Computers and Geotechnics 32 (2005) 139–152

www.elsevier.com/locate/compgeo

Simple genetic algorithm search for critical non-circular

failure surface in slope stability analysis
Ali R. Zolfaghari *, Andrew C. Heath, Paul F. McCombie
Department of Architecture & Civil Engineering, University of Bath, BA2 7AY, UK

Received 3 September 2004; received in revised form 4 January 2005; accepted 7 February 2005
Available online 19 March 2005

Abstract

Soil slope stability problems in engineering works are usually analysed using limit equilibrium methods. A number of existing
methods are based on finding the critical circular failure surface for homogeneous soils, but failure surfaces tend to be non-circular
for layered slopes. A simple genetic algorithm is presented to search the critical non-circular failure surface in slope stability analysis
and is used to solve the Morgenstern–Price method to find the factor of safety. The pseudo-static horizontal and vertical forces due
to earthquake and surcharge load due to existing buildings and structures on natural slopes are included in the analysis.
Ó 2005 Elsevier Ltd. All rights reserved.

1. Introduction the independent variables x describing the geometry of

the failure surface. He employed the simplex method
There are many different ways to compute the factor as the optimisation technique. Celestina and Duncan
of safety of earth dams or natural slopes including limit [5] used the same approach for non-circular failure sur-
equilibrium, finite element and finite difference methods. faces, but used the alternating-variable optimisation
In recent years the finite element method has been used technique. Li and White [8] proposed a more efficient
for slope stability analysis, but limit equilibrium meth- one-dimensional optimisation technique to replace the
ods are still common practice. quadratic interpolation method, which Celestina and
Many methods have been presented to compute the Duncan [5] used in the alternating-variable technique.
factor of safety using limit equilibrium with a circular Baker [1] defined the failure surface by a number of nodal
failure surface [3,9,12]. A simple circular failure surface points connected by straight lines. The vertical coordi-
method is sufficient for a slope in a homogenous soil nates of the nodal points are the variables in Baker
s
layer, while for a heterogeneous multi-soil layers slope, method and the dynamic programming technique is em-
a non-circular failure surface method should be consid- ployed as the optimisation method. Bolton et al. [4] de-
ered as circular methods can over predict the factor of fined a global optimisation algorithm for finding the
safety. Limit equilibrium has also been used for non-cir- critical failure surface by nodal points connected by
cular failure surfaces [1,2,4,5,8,11], and some of these straight lines for any shape of failure. Bardet and Kapus-
methods are summarized below. kar [2] presented a simple method of optimisation to
Nguyen [11] developed a method where the factor of search the critical failure surface using the downhill sim-
safety is formulated as a multivariate function F(x) with plex algorithm. A large number of computations are
needed to find the critical failure surface, as an arbitrary
nodal coordinate could be irrelevant among the rest of
*
Corresponding author. Fax: +01225 386691.
created nodal coordinates. For example, a fluctuated
E-mail addresses: abparz@bath.ac.uk (A.R. Zolfaghari), absah@ failure surface shape could be created using nodal coor-
bath.ac.uk (A.C. Heath), abspfm@bath.ac.uk (P.F. McCombie). dinates leading to an unrealistic failure surface (Fig. 1).

0266-352X/$ - see front matter Ó 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compgeo.2005.02.001
140 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152

Nomenclature

ai failure-line slope (slope of base of slice) gi slope of inter-slice force Zi

af i failure-line slope of slice with respect to verti- gi + 1 slope of inter-slice force Zi + 1
cal P total reaction normal to base of slice
Dafi difference between two subsequent base slice P0 reaction (due to effective stress) normal to
failure slopes ðDaf2 ¼ af2
af1 Þ base of slice
ah i failure-line slope of slice with respect to hor- Qi resultant of pair of inter-slice forces Zi and
izontal Zi + 1
ah pseudo-static horizontal coefficient of earth- q surcharge along the slope
quake S total shear force available
av pseudo-static vertical coefficient of earth- u pore-pressure
quake Sm total shear force mobilized
b inclination of the chord connecting the two W weight of slice
ends of the natural slope xqi ; y qi coordinates of acting point of inter-slice
b width of slice resultant forces, Qi
c0 cohesion (for effective stress) x0i normalized coordinate of slice i
F factor of safety X ei ; Y ei coordinates of mid point of slice base
/0 angle of shearing resistance (for effective X gi ; Y gi coordinates centre of gravity of slice
stress) X si ; Y si coordinates of left side of slice at ground sur-
/0m mobilized angle of shearing resistance face
ðtan /0m ¼ tan /0 =F Þ X sðiþ1Þ ; Y sðiþ1Þ coordinates of right side of slice at
hQ height of intersection of resultant inter-slice ground surface
forces X fi ; Y fi coordinates of left side of slice on failure sur-
hG height of centre of gravity of slice face
Nslice total number of slices, default value is 150 X fðiþ1Þ ; Y fðiþ1Þ coordinates of right side of slice on fail-
c bulk density ure surface
ci average bulk density of each slice Zi inter-slice force on left side of slice i
hi slope of resultant (Qi) of pair of inter-slice Zi + 1 inter-slice force on right side of slice i
forces

