentials
and the global conductivity matrix from equation (3.2). Assuming no sources or sinks
in the flow regime, the only non-zero values in Q correspond to those freedoms on the

Figure 3. Deterministic flow net: nf = 5, nd = 20

upstream and downstream boundaries. These values were summed to give the flow rate
defined by
Q in m3 /s/m, leading to a non-dimensional flow rate Q
= Q/(k H)
Q

(4)

where k is the (isotropic) mean permeability and H is the total head difference between
the up- and downstream sides.
The uplift force on the base of the dam U was computed by integrating the pressure
distribution along the base of the dam between the cut-off walls. This quantity was
easily deduced from the potential values at the nodes along this line together with a
simple numerical integration scheme (Repeated Trapezium Rule). A non-dimensional
was defined:
uplift force U
= U/(Hw L)
U

(5)

where w is the unit weight of water and L is the distance between the cut-off walls. U
is the uplift force expressed as a proportion of buoyancy force that would occur if the
dam was submerged in water alone.
The exit gradient ie is the rate of change of head at the exit point closest to the
dam at the downstream end. This was calculated using a 4-point backward difference
numerical differentiation formula of the form:
1
(110 181 + 92 23 )
(6)
6b
where the i values correspond to the piezometric head at the four nodes vertically below
the exit point as shown in Figure 4, and b is the constant vertical distance between nodes.
It may be noted that the downstream potential head is fixed equal to zero, thus 0 = 0.0
m. The use of this four-point formula was arbitrary, and was considered a compromise
between the use of very low order formulae which would be too sensitive to random
fluctuations in the potential, and high order formulae which would involve the use of
correspondingly high order interpolation polynomials which would be hard to justify
physically.
Referring to Figures 1 and 2, the constants described above were given the following
values: H = 10m, k = 105 m/s, w = 9.81kN/m3 , L = 6m, b = 0.2m.
ie

Dam

Downstream

o
1

2
Cut-off wall

Figure 4. Detail of downstream cut-off wall

and a deterministic analysis using the mesh of Figure 2 led to the following output
= 0.226, U
= 0.671, ie = 0.688.
quantities: Q
This value of ie would be considered unacceptable in a real design situation, bearing
in mind that the critical hydraulic gradient for most soils approximately equals unity.
The value of ie is proportional to the head difference H however, which in this case for
simplicity and convenience of normalisation has been set to 10 m as shown above.
These results will be compared with output from the stochastic analyses described in
the next section.

Stochastic Analyses

In all the 2-d stochastic analyses that follow, the soil was assumed to be isotropic with
a mean permeability of k = 105 m/s. More specifically, the random fields were generated such that the target (geometric) mean permeability of each finite element was
held constant at 105 m/s. Parametric studies were performed relating to the effect of
varying the standard deviation (k ) and the scale of fluctuation (k ) of the permeabil U
and ie
ity field. Following 1000 realisations, statistics relating to output quantities Q,
were calculated.

6.1

Single realisation

Before discussing the results from multiple realisations, an example of what a flow
net might look like for a single realisation is given in Figure 5 for permeability statistics
k /k = 1, and k = 1.0m.
In Figure 5, the flow net is superimposed on a grey-scale which indicates the spatial
distribution of the permeability values. Dark areas correspond to low permeability and
light areas to high permeability. The streamlines clearly try to avoid the low permeability zones, but this is not always possible as some realisations may generate a complete
blockage of low permeability material in certain parts of the flow regime. This type of
blockage is most likely to occur where the flow route is compressed, such as under a
cut-off wall. Flow in these (dark) low permeability zones is characterised by the streamlines moving further apart and the equipotentials moving closer together. Conversely,
flow in the (light) high permeability zones is characterised by the equipotentials moving
further apart and the streamlines moving closer together.

Figure 5. Stochastic flow net for a typical realisation

Although local variations in the permeability have an obvious effect on the local paths
taken by the water as it flows downstream, globally the stochastic and deterministic flow
nets exhibit many similarities. The flow is predominantly in a downstream direction, with
the fluid flowing down, under and around the cut-off walls. For this reason the statistics
of the output quantities might be expected to be rather insensitive to the geometry of
the problem (e.g. length of walls etc.), and qualitatively similar to the properties of a
1-d flow problem.
6.2

Statistics of the potential field

Figures 6 gives contours of the mean and standard deviation of the potential field
following 1000 realisations for the case where k = 1.0m. Here we have follow an approach
used by Smith and Freeze (1979a, 1979b) who presented the results of a series of numerical
experiments on both 1-d and 2-d confined flow problems.
The mean potential values given in Figures 6a are very similar to those obtained in the
deterministic analysis summarised in the flow net of Figure 3. The standard deviation
of the potentials given in Figures 6b indicate the zones in which the greatest uncertainty

