You are on page 1of 6

Computers and Geotechnics 52 (2013) 16

Contents lists available at SciVerse ScienceDirect

Computers and Geotechnics


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

Technical Communication

Treatment of transitional element with the Monte Carlo method for


FEM-based seepage analysis
Yu-xin Jie a, Xu-dong Fu a, Gang Deng b,
a
b

State Key Laboratory of Hydroscience and Engineering, Tsinghua University, Beijing 100084, China
State Key Laboratory of Simulation and Regulation of Water Cycle in River Basin, China Institute of Water Resources and Hydropower Research, Beijing 100048, China

a r t i c l e

i n f o

Article history:
Received 2 January 2013
Received in revised form 15 February 2013
Accepted 21 February 2013
Available online 30 March 2013
Keywords:
Monte Carlo method
Numerical integration
Seepage
Finite element method (FEM)
Free surface

a b s t r a c t
In this paper, The Monte Carlo method is incorporated into the nite element method (FEM) to conduct
seepage analysis with a free surface. For the transitional element cut by the free surface, it is used to calculate the composite permeability coefcient, as well as to perform the integration directly using Monte
Carlo integration. This new algorithm requires less iteration procedures for convergence. The convergence of the method is also proved for cases where there is a signicant difference between the permeability coefcient above and below the free surface.
2013 Elsevier Ltd. All rights reserved.

1. Introduction
One of the major difculties in FEM-based seepage analysis of
levees and dams is that the free surface is not known a priori. In
general, two approaches have been developed to deal with such
problems. One assumes a free surface initially and adjusts it iteratively, resulting in a moving-mesh procedure that continues until
the boundary conditions are sufciently satised on the free surface [1]. The other is a xed mesh method in which the nite element meshes are xed during the iterations. For example, Desai
presented residual ow schemes [2,3], while Bathe and Khoshgoftaar [4] proposed a model considering meshes on the free surface
to be composed of composite materials. Saturated and unsaturated
theory can also be used to solve this problem with a xed mesh, in
which the zero pressure contour serves as the free surface [5,6].
There are many other techniques that are also used to solve these
problems, such as the variational inequality method, the numerical
manifold method, the boundary-tted coordinate transformation
method, the nite-volume method, and the meshless method [7
11].
The xed mesh nite element method is more convenient when
the seepage problem is coupled with stressstrain analysis, e.g., in
consolidation analysis with a free surface. In the method proposed
by Bathe and Khoshgoftaar [4], the composite permeability coefcient of the transitional element cut by the free surface is deter Corresponding author. Tel.: +86 10 68785126.
E-mail address: dgang@iwhr.com (G. Deng).
0266-352X/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.compgeo.2013.02.005

mined by the ratio between the parts above and below the free
surface. For solving the problem with saturated and unsaturated
theory, there is also a marked difference in the permeability between the upper and lower parts of the transitional element. In
fact, Bathe and Khoshgoftaars method can be considered as a special case of saturated and unsaturated theory. The method for estimating the composite permeability correctly is crucial in the
numerical analysis. If the permeability coefcient is not suitably
calculated, iterations may fail leading to a wrong result being
obtained.
In this paper, we deal with the problem through integrating the
Monte Carlo method with nite element analysis. This proposed
method can be used to calculate the composite permeability coefcient of the transitional element cut by the free surface, or to do
the integration directly by Monte Carlo integration. A program
written by the authors was used to analyze several seepage problems with a free surface, and the computational examples show
that this method improves the convergence rate greatly and simplies the program.

2. The Monte Carlo method


The Monte Carlo method approximates solutions to mathematical or physical problems by using statistical sampling theory. It is
based on the law of large numbers in probability theory [12,13],
and is promising in many aspects and simple to implement.