In this study, instead of searching and optimising searching, because the slope of the failure surface of
along nodal y-coordinates, the search through failure- each slice (a in Fig. 2) is related to the slope of adjacent
line slopes using a simple genetic algorithm is presented. slices. In a slope stability analysis where the failing mass
The search through failure-line slopes is much more effi- moves from left to right, the angle of the base of a slice
cient and quicker to solve than nodal y-coordinates (a) is usually shifted counter-clockwise when moving

(x1,y1)

(x2,y2) (xn,yn)

(x3,y3)

(xn-1,yn-1)

Fig. 1. A fluctuating failure surface as a case in a Nodal Optimization Method.

 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152 141

from left to right (e.g., Fig. 2). Searching for a failure circular failure surface if this is the critical failure surface
surface using nodal y-coordinates cannot include this as- for the particular problem.
pect easily, possibly resulting in an unrealistic failure
surface as shown in Fig. 1.
This paper presents a simple computation format of 2. Presenting a simple solution for Morgenstern–Price
the Morgenstern–Price method for the non-circular method
slope stability analysis with pseudo-static horizontal
and vertical forces due to earthquake loading. The main Although similar to the Spencer method [12], the
reason for using the Morgenstern–Price is that it is a so- Morgenstern–Price method [10] was selected for the
called rigorous solution (i.e., rigorous in that both force analysis using the simple computation format.
and moment equilibrium are satisfied if one can make Fig. 2 shows a natural slope with a head to toe angle
certain assumptions), and produces realistic answers of b. Fig. 3 shows details of inter-slice forces for the slice
for surfaces which require significant internal distortion number i. As described in more detail in Appendix A,
of the sliding mass of soil. The option of surcharge load- the resultant of interslice forces in each slice can be writ-
ing is also included, which can be used to model the ef- ten as follows:

0
c0 bsec a
F
þ tanF / ðW  cos a
W  av  cos a
W  ah  sin a
u  b  sec a þ q  b  cos aÞ
W  sin a þ W  av  sin a
W  ah  cos a
q  b  sin a
Qi ¼ 0
cosða
hi Þ  ð1 þ tanða
hi Þ  tanF / Þ
ð1Þ

fect of buildings on slope stability. An important feature In order to satisfy equilibrium equations, the summa-
is that no assumptions are required with regards to the tion of resultant interslice forces and overall moment
geometry of the failure surface and no restrictions are over an optional point must be zero. In this case the mo-
placed on the positions of the initiation and termination ments about the origin (x = 0, y = 0) are taken to be
point of the failure surface. zero:
The simple genetic algorithm used in this study has X
two purposes: firstly to find the critical non-circular fail- ðQi  cos hi Þ ¼ 0 ð2Þ
ure surface in finite or infinite slopes, and secondly to
X
solve the Morgenstern–Price method to find the factor ðQi  sin hi Þ ¼ 0 ð3Þ
of safety corresponding to the critical failure surface.
As circular failure surfaces are a subset of more general X X
ðMÞ ¼ ðQi  cos hi  Y qi þ Qi  sin hi  X qi Þ ¼ 0 ð4Þ
non-circular surfaces, the proposed method will find a

x1 b

β
α
yi

xi

Fig. 2. General failure surface in a slope stability analysis.

142 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152

q b

Slice No: i
A
c'.b
/F/
q.b B co
sα
ηi
Zi W.av W.av q.b S/F P't
an
W.ah η W.ah φm
i+ G
W 1

θi
F
φm

hG
S/F W Q W

P'
S/F α Zi+1 α P

hQ
D

sα
E

/co
W.av

u.b
α

θi
C Q
W.ah
P P

Coordinates of point A: Xsi, Ysi Coordinates of point B: Xs(i+1) , Ys(i+1)

Coordinates of point C: Xf(i+1) , Yf(i+1) Coordinates of point D: Xfi, Yfi
Coordinates of point E: Xei, Yei Coordinates of point F: Xqi , Yqi
Coordinates of point G: Xgi, Ygi
Fig. 3. (a)–(c) Inter-slice forces in slice number i. Coordinates of point A: X si ; Y si . Coordinates of point B: X sðiþ1Þ ; Y sðiþ1Þ . Coordinates of point C:
X fðiþ1Þ ; Y fðiþ1Þ . Coordinates of point D: X fi ; Y fi . Coordinates of point E: X ei ; Y ei . Coordinates of point F: X qi ; Y qi . Coordinates of point G: X gi ; Y gi .