(a)
0.16

0.16

0.32
0.48
0.32

0.48

0.64

0.80

0.80

0.64

(b)
Figure 6. (a) Contours of mean potential values (contour interval=0.5m), (b) Contours
of standard deviation of potential values (contour interval=0.16m), k /k = 1, k = 1m

exists regarding the potential values. It should be recalled that the up- and down- stream
(boundary) potentials are deterministic, so the standard deviation of the potentials on
these boundaries equals zero. The greatest values of standard deviation occur in the
middle of the flow regime, which in the case considered here represents the zone beneath
the dam and between the cut-off walls. The standard deviation is virtually constant in
this zone. The statistics of the potential field shown in Figure 6 are closely related to
the statistics of the uplift force as will be considered in the next section.
6.3

Parameteric studies

The parametric studies based on the mesh of Figure 2 were designed to show the
effect of the permeabilitys standard deviation, k , and scale of fluctuation, k , on the
U
and ie . In all cases the mean permeability, k , was maintained
output quantities Q,
constant at 105 m/s.
Instead of plotting k directly, the dimensionless coefficient of variation of permeability was used, and the following values were considered: k /k = 0.125, 0.25, 0.50, 1, 2,
4, 8, 16.0 together with scales of fluctuation values given by k = 0, 1, 2, 4, 8, and m
All permutations of these values were analysed, and the results summarised in Figures
7,8 and 9 in the form of the logarithm (base 10) of k /k vs. the means and standard
U
,and ie , denoted (Q , Q ), (U , U ), and (i , i ) respectively.
deviations of Q,
e
e

deterministic 0.226
and k = 0, k =

(a)

Q
k =8m

} k <8m

log ( k / k )
(b)

Q
k =8m

k =

k =4m
k =2m
k =1m

log ( k / k )
Figure 7. Coefficient of variation of permeability plotted against, (a) mean flow rate;
(b) standard deviation of flow rate

10

Flow rate Figure 7a shows a significant fall in Q as k /k increases for k < 8 m.

approaches the
As the scale of fluctuation approaches infinity, the expected value of Q
constant 0.226. This curve is also shown in Figure 7a, although it should be noted it
has been obtained through theory rather than simulation. In agreement with this result,
the curve k = 8 m shows a less marked reduction in Q with increasing coefficient of
variation k /k . However, over typical scales of fluctuation, the effect on average flow
rate is slight. The decrease in flow rate as a function of the variability of the soil mass is
an important observation from the point of view of design. Traditional design practice
may very well be relying on this variability to reduce flow rates on average. It also implies
that ensuring higher uniformity in the substrate may be unwarranted unless the mean
permeability is known to be substantially reduced and/or the permeability throughout
the site is carefully measured.
Figure 7b shows the behaviour of Q as a function of k /k . Of particular note is
that Q reaches a maximum corresponding to k /k in the range 1.0 - 2.0 for k 8
m. Again the theoretical result corresponding to k = has been plotted on the figure,
showing a continuous increase with k /k .
occurs under rather typiIn general, it appears that the greatest variability in Q
cal conditions: scales of fluctuation between 1 and 4 m and coefficient of variation of
permeability of around 1 or 2.
Uplift forces Figures 8a and 8b show the relationship between uplift force parameters
U and U and input permeability parameters k /k and k . From Figure 8a, U is
relatively insensitive to the parametric changes. There is a gradual fall in U as both
k /k and k increase. The greatest reduction being about 10% of the deterministic
value of 0.671 when k /k = 16.0 and k = 8.0 m. The insensitivity of the uplift force
to the permeability input statistics might have been predicted from Figures 3, and 6a
in which the contours of (mean) piezometric head are virtually the same in both the
deterministic and stochastic analyses.
Figure 8b shows that U consistently rises as both k /k and k increase. It is known
that in the limit as k , U 0 since under those conditions, the permeability field
becomes completely uniform. Some hint of this increase followed by a decrease is seen
from Figure 8b in that the largest increases are for k = 0 to k = 1 while the increase
from k = 4 to k = 8 is much smaller.
The actual value of U for a given set of k /k and k could easily be deduced from
the standard deviation of the potential values. Figure 6b gave contours of the standard
deviation of the potential values throughout the flow domain for the particular value
k /k = 1.0 and k = 1.0m. In Figure 6b, the potential standard deviation beneath the
dam was approximately constant and equal to 0.8 m. After non- dimensionalisation by
dividing by H = 10 m, these values closely agree with the corresponding values in the
graphs of Figure 8b.
The magnitude of the standard deviation of the uplift force given in Figure 8b across
the range of parameters considered was not very great. The implication is that this
quantity can be estimated with a reasonable degree of confidence. The explanation lies
in the fact that the uplift force is calculated using potential values over quite a large
number of nodes beneath the dam. This averaging process would tend to damp out

11

 k =1m

deterministic 0.671
and k = 0, k =

k =2m
k =4m

k =8m

(a)

log ( k / k )

k =8m

k =4m
k =2m
k =1m

(b)
log ( k / k )
Figure 8. Coefficient of variation of permeability plotted against, (a) mean uplift force;
(b) standard deviation of uplift force

12

fluctuations in the potential values that would be observed on a local scale, resulting in
a variance reduction.

