CRITICAL FAILURE SURFACE LOCATION USING SIMPLE

GENETIC ALGORITHM AND MULTIPLE WEDGE SLOPE

STABILITY

Paul McCombie
Department of Architecture and Civil Engineering, University of Bath, United Kingdom

ABSTRACT: Established slope stability analysis methods cope well with moderately non-
circular shear surfaces, and the Simple Genetic Algorithm (SGA) has been used successfully
to find the critical slip surface. In complex situations, for example with concentrated soil
reinforcement, strong or weak layers, or concentrated applied loading, the assumptions made
to allow ‘rigorous’ analysis are likely to be simplistic. The results of individual analyses
could be in error, and the location of the critical slip surface is thus likely to be wrong. The
factor of safety could be overestimated because of the impossibility of full shear strength
being mobilised all along a shear surface at once, or underestimated because the need to
distort the sliding mass of soil as movement takes place is not properly accounted for. In such
cases, an engineering judgment has been necessary to identify the potential problem (if it has
been recognised at all), and hence an automated optimisation routine cannot work. The
multiple wedge method as used here avoids these problems, and so can be used successfully
with SGA to identify the critical surface. This is demonstrated here.

1 INTRODUCTION
The Simple Genetic Algorithm (SGA) has been used by a number of workers for the task
of locating the critical slip surface in slope stability analysis (Goh 1999; Goh 2000;
McCombie and Wilkinson 2002). In the case of circular failure surfaces the optimisation
problem that this represents is relatively straightforward, the search space being only three
dimensional (circle centre co-ordinates and radius), whereas for non-circular surfaces the
dimensions of the search space are equal to the number of lines defining the surface, if the
surface begins and ends at the ground surface. This represents a much more serious
optimisation problem. The use of SGA in tackling this problem is well described by Goh
(1999), in which the method of analysis used is as described by Donald and Giam (1989).
SGA has also been used by Zolfaghari et al (2005), using the method of Morgenstern and
Price (1965).
Methods of stability analysis for non-circular failure surfaces tend to be either methods of
slices (e.g. Janbu, 1957, Morgenstern and Price, 1965, Spencer, 1967) or methods of wedges
(e.g. Sarma, 1979, Donald and Giam, 1989), though composite analyses using finite element
analysis have been developed (e.g. Krahn, 2003). Rather than simply determining a factor of
safety against failure, which is what is needed in practice, most methods attempt a so-called
'rigorous' analysis which demonstrates equilibrium of both forces and moments, requiring
substantial assumptions to be made before obtaining the solution, which may be critically
dependent upon these assumptions. These assumptions usually relate to the degree of shear
force mobilisation on slice or wedge boundaries, and normal stress distribution. All methods
require an estimate to be made of the shape of the slip surface, but multiple wedge methods
also require the orientation of inter-wedge boundaries to be determined.
The method presented by McCombie (2009) avoids both of these problems by not
attempting a full 'rigorous' solution, simply using force equilibrium to obtain a factor of
safety, and by considering kinematics to determine the orientation of inter-wedge boundaries.
This makes the method particularly well suited for optimisation routines. The fact that
interwedge boundary angles are generated rather than specified reduces the number of
unknowns by approximately one third. Because there is no need to explore different
assumptions to allow 'rigorous' analysis, the process is very much more straightforward,
especially for significantly non-uniform soil masses or loading, for which the routinely used
assumptions are particularly likely to be seriously wrong. In adopting this approach, the
method attempts to reflect the unsurpassed elegance of Bishop's simplified method, which is
very widely used in practice because its solution is not crucially dependent upon assumptions
made on the basis of very little information (if any at all). Bishop's method focuses on
obtaining the required factor of safety, via a very simple assumption about the overall effect
of the unknowns for circular failure surfaces which has been justified by successful use in
practice over very many years. To be precise, the determination of the shear strength on the
slip surface is not dependent upon knowledge of the magnitude or distribution of the normal
forces on the sides of the slices; the effect of shear forces on individual slices was discounted
by Bishop in the simplified method, though the actual consequence of this simplification was
to neglect the overall effect of shear forces for all the slices, a much more robust assumption
which only becomes seriously erroneous if there is heavy concentrated loading. Shear force
and normal force distribution on the sides of the slices are simply accepted as unknown, and
not necessary for obtaining the factor of safety.