To find the factor of safety, Eq. (2) or (Eq. (3)) and Eq. tool. A simple genetic algorithm (SGA) uses strings of
(4) need to be solved. There will be two equations and binary coding, 0 and 1, to encode whatever information
two unknowns such as k and F. is needed to define a distinct solution to a problem. This
If it is assumed a pair of (F*, k*) is one of the answers solution may then be tested to produce a fitness value.
to the two equations f(F, k) = 0 and g(F, k) = 0, then For example, if the goal is to find three unknown values
(F*, k*) will also be an answer to the following equation: such as x, y, z, then each chromosome will be a string of
jf ðF ; kÞj þ jgðF ; kÞj ¼ 0 ð5Þ binary digits of x, y, and z. Comparison may be made
with biological coding, which uses units with four possi-
Using the above algebraic theory, Eqs. (2) and (4) can be ble values in DNA. Clearly, real variables such as coor-
written as follows: dinates need to be expressed as integer values by

X

X





breaking up the possible range into a number of steps,

M
þ
ðQi  cos hi Þ

so for example xint = round(256(x
xmin)/(xmax
xmin))

X



for an eight bit binary code.
¼
ðQi  cos hi  Y i þ Qi  sin hi  X i Þ


X
The simple genetic algorithm works in two main


stages: creating the initial population and reproduction.
þ
ðQi  cos hi Þ
¼ 0 ð6Þ
First the initial population is created and each number is
Eq. (6) is now solved using the simple genetic algorithm stored in a chromosome in the binary format. A fitness
that is explained in Section 3.1. value associated with each chromosome is calculated.
The population is then sorted in descending or ascend-
ing order according to the fitness value. A classical
3. Simple genetic algorithm way of ensuring that the best solution is never lost is
to copy the best individual to the next generation. Then
The simple genetic algorithm refers to a model intro- half of the population is selected for the reproduction
duced and investigated by Holland [7] which uses the process. A crossover process with a probability of
concepts of genetics in a specific way as an optimisation 0.7–0.9 is applied to two selected parent chromosomes.

(a): chromosome of F and λ (b): chromosome of x 1, αf1, ∆αf2, ∆αf3,… , ∆αfn:

F λ X1 αf1 ∆αf2 ∆αf3 … ∆αfn

1 1 0 1 0 1 1 0 1 1 1 0 1 0 0 1 1 1 1 0 1 0 0 1 0 1 1 0 0 1 1 0 1 0 1 1 0 0 0 1 1 1…1 0 1 0

Fig. 4. Examples of two chromosomes.

 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152 143

αf 1=αf 1 αh1=270+αf 1 α1=360-αh1

αh1 αf 2=αf 1+∆αf 2 αh2=270+αf 2 α2=360-αh2
αf 3=αf 2+∆αf 3 αh3=270+αf 3 α3=360-αh3
1
αf

∆α f 2
αf 2

∆αf 3
3
β
αf

αi

Fig. 5. Non-circular failure surface.

A random position along the length of the chromosome the possibility of significant shifts away from the solu-
is selected and the values of each binary string are ex- tions currently being converged on, that overcomes
changed or crossed by swapping all characters after this problems associated with local maxima or minima in
position. The two new chromosomes created are known analyses. Each binary value in a chromosome selected
as children of those parents. Mutation is applied to a for mutation is swapped with a probability of 0.008,
small proportion of chromosomes, thus introducing i.e., each 0 or 1 is changed into 1 or 0 with a probability

Fig. 6. Slope stability analysis program (SlopeSGA) in VB6.