Y.-x. Jie et al. / Computers and Geotechnics 52 (2013) 16

The Monte Carlo method provides a simple approach to conducting numerical integration called Monte Carlo integration. For
the integration

f xdx;

the value of I can be considered to be the product of b  a and the


average value of f x over a; b. The average value of f x can be obtained by generating a random sample x1 , . . ., xn that is uniformly
distributed over a; b. One then obtains:

^I b  a f x1    f xn :
n

If n ! 1, ^I will converge to I with probability 1. The error in the


integration is proportional to the standard error of ^I and is inverp
sely proportional to n, so we can achieve arbitrary accuracy with
sufciently many sample points.
The Monte Carlo integration needs a lot of sampling data, which
lowers its computational efciency in most cases in contrast to the
Gaussian integration. With the development of computer hardware, restrictions on the running speed have become less and less
important. The Monte Carlo method has become more and more
attractive due to its simplicity and adaptability, and the fact that
the probability and rate of convergence are independent of the
dimension of the problem. It has been used to calculate the area
or volume of a geometric object as well as the location of the centroid of a region. For example, we found that the Du et al.s method
for determining Centroidal Voronoi tessellations (CVTs) can be
considered as a probabilistic version of Lloyds method [14,15],
the essence of which is actually computing the centroids of Voronoi regions by the Monte Carlo integration.
In the eld of geomechanics, the Monte Carlo method has
mainly been used to solve problems involving spatial and temporal
variability of rock and soil parameters [1623], but has not yet
been directly used in seepage analysis with a free surface.
3. Saturated steady seepage
3.1. Governing equations
The differential equation governing steady-state saturated
water ow is







@
@h
@
@h
@
@h
kx

ky

kz
0;
@x
@x
@y
@y
@z
@z

where kx, ky, kz are the permeability coefcients in the x, y, and z


directions, h = z + p/cw is the water head, p is the pore water pressure, and cw is the unit weight of water. Note that here we assume
the coordinate system (x, y, z) corresponds to the principal directions of the hydraulic conductivity tensor.
The nite element equation is

Kfhg fQg;

where K is the element characteristic matrix, fhg is the vector of


nodal heads, and fQ g is the element ux vector.
3.2. Treatment of transitional elements
In the adjusting permeability method proposed by Bathe and
Khoshgoftaar [4], the permeability coefcient of the transitional
element is adjusted to approximate the real seepage eld. For the
portion of the element below the free surface, the soil is supposed
to be saturated with a permeability coefcient of k, while for the
portion of the element above the free surface the permeability
coefcient is gk, generally with g = 1/1000. That is,

g 1; p P 0;
g 1=1000; p < 0;

where p is the pore water pressure.


However, there are often difculties with the convergence of
this method, as the abrupt jump in the permeability coefcient
leads to oscillation of the solution during iterations. There are
two general approaches to solving this problem. One is to replace
the step function for g with a continuous function (see Fig. 1), to
control the change in the permeability coefcient during the iterations. The other is to use more Gaussian integration points to capture the change in permeability in the transitional element cut by
the free surface. For the transitional elements in Fig. 2, if all of the
integration points are above the free surface (see Fig. 2a), the element is considered to be an unsaturated material and the inuence
of the lower part of the element is totally omitted. Conversely, if all
of the integration points are below the free surface (see Fig. 2b), the
element is considered to be a saturated material and the inuence
of the upper part of the element is omitted. In general, this omission may be remedied to some extend by adding more Gaussian
points. However, the effect of adding Gaussian points is limited.
Since the number of Gaussian points is unlikely to be extremely
large (e.g., up to 10,000 points), thus the situations indicated in
Fig. 2 will always occur.
An possible alternative method is to determine the composite
permeability coefcient of the transitional element according to
the volume ratio between the two parts above and below the free
surface, whereas it is troublesome to calculate the volume of the
irregular geometric parts as well as the pattern with which the
transitional element is cut by the free surface [24,25]. Therefore,
the Monte Carlo method seems promising in these cases.
The Monte Carlo method is feasible for the following two
approaches:
(1) The Monte Carlo method is used to perform numerical integration on the transitional elements, while Gaussian integration is used for the other elements.
(2) Gaussian integration is used on all the elements. The permeability coefcient of the elements below and above the free
surface are specied to be k and gk, respectively, while for
the transitional elements, the composite permeability coefcient is determined by the Monte Carlo method. For example, as shown in Fig. 3, we generate n random points in the
element. Assume that the number of points below the free
surface is n1, and that the number above the free surface is
n2 (n = n1 + n2). Then the composite permeability coefcient
of the transitional element is thought to be