2 METHOD OF ANALYSIS
A full description of the method of analysis used in this work is presented in McCombie
(2009). The method uses kinematics to determine the orientation of inter-wedge boundaries,
as illustrated for a four wedge mechanism in Figure 1. Hodographs are used to describe
relative motions (Donald and Chen, 1997). The motions of each of wedge relative to the
underlying ground are shown as vectors (V1..4). The hodograph on the left in dotted lines
shows the case for no dilation. If the displacement along the shear surface is uniform, as
would be expected in the simplest situations, then the lengths of each of these vectors will be
the same. The vectors (V12..34) then show the relative displacement between each wedge.
The orientation of the wedge boundaries must correspond to the directions of these vectors,
which will bisect the angle formed by the bases of the wedges. If dilation occurs as the
displacement takes place and the shear strength is mobilised, then the directions of the
displacements will not be the same as the orientations of the surfaces, as shown in the
hodograph on the right, in which ψ denotes the angle of dilation. It can be seen that the inter-
wedge boundaries must be rotated backwards by twice the angle of dilation.
It may be noted that in a classic plasticity approach, the angle of dilation would be set at
the angle of friction, and displacements along the shear surface could probably no longer be
uniform but would become progressively larger towards the toe, so that peak friction might
be developed at the toe well before it is developed at the crest. Such an approach could in
any case only apply if the displacements were very small, because once they reached an order
similar to the size of the particles the angle of dilation would have dropped to a function of
the difference between critical state and peak angles of friction (0.8 is used here, after Bolton,
1986). The need for perfectly plastic behaviour would therefore never arise, as the approach
would already have failed to represent the real mechanisms of slope instability being
considered.

2ψ
2
3
4 2ψ
2ψ

V34 V4 V34 V4
ψ
ψ
V3 V3 ψ

V V ψ
23 23
V2 ψ V2 ψ

V V
1 1
V12 V12 ψ

ψ =0

Fig. 1. Hodographs for a four wedge mechanism. The thick lines in the hodograph on the right denote the
orientations of the surfaces.

This approach works well until a situation develops in which wedge boundaries overlap
(Figure 2). It is not sufficient to simply disallow such configurations, however, as there may
be constraints on the position of the slip surface which force this situation to arise. It could
be that the wedge boundaries would indeed intersect below the ground surface, perhaps
resulting in a block of soil which rotates as the sliding wedges pass beneath it. However, it
seems more likely that the displacements along the shear surface will not be uniform, which
is what must happen if the boundaries are rotated to remove overlaps. This implies that the
full shear strength cannot be mobilised everywhere along the shear surface at the same time
unless the soil is perfectly plastic.
The method uses the relative displacements on the slip surface and inter-wedge boundaries
to determine the shear force mobilised. The degree of mobilisation of both frictional and
cohesive shear strength can be simply related to the displacement on the surface by reference
to shear box testing, for example, and modeled in as simple or as complex a way as the
engineer requires. This would be expected to include some post-peak behaviour, so that the
shear strength reduces as deformations become large. For a simple analysis in which the
peak strength is not reached anywhere, a reliable factor of safety will be obtained using a
simple linear development of strength with displacement up to peak, and even a simple
representation of post-peak behaviour will give an indication of what might happen if the
slope is marginally stable or will fail.

Wedge boundaries for

constant velocity 1 1
overlap

2 2

5 5 Boundaries
4 4 rotated to
remove overlaps
3 3

Dilation is zero for 4

clarity
3
Velocities are reduced
where surface turns a
tight corner

2
1

Fig. 2. Overlapping inter-wedge boundaries.

The analysis is carried out by the application of increments of displacement until force
equilibrium can be obtained for all the wedges. Each wedge is analysed in turn. The
frictional shear force is related to the effective normal force on a boundary by the mobilised
angle of friction, as shown in figure 3.