144 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152

of 8/1000. The fitness values of the new population, ulation can converge on all the solutions at the same
which include both children and parent chromosomes, time. It may be necessary to adjust the parameters of
are then calculated. the search to ensure that local maxima (or minima, in
The process of reproduction, crossover, mutation and this case) are not rejected too early.
evaluation is repeated as a cycle of generation. A num- McCombie and Wilkinson [9] developed a simple ge-
ber of cycles are performed until an optimal solution netic algorithm to search for the minimum factor of
is determined. Goldberg [6] refers to this basic imple- safety of a circular failure surface in slope stability
mentation as a simple genetic algorithm (SGA). The analysis. They presented a three-dimensional chromo-
selection of the part of each generation to go into the some coding containing the x and y coordinates of
reproduction process means that the population as a the centre of a circle and the radius of a circular failure
whole tends to get fitter. This might mean that the pop- surface. They also showed that replacing the radius
ulation tends towards being identical, or if there are a with a tangent level, or with the coordinates of a point
number of solutions with a similar fitness, then the pop- the circle had to pass through (thus creating a four

Start

Create population of size N of F and λ

Convert F and λ into binary format and store them in a chromosome

Solve equation 6: Fitness = Σ ( Qi . cos θ i .Yi + Qi . sin θ i . Xi ) + Σ ( Qi . cos θ i )

Sorting data ascending based on fitness value, taking top half of population for reproduction

Reproduction Process Selecting a pair of chromosomes as

parent 1 and parent 2

Rnd <
Crossover Process Yes Crossover
Probability No

Rnd <
Mutation Mutation Process
Probability Yes

No

Generating children 1 and children 2

Solve equation 6: Fitness = Σ ( Qi . cos θ i .Yi + Qi . sin θ i . Xi ) + Σ ( Qi . cos θ i )

No Is Required No of
Cycles exceeded?

Sorting data ascending based on fitness value

F and λ corresponding with the lowest fitness value are the solution

End

Fig. 7. Simple genetic algorithm to find factor of safety.

 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152 145

dimensional search space), would usually work better, expected to produce more efficient convergence by pre-
as the formulation of the problem becomes closer to venting unrealistic failure surfaces similar to those
what determines the fitness of each chromosome. In shown in Fig. 1.
addition, the use of a simple genetic algorithm was
found to be more efficient at solving slope stability 3.1. Using the simple genetic algorithm to solve Eq. (6) in
problems than ‘‘traditional’’ numerical optimisation order to find the lowest factor of safety of a non-circular
methods. While this previous research is only applica- failure surface
ble to circular failure surfaces, the solution algorithm
can be extended to non-circular failure surfaces. In this Eq. (6) shows there are two unknown variables (hi
case definition of the surface by the change of angle at and F). Using Eq. (14) or (15), hi can be replaced with
points along it is a more natural definition of the prob- k. As in any numerical solution technique, the initial
lem than using y co-ordinates, and can, therefore, be population (values) of F and k has a considerable effect

Start

Create population of size N of x1, αf and ∆αf (N equals to no of slice, Nslice-default=150)

Convert x1, αf and ∆αf into binary format and store them in a chromosome

Failure Morgenstern-Price
No surface hits the Yes SGA method to get F
F = 100
slope geometry

Sorting data ascending based on Factor of safety, taking top half of population for reproduction

Reproduction Process Selecting a pair of chromosomes as

parent 1 and parent 2

Yes Rnd <

Crossover Process Crossover
No
Probability

Rnd <
Crossover Mutation Process
Yes
Probability
Generating children 1 and children 2
No

No Failure surface hits Yes

F = 100 the slope geometry
Morgenstern-Price
SGA method to get F
Is
No Required No of
Cycles exceeded?
Yes
Sorting data ascending based on fitness value (factor of safety)

Critical failure surface is corresponding with the lowest factor of safety

End

Fig. 8. Simple genetic algorithm to find non-circular failure surface.

146 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152

X

on the convergence rate. The initial value of k is as-


Fitness ¼
ðQi  cos hi  Y i þ Qi  sin hi  X i Þ

sumed to be approximately 0.7 tan b [13].
X

Let N be the size of the population for F and k, and


þ
ðQi  cos hi Þ

0.5 the percentage of reproduction. In order to easily
convert F into binary code, in the initial population, val- The chromosomes are sorted into ascending order
ues of F are created in a range of 1–1000 as integer val- according to their fitness value and the half with the low-
ues. Values of k are created in a range around est fitness value is selected for the reproduction process.
100*0.7 tan b (0.7 tan b is a real value). These values are The crossover and mutation are applied to chromo-
converted into binary format and stored in a string as somes during the reproduction process as described
a chromosome (Fig. 4). The values of F and k are di- above. The population is subjected to a number of ge-
vided by 100 to obtain their real values in the calcula- netic cycles in order to find the F and k that minimize
tion. The fitness value is calculated for every the fitness value. All these calculations are performed
combination of F and k in the population using Eq. (6). in a slope stability analysis program, SlopeSGA that is

52
Y (m) Factor of Safety Bishop Morgenstern
A Circular 1.74 1.76
50
Non-Circular * 1.75
* Bishop Method is for a Circular Failure Surface
48
Circular-Failure
Non-Circular Failure
46
(m) X Y
A 15 50
44 B 32 41.5

42 B
C' =1500 kg/m2
40 φ' = 20
γ = 1900 kg/m3
38
10 15 20 25 30 X (m) 35

Fig. 9. An example of a natural slope with a homogenous soil layer.