 n1 k n2 k ;
k
1
2
n
n

where k1 = k and k2 = gk.


Obviously, n1 =n and n2 =n are approximations of the volume ratios V1/V and V2/V, respectively, where V1 is the volume below the
free surface, V2 is the volume above the free surface, and V = V1 + V2
 in Eq. (6) is
is the total volume of the transitional element. Thus k
actually determined by the volume ratio between the parts above
and below the free surface.
The advantage of the Monte Carlo method is that it avoids the
problems of calculating the volume of the transitional element as
well as how the element is cut by the free surface. The location
and the related permeability coefcient of the sample points can
be simply determined from the pore water pressure p. For example, if p > 0, the point is below the free surface, and if p < 0, the
point is above the free surface. Additionally, in comparison with
Gaussian points, the number of points can be very large by the
Monte Carlo method, e.g., 10,000, even 100,000 or 1,000,000 for

Y.-x. Jie et al. / Computers and Geotechnics 52 (2013) 16

k
k

k
p

(a) The step function

(b) The continuous function

Fig. 1. Permeability coefcient vs. pore water pressure.

Free surface

Free surface

(b) Integration points below the


free surface

(a) Integration points above the


free surface

Fig. 2. Free surface and the integration points.

V1
V2

Free surface

Fig. 3. Transitional element cut by the free surface.

an element, thus the situations indicated in Fig. 2 can be largely


avoided.
3.3. Examples
Fig. 4 shows a nite element model for a rectangular dam. The
model consists of 4-node linear elements. The width and the height
are 4 m and 6 m, respectively. The upstream (left) water level is
6 m, and the downstream (right) water level is 1 m. The permeability coefcient is 1  103 cm/s, while g in Eq. (5) is assigned to be
1/1000. The error tolerance is 0.1%.
The following three schemes are used and compared in the
analysis:
Scheme 1: Gaussian integration is used on all the elements, and
the number of Gaussian integration points in each element is
2  2.
Scheme 2: Monte Carlo integration is employed on the transitional elements, while Gaussian integration is still used on
other elements.
Scheme 3: Gaussian integration is used on all the elements as in
Scheme 1. For transitional element, the composite permeability
coefcient is obtained by the Monte Carlo method according to
Eq. (6).
The computational procedure is as follows:
1. Assume an initial release point on the downstream boundary.
2. Compute the water head fhg1 according to Eq. (4), in which the
permeability coefcient of each element is set to be k.

Fig. 4. Finite element mesh for a rectangular dam (in m).

3. Determine the relationship between each element and the free


surface for fhg1 : an element is below the free surface if all the
nodal water pressures are bigger than 0, and is above the free
surface if all the nodal water pressures are smaller than 0. An
element is a transitional element cut by the free surface if some
of the nodal water pressures are bigger than 0 and the others
are smaller than 0.
4. Compute the water head (fhg2 ) again, where the permeability
coefcient of elements below the free surface and those above
the free surface are set to be k and gk, respectively, and the
transitional elements are dealt with according to the three
schemes explained above.
5. If the difference between fhg2 and fhg1 is bigger than the error
tolerance, set fhg1
fhg1 fhg2 =2 to determine the permeability coefcient of each element.
6. Repeat steps 25 until the error tolerance is met.
7. Check the boundary conditions on the downstream boundary. If
the boundary conditions are not fullled, modify the release
point.
8. Repeat steps 27 until the boundary conditions are fullled.
In our analysis, the number of random points for the Monte Carlo method was set to be 10,000.
We found that convergence is usually achieved in 250 iterations, and generally less than 100 iterations, by Schemes 2 and 3,
and that the free surface obtained is quite smooth (see Fig. 5a).
On the other hand, convergence failed even after 1000 iterations

Y.-x. Jie et al. / Computers and Geotechnics 52 (2013) 16

(a) Schemes 2 and 3

(b) Scheme 1

Fig. 5. Calculated free surface (in m).