Directions known, hence

magnitudes calculated

Cohesion

Effective
normal
Vectors force and
frictional
Wedge shear force

Fig. 3. Determination of forces on the first wedge.

Because the directions of the combined effective normal and shear forces are known, the
unknown magnitudes on the two boundaries can be determined through the vector diagram
shown, which is equivalent to solving two simultaneous equations. For each subsequent
wedge there are similarly two unknown forces, until at the final wedge there needs to be an
out-of-balance force applied to maintain equilibrium. The applied displacements, and hence
mobilisation of friction, are increased until this out-of-balance force becomes zero. If the
displacement along the boundary is not uniform, some parts of the slip surface may have
passed peak strength while other parts have not yet reached peak strength. As more and more
of the surface passes the displacement for peak strength, it may become impossible to obtain
this zero out-of-balance force, and a factor of safety of less than one is obtained by
considering the forces at the displacement which attained the minimum out-of-balance force.
It is possible that equilibrium cannot be achieved at all, even though it might have been
could the strength be mobilised everywhere at once. This effect is generally not considered
in slope stability analysis; the term 'progressive failure' is normally taken to apply to large
slopes where compression of the soil results in the non-uniform displacements that lead to
uneven mobilisation of shear strength. The possibility that the geometry of the shear surface
alone may account for uneven mobilisation of shear strength has probably not been
recognised. This effect turns out to be sufficiently important for the failure of Carsington
Embankment (see below) that it can completely account for the observed failure, without any
need to call on compression of the soil. Given that non-circular failures are most likely to
occur in situations which are significantly non-homogenous, the effect of these conditions in
forcing the failure surface away from an approximation of a circle is probably very important.
The ability to account for these effects must be an important factor in the production of an
optimisation routine aimed at finding the critical failure surface.
There is some dependence in this method on the initial conditions in the slope, in that the
development of shear resistance is related to the displacement on a surface. For example, soil
towards the toe of an excavated slope would be expected to have a greater part of the shear
strength already mobilised due to some kind of Ko condition, whereas an overconsolidated
clay fill would probably hold substantial pore water suction, and so the degree to which
friction is mobilised would start small and slowly increase as pore water pressures increase
towards equilibrium values.