Fitness Value=
Abs(Qcosθ) + Abs(M)
(kg.m) Fs=1.75

10000.00

1000.00

100.00

10.00

1.00

0.10
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
λ

Fig. 10. Converging the fitness value in Morgenstern–Price method.

 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152 147

written in Visual Basic 6 by the first author, as depicted of populations for x1. Therefore, Nx1 cases of x1 are cre-
in Figs. 6 and 7. ated randomly. Now let Naf be the number of popula-
tions for af for each x1. Therefore, Naf cases of af1 are
3.2. Using the simple genetic algorithm to find critical randomly created for each x1. The initial value of af1
non-circular failure surface is chosen randomly around the Rankine failure angle
range. As the Rankine failure angle with respect to the
As the initial part of the simple genetic algorithm horizontal is 45 + //2, the af1 range will be 45
//2 or
method, a population of all searching parameters needs 30–45, for / equals to 30 and 0, respectively (Fig. 5).
to be created. The first x coordinate of the failure sur- These two values of x1 and af1 are converted into binary
face, x1, is created randomly. Let Nx1 be the number code and stored in a chromosome.

Minimum Factor of
Safety

2.09

2.04

1.99

1.94

1.89

1.84

1.79

1.74
0 5 10 15 20 25 30
Generation

Fig. 11. A comparison between random reproduction and the SGA.

Factor of
Safety
2.8
2.7
Simple Genetic Algorithm
2.6
Min-Factor of Safety=1.75
2.5
2.4
2.3
Simplex Method
2.2 Min-Factor of Safety=1.75
2.1
2
1.9
1.8
1.7
1.6
0 10 20 30 Time(sec) 40
Fig. 12. A comparison between the SGA and simplex method.
148 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152

After choosing x1 and af1 for a failure surface, the rest Now let Nslice be the total number of slices for a non-
of the failure-line slopes through the failure surface need circular failure surface. Dafi is randomly selected
to be chosen. As shown in Fig. 5, each failure-line slope through a reasonable range. In order to start with a suf-
has a relation to the previous one, so instead of choosing ficient population, 11 different categories of Dafi range
failure-line coordinates as in previous research, e.g. [1], are defined. For example, the failure-lines slopes can
the angular difference Dafi between each successive fail- be very rapid (Dafi between 5° and 15°), rapid (Dafi be-
ure-line slope is chosen randomly for all slices. This al- tween 0° and 10°), gentle (Dafi between 0° and 5°), and
lows the population of irrelevant failure-lines slopes to very gentle (Dafi between 0° and 3°). More cases are cre-
be ignored easily. ated, such as the failure surface continues horizontally

51
A Non-circular failure, FS-Morgenstern=1.24
Circular failure, FS-Bishop=1.475, FS-Morgenstern=1.5
1
E (m) X Y
B
A 15 50
2
B 19 48
F C 30.5 42.5
D 32 41.5
G 3
E 10 48.5
4 F 10 46.7
G 10 46.2

Layers 1 2 3 4
c' (kg/m2) 1500 1700 500 3500 C
φ' 20 21 10 28
D
γ (kg/m3) 1900 1900 1900 1900
41
10 15 20 m 25 30

Fig. 13. Comparison of a circular with a non-circular failure surface method.

55
* * * EQ & GW No EQ, No GW
Y (m)

Phreatic water table * * * EQ,(No GW) - - - - GW, (No EQ)

(m) X Y
A A 15 50
1 B 19 48
F B C 25 45
2 D 27 44
J E 32 41.5
G
F 10 48.1
C G 10 46.7
H H 10 45
3 D
I I 10 44
4 J 17 46.7

Layers 1 2 3 4
c' (kg/m2) 1500 1700 500 3500
φ' 20 21 10 28
γ (kg/m3) 1900 1900 1900 1900
35
10 15 20 25 30 X (m) 35 40
Fig. 14. Example of a natural slope with a complex soil layers.
 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152 149