Fig. 8. 3-D nite element mesh for a rectangular dam.

Fig. 6. Finite element mesh for a trapezoidal dam (in m).

Fig. 7. Calculated free surface by Schemes 2 and 3 (in m).

by Scheme 1. Fig. 5b also shows the free surface obtained by


Scheme 1. The curve is obviously rough, although the location of
the free surface and the release point are not incorrect. However,
the non-convergent results are theoretically unreliable and should
be rejected.
Fig. 6 is a nite element model for a trapezoidal dam. The upstream (left) water level is 18 m, and the downstream (right) water
level is 0 m. Again, convergence is achieved in less than 100 iterations by Schemes 2 and 3, and fails by Scheme 1. The free surface
obtained by Schemes 2 and 3 is again quite smooth (see Fig. 7).
Further calculations show that for the two examples in Figs. 4
and 6, convergence is achieved within 200 iterations by Schemes
2 and 3 for g = 104, while convergence is achieved within 200 iterations by Scheme 3 and within 500 iterations by Scheme 2 for
g = 105. Even in the extreme case of g = 1010, convergence can
still be achieved within 200 iterations by Scheme 3 although convergence difculties emerges for Scheme 2.
The Monte Carlo method is also feasible for analyzing seepage
problems in three dimensions. Fig. 8 presents a 3-D nite element
model for a rectangular dam, where the width and height in the xzplane are 4 m and 6 m, as in Fig. 4, and the thickness in the y direction is 2 m. The upstream (left) water level is 6 m, and the downstream (right) water level is 1 m, again as in Fig. 4. Convergence
fails after 1000 iterations by Scheme 1, while Scheme 3 continues

Fig. 9. Calculated 3-D free surface by Scheme 3.

to show good convergence characteristics. The free surface obtained by Scheme 3 is shown in Fig. 9.
4. Unsaturated unsteady seepage
For unsaturated seepage, the zero pressure contour serves as
the free surface. The permeability coefcient is a function of pore
water pressure (see Fig. 10). In fact, the curve of k vs. p in Fig. 1
can be considered to be a special case for unsaturated soil.
The differential equation governing unsaturated unsteady
water ow is







@
@h
@
@h
@
@h
@nSr

kx
ky
kz
;
@x
@x
@y
@y
@z
@z
@t

where n is the porosity of the soil, and Sr is the degree of saturation.


As does in Eq. (3), here we also assume the coordinate system

Y.-x. Jie et al. / Computers and Geotechnics 52 (2013) 16

p
Fig. 10. Permeability coefcient vs. pore water pressure for unsaturated soil.

The initial water level on the left side was 0.1 m, and it reached
0.3 m instantaneously during the test. The water level on the right
side was kept at 0.1 m throughout. The computed results are presented in Fig. 12, in which the curves are the free surfaces calculated at t = 30 s, 60 s, 120 s, 240 s, 600 s and 4800 s after the
water level reaches 0.3 m. Comparing the test results indicates that
the Monte Carlo method performs well for analyzing unsaturated
and unsteady seepage. Not surprisingly, a similar application is
to conduct consolidation analysis of unsaturated soil. In fact, the
numerical simulation of water-lling in an earth-rock dam reported in Ref. [27] has actually employed the Monte Carlo method
to deal with the transitional element.
5. Conclusions
The Monte Carlo method is used in this paper to improve the
convergence of seepage analysis. It is effective for regions where
permeability varies greatly on some transitional elements that
are cut by the free surface.