3 IMPLEMENTATION IN TENSARSLOPE
The method described above has been implemented in an experimental version of
TensarSlope, the reinforced soil slope stability analysis program belonging to Tensar
International. Details are given in McCombie (2009). The programming environment used
is Codegear (formally Borland) Delphi; the versatility of data types and object orientation
were particularly useful in programming this method and the SGA. The program already
allowed the definition of complex problems, with soils defined in zones rather than in layers,
for example, and scope for applying surcharge loads, horizontal line loads, soil
reinforcement, complex pore water pressure regimes, and external standing water. However,
the definition of soil strength was altered to allow use of both peak and critical state angles of
friction so that the angle of dilation could be deduced. A control was provided to allow the
user to view the detailed results for each individual wedge once an analysis was completed,
and to export them to a comma separated value file to facilitate checking.
A new form was developed to contain the processes necessary for the genetic algorithm,
including entering values for the bounds of the problem, and the parameters for the
optimisation. One option is produce each generation of chromosomes by a random process,
rather than following the genetic algorithm. This is still far more efficient than using
completely random generation of point co-ordinates, because the rational system described
below is used to generate a surface from each chromosome, but it allowed the optimisation
process itself to be evaluated.
4 SIMPLE GENETIC ALGORITHM
The SGA used here employs binary coding to represent the variables used to generate each
slip surface to be analysed. A simple way to generate a slip surface would be to encode the
co-ordinates of each wedge junction, but most of the surfaces thus produced would not be
feasible. The approach used has therefore been to use the binary chromosome to represent a
series of integers, each of which define where a value lies within a range of feasible values.
The surface is generated from one end to the other, and the range of feasible values takes into
account the geometry of the surface generated thus far. The encoding thus does not define
the co-ordinates directly, but rather the sequence of steps by which they must be generated.
This does result in the number of wedges used being fixed. However, in the method of
analysis adopted, if two wedges have their bases aligned then there will be no relative
movement between them and they will behave as a single wedge. It is therefore possible for
the search to in effect reduce the number of wedges if this optimises the solution.
The following information is therefore used to generate a surface:
- minimum and maximum x co-ordinates for the toe of the slip surface;
- minimum and maximum x co-ordinates for the crest of the slip surface;
- minimum and maximum angles for the slip surface at the toe;
- minimum and maximum angles for the slip surface at the crest.
These values apply to every surface, and would provide all the information required for a
two wedge mechanism. For each individual surface, a single number is used to define where
each actual x co-ordinate or angle lies within its range of possible values. This might be
thought of as a real number between 0 and 1, but is actually encoded as a binary value
between 0 and 255. This reduces the precision of the solution compared with using a floating
point value, but given that the range is restricted to realistic possibilities rather than
encompassing all possible values, this precision is more than sufficient. In real coded
algorithms, crossover would have to operate at data boundaries, and changes in actual
numbers can only occur through mutation. The binary coded approach adopted here has the
advantage of being able to produce substantial changes through splitting an integer encoding
in two at crossover, which far outweighs the potential for more precise definition of numbers
through a large number of random mutations in real coded algorithms.
The positions of the intervening wedge junctions are defined by two values at each
junction, which are each encoded as binary values in the same way. One value is used to
determine the x co-ordinate. For each node, a realistic range of values may be set using
whatever rule is considered appropriate, but in this instance a lower bound has been set at one
fifth of an equal share of the horizontal distance to the end of the surface, and the upper
bound is set at 1.5 times this equal share, up to a maximum of 0.8 times the distance to the
end of the surface.
The other value is used to define the angle of the wedge starting from that point. If the
lower bound for the angle is the angle of the base of the previous wedge, then a concave
surface is guaranteed; this might not be required, in which case a different lower bound may
be defined. Given a requirement for a concave failure surface, an upper bound for the angle
can be defined in terms of the angle to be achieved by the end of the failure surface; if this
were achieved before the last point, then remaining points would have to lie on a straight line,
and the range of permissible angles would be reduced to zero. The angle may also be
constrained by the need to reach the end of the slope within the required range of x co-
ordinates. As the last line on the surface has its position defined already, the penultimate
point may be found by the intersection of this line with the previous line. Checks may be
made that the surface lies beneath the ground at all points along its length, but provided any
inappropriate surface is given a high value of factor of safety, then the optimisation will
succeed.
The integer values thus required are combined into a chromosome on which the genetic
algorithm operates. The conventional way of encoding for the SGA has all the data
represented in a single binary string. This could take the form of examples a) or b) in Fig. 4.

a) sequential, x x x angle angle angle

x’s then angles

b) alternating, x angle x angle x angle

x’s and angles

c) parallel, x’s x x x x x x
and angles
angle angle angle angle angle angle

Fig. 4. Encoding the surface.

Parent 1 x x x x x x x x
angle angle angle angle angle angle angle angle

Parent 2 x x x x x x x x
angle angle angle angle angle angle angle angle

Child 1 x x x x x x x x
angle angle angle angle angle angle angle angle

Child 2 x x x x x x x x
angle angle angle angle angle angle angle angle

Fig. 5. Two point crossover for parallel chromosomes.

In the conventional approach, all the genetic operations take place on single chromosomes,
with each chromosome containing all the information required to define the member of the
population, or failure surface. In case a), a genetic process could result in either the angles
for a part of a surface being swapped, or the x co-ordinates, but never both angles and x co-
ordinates for the same section of slip surface. This constrain the power of the algorithm to
generate new solutions or optimise existing solutions. In case b), this problem is removed,
but there is still a possibility of separating an x-coordinate from its associated angle, and it
seems more appropriate to use two parallel chromosomes, representing each set of data.
These chromosomes are then manipulated in parallel, as if they were fixed together so that
they are cut and exchanged at the same places (figure 5). This method has been used in the
current work, and has the additional advantage of being relatively straightforward to program,
as the two chromosomes are read in parallel as the surface is generated.