Table 1
Results of slope stability analysis in example 3
Loading Method No. of slice Factor of safety (Morgenstern–Price)
No earthquake – No ground water Non-circular 40 1.48
ah = 0.1 – No ground water Non-circular 44 1.37
No earthquake – Ground water Non-circular 41 1.36
ah = 0.1 – Ground water Non-circular 45 0.98

70
(m) X Y
ah = 0, FS = 1.14 A 0 60
A
ah = 0.1, FS = 0.95 B 0 55
B
C C 0 53
D 97.17 0
E 100.83 0
35 F 110 0

Layers 1 2 3
c' (kg/m2) 4000 3000 5000 21
φ' 17 10 27
3
γ (kg/m3) 1900 1900 1900
0
m D E F
0 60 120
Fig. 15. Results of the infinite slope analysis in example 4.

once the failure-line slope becomes horizontal. The ele- All these calculations are performed in the slope sta-
ven cases make the simple genetic algorithm more suc- bility analysis program as depicted in Figs. 6 and 8.
cessful in finding the critical non-circular failure
surface. The values of Dafi are converted into binary
code and stored in the chromosome containing x1 and 4. Examples
af1 (Fig. 4).
The crossover and mutation are applied to the chro- For the purpose of illustration, four examples of nat-
mosomes in the reproduction process as described in ural slopes are analysed using the simple genetic algo-
Section 3. As the chromosome represents a search space rithm method proposed in this paper. The
of Nslice + 2, or usually 152 dimensions (based on the de- aforementioned slope stability analysis program is used
fault value where Nslice equals to 150), more crossover to analyse these examples as follows:
cases need to be applied in order to generate more (a) An example of a natural slope with a homogenous
new chromosomes. Five different cases of crossover soil layer, as shown in Fig. 9, is analysed. The factor of
are applied: one point crossover in the x1 coding of safety is calculated for the slope using the Bishop and
the chromosome, two point crossover in the x1 and af1 Morgenstern–Price methods for both circular and non-
coding parts of chromosome, one point crossover circular methods. Fig. 10 shows how the fitness value,
through the whole chromosome, one point and two abs(Q cos q) + abs(M) in Eq. (6), is changed and eventu-
points crossover in the Dafi coding of chromosome. ally converged to a negligible value as the solution con-
The fitness value is the factor of safety against failure verges to the minimum factor of safety. The simple
for the surface, which is calculated using the simple ge- genetic algorithm method is also compared to a random
netic algorithm method, as explained in Section 3.1. If reproduction approach in Fig. 11 and to the simplex
the created non-circular failure surface does not hit the method in Fig. 12.
geometry of the natural slope, a large value is given to In Fig. 11, comparison was made between random
the fitness value and the failure surface is subsequently reproduction and the simple genetic algorithm over 30
excluded from the population through the ‘‘survival of cycles. In the random reproduction, a population of
the fittest’’ characteristic of the genetic algorithm. A 100 was set, and then the best of 50% of the population
number of genetic cycles are calculated in order to find is kept after each reproduction. The thick lines show the
the critical non-circular failure surface with the lowest lowest factor of safety achieved after each set of 100
factor of safety. analyses of randomly generated surfaces (eventual
150 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152