Fig. 11. Finite element mesh for the sand box model test (in m).

Table 1
Degree of saturation Sr and relative permeability coefcient kr of the sand.
Pore water pressure, p
(kPa)

Degree of
saturation, Sr

Relative permeability
coefcient, kr

0
0.325
0.65
0.91
1.0
10.0

1
0.82
0.64
0.50
0.45
0.18

1
0.97
0.94
0.20
0.13
0.01

Tested
Calculated
4800s
600s
240s
120s
30s

60s

Fig. 12. Calculated free surfaces (in cm).

(x, y, z) corresponds to the principal directions of the hydraulic conductivity tensor.


In numerical analysis of unsaturated soil, the permeability coefcient is also adjusted according to the calculated pore water pressure, not only for the transitional element but also for an element
with negative pore water pressure. Consequently, a convergence
problem occurs when there is a big change in the permeability
coefcient during iterations. Again, the Monte Carlo method can
be helpful in these cases, and we can use it to compute the composite permeability coefcient or to perform integration on the
element directly.
Fig. 11 shows the nite element mesh for the sand box model
test done by Akai et al. [26]. The model was 3.15 m long, 0.23 m
wide and 0.33 m high. Porosity of the uniform sand was 0.44,
and the saturated permeability coefcient was 3.3  103 m/s.
The variation in degree of saturation Sr and the relative permeability coefcient kr vs. pore water pressure p are shown in Table 1.

(1) The Monte Carlo method is a simple but effective method. It


can be used to calculate the area or volume of an irregular
region, and be used as an alternative to Gaussian integration.
The Monte Carlo method can sample a huge number of
points, which makes it able to capture the information
involved in an element in detail. The property of what the
Monte Carlo method has facilities its application to regions
where exists great gradient in permeability. The examination has proved its applicability to seepage problems with
free surfaces and those unsaturated and unsteady ows
where permeability varies greatly on some elements.
(2) A sufcient number of random points is necessary, preferably at least 10,000. Though a large number of points require
more time at each step, the decreased number of iterations
actually saves time. It is thus more efcient as a whole in
many cases.
(3) The Monte Carlo method may also be useful for other applications such as consolidation analysis and stressstrain
analysis when the material properties vary greatly on some
elements. Further development and application of this
method looks promising.
(4) Though there are only linear elements employed in this
study, the proposed method may still be suitable for highorder elements. The theorem of Monte Carlo integration
and the Monte Carlo based volume calculation is independent of the element type. By using the corresponding interpolation method to determine the water head (or pore water
pressure) in the element, the Monte Carlo method can be
directly employed in the numerical analysis with high-order
elements.

Acknowledgments
The supports of Natural Science Foundation of China
(51039003), National Basic Research Program of China (973 Program 2013CB036402) and the State Key Laboratory of Hydroscience and Engineering (2013-KY-4) are gratefully acknowledged.
References
[1] Desai CS. Finite element methods for ow in porous media. In: Gallagher RH
et al., editors. Finite elements in uids. UK: Wiley; 1975 [chapter 8].
[2] Desai CS. Finite element residual schemes for unconned ow. Int J Num Meth
Eng 1976;10:14158.
[3] Desai CS, Li GC. A residual ow procedure and application for free surface ow
in porous media. Adv Water Resour 1983;6:2735.