5 DEMONSTRATION
The analysis method and the optimisation process have been evaluated using the example of
the Carsington Embankment failure, using parameters from work by Potts et al. (1990), as
described in McCombie (2009). An image from the computer screen is shown in Fig. 6 after
66 generations of a population of 40, which took under two minutes. The entire population of
surfaces is shown, indicating the scope for significantly divergent solutions to be generated
even after a large number of generations.

Fig. 6. SGA analysis of Carsington Embankment failure.

A second screen shot is shown in Fig. 7, following a final stepwise adjustment of the slip
surface position, which gives a slight improvement in the resolution for this example in
which the weak layer is very thin. This figure also allows the final surface to be seen clearly
on its own.
 Fig. 7. Final refinement of the analysis.

The critical mechanism as found by the SGA corresponds closely to the pattern of vectors
from the finite element analysis, as described by McCombie (2009). The resulting factor of
safety of 1.000 found in an analysis carried out for this paper also corresponds with the result
reported by Potts et al. (1990), and the fact that the embankment actually failed on a surface
very close to that described above. Potts et al. (1990) reported that limit equilibrium
methods available at the time could not achieve this result.

6 CONCLUSIONS
The SGA has been shown to work well in finding an accurate optimum solution for a
relatively difficult problem, when used with the multiple wedge method as described by
McCombie (2009). Importantly, both the analytical procedure and the optimisation routine
are easy and quick to use for routine design work.

ACKNOWLEDGEMENT
The author wishes to express his particular thanks to Tensar International Limited for their
permission to use the program WinSlope, later TensarSlope, in the work reported here.

REFERENCES
Bolton, M.D. (1986). "The strength and dilatancy of sands". Géotechnique Vol. 36(1), 65-
78.
Donald, I.B. and Chen, Z. (1999). "Modern Wedge Methods of Slope Analysis – Their Power
and Their Versatility". Proceedings of 2nd Int. Conf. on Landslides, Slope Stability &
The Safety of Infra-Structures, Singapore, 13-24.
Donald, I.B. and Giam, P.S.K. (1989), “Improved comprehensive limit equilibrium stability
analysis”. Department of Civil Engineering Report No. 1/1989, Monash University,
Melbourne, Australia.
Goh, A.T.C. (1999). "Genetic algorithm search for critical slip surface in multiple wedge
stability analysis". Can. Geotech. J., Vol. 36, 382-391.
Goh, A.T.C. (2000). "Search for critical slip circle using genetic algorithms". Civil Eng, and
Env. Syst., Vol. 17, 181-211.
Krahn, J. (2003). "The 2001 R.M. Hardy Lecture: The limits of limit equilibrium analyses".
Can. Geotech. J., Vol. 40, 643-660.
Janbu, N. (1957). "Earth pressures and bearing capacity calculations by generalized
procedure of slices ". Proceedings 4th Conference of ISSMFE, London, Vol. 2, 207-212.
McCombie, P. (2009). "Displacement based multiple wedge slope stability analysis".
Computers and Geotechnics, Vol. 36(1-2). 332-341.
McCombie, P. and Wilkinson, P. (2002). "The use of the simple genetic algorithm in finding
the critical factor of safety in slope stability analysis". Computers and Geotechnics, Vol.
29(8), 699-714.
Morgenstern, N.R., and Price, V.E. (1965). "The analysis of the stability of general slip
surfaces". Géotechnique, Vol. 15(1), 79-93.
Potts, D.M., Dounias, G.T. & Vaughan, P.R. (1990). "Finite element analysis of progressive
failure of Carsington embankment". Géotechnique Vol. 40 (1), 79-101.
Sarma, S.K. (1979). "Stability analysis of embankments and slopes". J. Geotech. Engng Div.
Am. Soc. Civ. Engrs., Vol. 105(GT12), 1511-1524.
Spencer, E. (1967). "A method of analysis of embankments assuming parallel interslice
forces ". Géotechnique, Vol. 17(1), 11-26.
Zolfaghari AR, Heath AC, McCombie PF. (2005). "Simple genetic algorithm search for
critical non-circular failure surface in slope stability analysis". Computers and
Geotechnics, Vol. 32(3), 139-152.