minimum 1.86), whilst the thin lines show the corre- However, relating the slopes of adjacent slices results
sponding values with genetic reproduction producing in greatly increased computational efficiency. The pre-
each generation (minimum 1.75). The increased conver- sented simple genetic algorithm method can be applied
gence rate for the simple genetic algorithm is clear. The to find the non-circular failure surface with the lowest
code is written in Visual Basic 6 which is slow compared factor of safety very quickly compared to random or
to other packages, but this does allow easy comparison simplex method approaches. The Morgenstern–Price
of the relative efficiency of different methods to by the method can be easily solved with the simple genetic algo-
time taken for each run. The simple genetic algorithm rithm in order to obtain the factor of safety for a variety
run took 9 s for the 30 generations. of slope geometries and loading conditions. The results
Fig. 12 shows the results from solving the same exam- of this study suggest that the presented searching method
ple using the simplex method presented by Bardet and could be used in order to analyse the stability of earth
Kapuskar [2]. In this case this method was modified dams, finite or effectively infinite natural slopes and any
by the fact that slope of the base of any slice is related other geotechnical problem with layered or unlayered
to the slope of adjacent slices, as described above. This geology. The option of a surcharge load and pseudo-static
code was also written in Visual Basic 6, and took 41 s horizontal and vertical forces due to earthquake loading
to find the non-circular failure surface with the factor are included to enable a comprehensive evaluation of
of safety of 1.75. The comparison between the simple ge- slope stability.
netic algorithm and the simplex method, Fig. 12, shows For a slope in a homogenous material, the non-lin-
the rapid convergence of the simple genetic algorithm ear algorithm approximates a circular failure surface
compared to the simplex method. As the SGA is started and predicts a similar factor of safety to that for a cir-
with a wider population (i.e., 100 population), therefore, cular failure surface. For a slope with a layered struc-
the factor of safety in this method is a lower value at the ture, the circular methods can over predict the factor
beginning of the search compare to the simplex method of safety, which might lead to unconservative estimates
(Fig. 12). of slope stability. In these cases, non-circular slope sta-
(b) An example of a slope with complex soil layering, bility analysis is essential for reliable assessment of
as shown in Fig. 13, is analysed using the proposed SGA stability.
method. In this example the difference between the fac-
tor of safety in circular and non-circular failure surface
methods is presented. As expected, the non-circular fail- Appendix A. Complete formulation of the solution for the
ure surface is drawn towards the weakest layer, resulting Morgenstern–Price method
in a lower factor of safety than that for the circular
surface. (a) Weight of slice number i:
(c) Another natural slope with complex soil layering,  
ðY sðiþ1Þ
Y fðiþ1Þ Þ þ ðY si
Y fi Þ
as shown in Fig. 14, is analysed using the presented W ¼  b  ci
method. Four loading cases are considered: water pres- 2
sure and earthquake loading; water pressure and no ðA:1Þ
earthquake loading; earthquake loading and no water (b) Total reaction of normal in the base of the slice, P,
pressure; and absence of both water pressure and earth- which can be presented as:
quake loading. A pseudo-static horizontal coefficient of 1. Force (P 0 ) due to the effective stress.
earthquake loading of 0.1 is used in this analysis. The re- 2. Force (u Æ b Æ sec a) due to the pore pressure (u)
sults are presented in Table 1. As expected, an increase
in water pressure resulted in a decreased factor of safety, P ¼ P 0 þ u  b  sec a ðA:2Þ
and an increase in pseudo-static horizontal earthquake
loading decreased the factor of safety. (c) Mobilized shear force (Sm = S/F)
(d) An infinite slope, as shown in Fig. 15, is analysed
using the simple genetic algorithm. In this example, a S ¼ c0 b  sec a þ P 0  tan /0 ðA:3Þ
soft layer is located between two layers with higher
S m ¼ c0  b  sec a=F þ P 0  tan /0 =F ðA:4Þ
strength. A pseudo-static horizontal coefficient of earth-
quake loading of 0.1 is assumed. (d) Pseudo-static horizontal force due to earthquake,
ah Æ W
(e) Pseudo-static vertical force due to earthquake,
5. Conclusions av Æ W.
(f) Surcharge force in the slice due to surcharge load
Many previous approaches to determining the non- along the natural slope, q Æ b.
linear failure surface assumed the slope of the base of (g) Inter-slice forces Zi and Zi + 1 with horizontal angle
any slice is independent of the slope of adjacent slices. of gi and gi + 1, respectively.
 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152 151

(h) Qi, Resultant force of Zi and Zi + 1 forces. This the angle of inter-slice resultant force is equal for all
resultant force, Qi that acts with a horizontal angle the slices through the failure surface, Eq. (A.15).
of hi (Fig. 3), is calculated from equilibrium condi- Now let hi be the horizontal angle of inter-slice resul-
tion in two perpendicular directions, using Eqs. tant force Qi.
(A.5) and (A.6). tan hi ¼ k  f ðx0i Þ ðA:13Þ
For equilibrium in each slice, the sum of inter-slice Two cases for f ðx0i Þ will be as follows:
forces in the P and S/F direction must be zero. The pro- ðaÞ f ðx0i Þ ¼ sinðx0i Þ then hi ¼ arctanðk  sinðx0i ÞÞ
cess is shown in Eqs. (A.5) and (A.6), respectively,
ðMorgenstern–Price methodÞ ðA:14Þ
P þ Qi  sinða
hi Þ
W  cos a þ W  av  cos a
þ W  ah  sin a
q  b  cos a ¼ 0 ðA:5Þ ðbÞ f ðx0i Þ ¼ 1 then hi ¼ arctanðkÞ
ðSpencer methodÞ ðA:15Þ
S=F
Qi  cosða
hi Þ
W  sin a þ W  av  sin a
Now to find the factor of safety, Eq. (A.10) (or

W  ah  cos a
q  b  sin a ¼ 0 ðA:6Þ
Eq. (A.11)) and Eq. (A.12) need to be solved. There
Substituting Eqs. (A.3) and (A.4) in Eqs. (A.5) and will be two equations and two unknowns such as k
(A.6): and F.