Exit gradients This quantity is based on the first derivative of piezometric head or
potential with respect to distance at the exit point closest to the downstream end of the
dam. It is well known that in a deterministic approach, the largest value of ie , and hence
the most critical, lies at the exit point of the uppermost (and shortest) streamline. While
for a single realisation of a stochastic analysis this may not be the case, on average the
location of the critical exit gradient is expected to occur at the deterministic location.
As ie is based on a first derivative at a particular location within the mesh (see
Figure 4), it can be expected to be the most susceptible to local variations generated
by the stochastic approach. In order to average the calculation of ie over a few nodes,
it was decided to use a 4-point (backward) finite difference scheme as given previously
in equation (5.3). This is equivalent to fitting a cubic polynomial over the potential
values calculated at the four nodes closest to the exit point adjacent to the downstream
cut-off wall. The cubic is then differentiated at the required point to estimate ie . Note
then that the gradient is estimated by studying the fluctuations over a length of 0.6 m
vertically (the elements are 0.2 m by 0.2 m in size). This length will be referred to as
the differentiation length in the following.
The variation of ie and ie , over the range of parameters considered are given in
Figures 9a and 9b. The sensitivity of ie to k /k is clearly demonstrated. In Figure 9a,
ie agrees quite closely with the deterministic value of 0.688 for values of k /k in the
range 0.0 to 1.0, but larger values start to show significant instability and divergence. It
is interesting to note that for k 1, the tendency is for ie to fall below the deterministic
value of ie as k /k is increased, whereas for larger values of k it tends to increase above
the deterministic value. The scales 0 and 1 are less than and of the same magnitude as the
differentiation length of 0.6 m used to estimate the exit gradient, respectively, while the
scales 2, 4, and 8 are substantially greater. If this has some bearing on the divergence
phenomena seen in Figure 9a it calls into some question the use of a differentiation
length to estimate the derivative at a point. Suffice to say that there may be some
conflict between the numerical estimation method and random field theory regarding the
exit gradient that needs further investigation.
Figure 9b indicates the relatively large values of ie , which grow rapidly as k /k is
increased. The influence of k in this case is not so great, with the results corresponding
to k = 1.0, 2.0, 4.0 and 8.0 m being quite closely grouped. It is noted that theoretically
as k and k 0, ie 0.688 and ie 0. Their appears to be some evidence of
a reduction in ie as k increases, which is in agreement with the theoretical result. For
scales of fluctuation negligible relative to the differentiation length, that is k = 0, the
variability in ie is much higher than that for other scales at all but the highest permeability variance. This is perhaps to be expected, since k = 0 yields large fluctuations in
permeability within the differentiation length.

13

 k =8m

ie

k =4m
k =2m
deterministic 0.68
and k = 0, k =

k =1m

(a)
log ( k / k )

ie

k =4,8m

k =2m
k =1m

(b)
log ( k / k )
Figure 9. Coefficient of variation of permeability plotted against, (a) mean exit gradient;
(b) standard deviation of exit gradient

14

Concluding Remarks

A range of parametric studies have been performed relating to flow beneath a water
retaining structure with two cut-off walls founded on a stochastic soil. Random field
concepts were used to generate permeability fields having predefined mean, standard
deviation and correlation structure. These values were mapped onto a finite element
mesh consisting of 1400 elements, and, for each set of parameters, 1000 realisations of
the boundary value problem were analysed. In all cases, the target mean permeability
of each finite element was held constant and parametric studies were performed over a
range of values of coefficient of variation and scale of fluctuation.
The three output quantities under scrutiny were the flow rate, the uplift force and
the exit gradient; the first two of these being non-dimensionalised for convenience of
presentation.
The mean flow rate was found to be relatively insensitive to typical scales of fluctuation, but fell consistently as the variance of of the permeability was increased. This
observation may be of some importance in the design of such water retaining structures.
The standard deviation of the flow rate consistently increased with the scale of fluctuation, but rose and then fell again as the coefficient of variation was increased. These
maxima are currently the subject of further investigations by the Authors.
The mean uplift force was rather insensitive to the parametric variations, falling by
only about 10% in the worst case. The relatively small variability of uplift force was due
to a damping out of local variations inherent in the random field by the averaging of
potential values over the nodes along the full length of the base of the dam. Nevertheless, the standard deviation of the uplift force rose consistently with increasing scale of
fluctuation and coefficient of variation, as was to be expected from the contour plots of
the standard deviation of the potential values across the flow domain.
The mean exit gradient was much more sensitive to the statistics of the input field.
Being based on a first derivative of piezometric head with respect to length at the exit
point, this quantity is highly sensitive to local variations inherent in the potential values
generated by the random field. Some local averaging was introduced by the use of fourpoint numerical differentiation formula, however the fluctuation in mean values was still
considerable and the standard deviation values high.