Y.-x. Jie et al. / Computers and Geotechnics 52 (2013) 16

[4] Bathe KJ, Khoshgoftaar MR. Finite element free surface seepage analysis
without mesh iteration. Int J Num Anal Meth Geomech 1979;3:1322.
[5] Lam L, Fredlund DG. Saturatedunsaturated transient nite element
seepage model for geotechnical engineering. Adv Water Resour 1984;7:1326.
[6] Zhang JH, Xu QJ, Chen ZY. Seepage analysis based on the unied unsaturated
soil theory. Mech Res Commun 2001;28(1):10712.
[7] Darbandi M, Torabi SO, Saadat M, Daghighi Y, Jarrahbashi D. A moving-mesh
nite-volume method to solve free-surface seepage problem in arbitrary
geometries. Int J Num Anal Meth Geomech 2007;31:160929.
[8] Jiang QH, Deng SS, Zhou CB, Lu WB. Modeling unconned seepage ow using
three-dimensional
numerical
manifold
method.
J
Hydrodyn
2010;22(4):55461.
[9] Jie YX, Jie GZ, Mao ZY, Li GX. Seepage analysis based on boundarytted
coordinate
transformation
method.
Comput
Geotech
2004;31(4):27983.
[10] Li GX, Ge JH, Jie YX. Free surface seepage analysis based on the element-free
method. Mech Res Commun 2003;30:919.
[11] Zheng H, Liu DF, Lee CF, Tham LG. A new formulation of Signorinis type for
seepage problems with free surfaces. Int J Num Meth Eng 2005;64(1):116.
[12] Gao MK. Practical probability and statistics: probability and its
application. Beijing: Defence Industry Press; 1988 [in Chinese].
[13] Gong GL, Qian MP. Applied random process and random model in algorithm
and intelligent computation. Beijing: Tsinghua University Press; 2004 [in
Chinese].
[14] Du Q, Gunzburger M, Ju L. Meshfree, probabilistic determination of point sets
and support regions for meshless computing. Comput Meth Appl Mech Eng
2002;191:134966.
[15] Ju L, Du Q, Gunzburger M. Probabilistic methods for centroidal Voronoi
tessellations and their parallel implementations. Parallel Comput
2002;28:1477500.

[16] Bruen MP, Osman YZ. Sensitivity of streamaquifer seepage to spatial


variability of the saturated hydraulic conductivity of the aquifer. J Hydrol
2004;293:289302.
[17] Chu-Agor ML, Wilson GV, Fox GA. Numerical modeling of bank instability by
seepage erosion undercutting of layered streambanks. J Hydrol Eng
2008;13(12):113345.
[18] Grifths DV, Fenton GA. Three-dimensional seepage through spatially random
soil. J Geotech Geoenviron Eng 1997;123(2):15360.
[19] Grifths DV, Fenton GA. Probabilistic analysis of exit gradients due to steady
seepage. J Geotech Geoenviron Eng 1998;124(9):78997.
[20] Haldar S, Babu GLS. Effect of soil spatial variability on the response of laterally
loaded pile in undrained clay. Comput Geotech 2008;35(4):53747.
[21] Jeyisanker K, Gunaratne M. Analysis of water seepage in a pavement system
using the particulate approach. Comput Geotech 2009;36(4):64154.
[22] Keery J, Binley A, Crook N, Smith JWN. Temporal and spatial variability of
groundwatersurface water uxes: development and application of an
analytical method using temperature time series. J Hydrol 2007;336:116.
[23] Srivastava A, Babu GLS, Haldar S. Inuence of spatial variability of permeability
property on steady state seepage ow and slope stability analysis. Eng Geol
2010;110:93101.
[24] Zheng GL, Wang Y, Pei LH. High precision numerical integral on free surface
boundary of 3-D unsteady seepage problem. Adv Water Sci 2008;19(1):1238
[in Chinese].
[25] Zhou CB, Xiong WL, Liang YG. A new method for the solution of unconned
seepage eld. J Hydrodyn 1996;11(5):52834 [in Chinese].
[26] Akai K, Onishi Y, Nishigaki M. Finite element analysis on saturated
unsaturated seepage problem. Proc Civil Eng Soc 1977;264:8796 [in
Japanese].
[27] Jie YX, Wen YF, Deng G, Chen R, Xu ZP. Impact of soil deformation on phreatic
line in earth-ll dams. Comput Geosci 2012;46:4450.

You might also like