0
c0 bsec a
F
þ tanF / ðW  cos a
W  av  cos a
W  ah  sin a
u  b  sec a þ q  b  cos aÞ
W  sin a þ W  av  sin a
W  ah  cos a
q  b  sin a
Qi ¼ 0
cosða
hi Þ  ð1 þ tanða
hi Þ  tanF / Þ
ðA:7Þ

For moment equilibrium in each slice, take the mo- Let assume a pair of (F*, k*) is one of the answers to
ment about the point E equal zero (Fig. 3): the two equations f(F, k) = 0 and g(F, k) = 0, therefore,
X (F*, k*) will also be an answer to the following equation:
M E ¼ Q  cos h  hQ
W  ah  hG ¼ 0 ðA:8Þ
jf ðF ; kÞj þ jgðF ; kÞj ¼ 0 ðA:16Þ
After calculating Q for each slice, hQ is calculated
using Eq. (A.8), then the y coordinate of point F in Using the above algebraic theory, Eqs. (A.10) and
Fig. 3, Y qi is: (A.12) can be written as follows:

X

X






Y qi ¼ Y ei þ hQ ðA:9Þ
M
þ
ðQi  cos hi Þ


X

Now for overall equilibrium in the natural slope, the


¼
ðQi  cos hi Þ  Y i þ Qi  sin hi  X i Þ

sum of horizontal and vertical inter-slice forces must be
X

zero.


þ
ðQi  cos hi Þ
¼ 0 ðA:17Þ
X
ðQi  cos hi Þ ¼ 0 ðA:10Þ Eq. (A.17) is now solved using the genetic algorithm that
X is explained in Section 3.1.
ðQi  sin hi Þ ¼ 0 ðA:11Þ
Furthermore, the sum of the overall moments about an References
arbitrary point must be zero, in this case let the mo-
ments about the origin (x = 0, y = 0) be zero: [1] Baker R. Determination of critical slip surface in slope stability
X X computations. Int J Numer Anal Meth Geomech 1980;4:333–59.
ðMÞ ¼ ðQi  cos hi  Y qi þ Qi  sin hi  X qi Þ ¼ 0 [2] Bardet JP, Kapuskar MM. A simplex analysis of slope stability.
Comput Geotech 1989;8:329–48.
ðA:12Þ [3] Bishop AW. The use of the slip circle in the stability analysis of
The vertical inter-slice force divided by the horizontal earth slopes. Geotechnique 1955;5(1):7–17.
[4] Bolton Hermanus PJ, Heymann G, Groenwold A. Global search
inter-slice force can be defined in terms of k  f ðx0i Þ for critical failure surface in slope stability analysis. Eng Opt
(Morgenstern–Price method [10]), where x0i is the line- 2003;35(1):51–65.
arly normalized xi coordinate with values at the two [5] Celestino TB, Duncan JM. Simplified search for non-circular slip
ends of the failure surface equal to zero and p, respec- surface. In: Proceedings of the 10th international conference on
tively. In this case, f ðx0i Þ is by convention assumed to soil mechanics and foundation engineering; 1981.
[6] Goldberg DE. Genetic algorithms in search, optimisation, and
be equal to sinðx0i Þ, therefore, the overall shape of machine learning. Addison-Wesley; 1989.
f ðx0i Þ on the failure surface is a half sin shape. In the [7] Holland J. Adaptation in natural and artificial systems. Univer-
Spencer method [10], f ðx0i Þ is equal to 1, therefore, sity of Michigan Press; 1975.
152 A.R. Zolfaghari et al. / Computers and Geotechnics 32 (2005) 139–152

[8] Li KS, White W. Rapid evaluation of the critical slip surface in [11] Nguyen VU. Determination of critical slope failure surface. J
slope stability problems. Int J Numer Anal Meth Geomech Geotech Eng 1985;111(2):238–50.
1987;11:440–73. [12] Spencer E. A method of analysis for stability of embank-
[9] McCombie P, Wilkinson P. The use of the simple genetic ments using parallel inter-slice forces. Geotechnique 1967;
algorithm in finding the critical factor of safety in slope stability 17:11–26.
analysis. Comput Geotech 2002;29:699–714. [13] Zhu DY, Lee CF, Qian QH, Zou ZS, Sun F. A new procedure for
[10] Morgenstern NR, Price VE. The analysis of the stability of computing the factor of safety using the Morgenstern–Price
general slip surfaces. Geotechnique 1965;15:79–93. method. Can Geotech J 2001;38(4):882–8.