Acknowledgement

The work described in this paper was supported in part by a travel grant (ref. 07-04-032) from the Association of Universities and Colleges of Canada.

15

Notation

b
g
gi
H
ie
k
ki
kx , ky
L
nf
nd

Q, Q

U, U
x, y
w
k
k , k
ln k , ln k
ie , ie
Q , Q
U , U
k /k

0 , 1 , 2 , 3

vertical spacing between nodes

standard Gaussian random field
local average of g over the ith element
head difference between up- and downstream levels
exit hydraulic gradient
permeability
permeability assigned to the ith element
permeability in x- and y-directions
width of dam
number of flow channels
number of equipotential drops
flow rate and non-dimensional flow rate
uplift force and non-dimensional uplift force
Cartesian coordinates
unit weight of water
scale of fluctuation of permeability
mean and standard deviation of permeability
mean and standard deviation of ln ki
mean and standard deviation of exit gradient
mean and standard deviation of dimensionless flow rate
mean and standard deviation of dimensionless uplift force
coefficient of variation of permeability
potential or piezometric head
potential values near exit point

K
Q

global conductivity matrix

global flow vector
global potential vector

Bibliography
G.B. Baecher and T.S. Ingra. Stochastic FEM in settlement predictions. J Geotech Eng,
ASCE, 107(GT4):449463, 1981.
A. Casagrande. Seepage through dams. Journal of the New England Water Works
Association, L1(2), 1937.
H.R. Cedergren. Seepage, drainage and flow nets. John Wiley and Sons, Chichester, New
York, 1967.
G.A. Fenton. Simulation and analysis of random fields. PhD thesis, Department of Civil
Engineering and Operations Research, Princeton University, 1990.

16

G.A. Fenton and E.H. Vanmarcke. Simulation of random fields via local average subdivision. J Eng Mech, ASCE, 116(8):17331749, 1990.
D. V. Griffiths. Rationalised charts for the method of fragments applied to confined
seepage. Geotechnique, 34(2):229238, 1984.
M.E. Harr. Groundwater and seepage. McGraw Hill, London, New York, 1962.
R.J. Hoeksema and P.K. Kitanidis. Analysis of the spatial structure of properties of
selected aquifers. Water Resour Res, 21(4):563572, 1985.
K. Ishii and M. Suzuki. Stochastic finite element method for slope stability analysis.
Structural Safety, 4:111129, 1987.
N.N. Pavlovsky. Motion of water under dams. In Proceedings of the 1st Congress on
Large Dams, Stockholm, pages 179192. 1933.
G. Righetti and K. Harrop-Williams. Finite element analysis of random soil media. J
Geotech Eng, ASCE, 114(GT1):5975, 1981.
I. M. Smith and D. V. Griffiths. Programming the Finite Element Method. John Wiley
and Sons, Chichester, New York, 4th edition, 2004.
L. Smith and R.A. Freeze. Stochastic analysis of steady state groundwater flow in a
bounded domain, 1. One-dimensional simulations. Water Resour Res, 15(3):521528,
1979a.
L. Smith and R.A. Freeze. Stochastic analysis of steady state groundwater flow in a
bounded domain, 2. Two-dimensional simulations. Water Resour Res, 15(6):1543
1559, 1979b.
E.A. Sudicky. A natural gradient experiment on solute transport in a sand aquifer:
Spatial variability of hydraulic conductivity and its role in the dispersion process.
Water Resour Res, 22(13):20692083, 1986.
E.H. Vanmarcke. Random fields: Analysis and synthesis. The MIT Press, Cambridge,
Mass., 1984.
E.H. Vanmarcke and M. Grigoriu. Stochastic finite element analysis of simple beams. J
Eng Mech, ASCE, 109(5):12031214, 1983.
A. Verruijt. Theory of groundwater flow. Macmillan and Co. Ltd, London, 1970.

17