A numerical method for solving the equations of stability of general slip surfaces

By N. R. Morgenstern* and V. E. Pricef

A practical method of solving the equations of statical equilibrium to obtain the factor of safety of earth slopes, with slip surfaces of arbitrary shape, is described. The basic mathematical problem is the solution of a pair of. simultaneous non-linear equations which arise from the boundary conditions on the solution of an ordinary differential equation.

Downloaded from http://comjnl.oxfordjournals.org/ at Pontif-cia Universidade Catlica do Rio de Janeiro on March 21, 2014

1. Introduction

A method has been developed by Morgenstern and Price (1965) for analyzing the stability of earth slopes with regard to slips along general surfaces which can be considered as a sequence of planes. The method is believed to be the first in which all the equations of statical equilibrium are satisfied. It has been used satisfactorily to determine factors of safety for several thousand slip surfaces on more than forty different cross-sections of slopes. A set of equations was derived which determine the factor of safety F, an overall force ratio scaling constant A, and the internal forces. In this paper we describe the numerical technique for solving these equations, which has been used in a program for a digital computer. The equations are equivalent to a pair of non-linear simultaneous equations in the unknowns A and F, and the principles of the method of solving them may be applied to any set of non-linear simultaneous equations, when the first derivatives with respect to the unknowns can be evaluated. Many of the preliminary experiments were performed on the Deuce computer using the Alphacode language, but most of the numerical results have been obtained on an I.C.T. 1905 using a program which was written in a dialect of FORTRAN similar to FORTRAN II. The program is at present being converted to FORTRAN as specified by the American Standards Association. In 2 there is a summary of the equations which are to be solved, and in 3 the principles of the numerical method of solving the equations are described, the detailed formulae being given in 4. It is necessary to evaluate some integrals which involve some slight numerical complexity, and the method for these is described in 5.
2. Summary of the equations

y = y(x)

L V = V t ( l )

Fig. 1.

Potential sliding mass

and shearing forcej respectively in a general vertical plane with coordinate x, which intersects the slip surface at (x, y) and the resultant E at (x, y,) then
dE

(1)
(2) (3)

Zx + L)E) and where L=Xn X= XfE

F

(4)
A tan 4>'

F~ . tan.

(5) (6)

and tan (7)

t The most recent version of the program works in terms of effective E and X forces and also incorporates earthquake loading represented by a horizontal body force expressed as some percentage of gravity. These differences only complicate the coefficients in equation (2) and in no way alter the material presented here.

We consider a cross-section of the slope of unit thickness, and take axes horizontal such that sliding tends to occu: ill the negative x direction and Oy vertically downwa. ds, the origin being arbitrary (see Fig. 1). If E and X denote the resultant total normal force

* Lecturer in Civil Engineering, Imperial College of Science and Technology, Exhibition Road, London, S.W. 1. t Reader in Numerical Analysis, The City University, St. John St., London, E.C.I.
388

Slip surfaces where and

/ = kx + m
dx

(8)

(9)

Hence

SA = DA
F

(16)
IF DF DA

The unknowns in these equations are the constants A and F, and the functions E, X and yt. All the other quantities are known functions of x depending on the chosen slip surface, the geometry of the slope, the pore pressures in the soil mass and the properties of the types of soils from which it is constructed. Equation (3) is an assumption which is made concerning the internal forces which makes the problem statically determinate. A discussion of the choice of the function f(x), which must be specified, was given by Morgenstern and Price (1965). By assuming that the boundaries of the different soil types are linear, that f{x) may be approximated by a sequence of linear functions, and that the slip surface consists of a sequence of planes, then the potential sliding mass may be divided into n slices by vertical planes with coordinates x0, xu . . ., xn such that in any slice K, L, N and P are constants. We do not assume that any of these slices are necessarily thin. The boundary conditions to be satisfied are normally: E == 0 E == 0 and where M == 0 M-when x x0 when x = xn when x = xn (10) (11) (12) (13)

and

SF = DA

"D A
DF DF DA

(17)

Thus starting with estimated or arbitrary values of A and F, by evaluating En, Mn and their derivatives with respect to A and F, new estimates may be found, and so on, until the changes in A and F are sufficiently small. However, if this iterative process is used without additional controls to the values of A and F, then even with reasonable initial approximations the process will in many cases either not converge, ox converge to give values of A and F which are physically unacceptable. There are three additional methods of controlling the values of A and F as follows. (a) If (L + Kx) is zero for any value of x, then it may be seen from equation (2) that E is in general infinite at that point, which is physically unacceptable. This occurs if (L + Kx) changes sign within any slice. In addition, it is reasonable to expect that (L + Kx) will not change sign at any of the points xu x2,. . ., x n _i where it is discontinuous, since any discontinuity in the soil characteristics, or the slope of a line, may be considered as the limit of a rapidly varying continuous quantity. Thus (Z, + Kx) should be of the same sign for the complete range of x. It has been found in practice that this function should be positive, because the contribution of 1 to L in equation (5) is often the dominant term. Thus we define a feasible region for (A, F) such that for all points in this region the value of (Z, + Kx) is positive for all x. By restricting the values of (A, F) to the feasible region, unique solutions have been found. This has been tested in several experiments by starting with different initial values of A and F, and always the solution has been independent of the starting values. In some of the preliminary experiments, when this control was not applied, it was found that there were often several real solutions to the equations, but that not more than one solution was physically feasible. If in any iteration (L + Kx) becomes negative for any value of x, then the current values A[ and Ft are replaced by K*o + A,) and K ^ o + ^ i ) where (A 0 ,F 0 ) is the previous value in the feasible region. It is necessary that the initial values of A and F should be in the feasible region. We have found in practice that a satisfactory starting point in general is F= 1-5 and A = 0-3/fmax. If during the first iteration, it is found that this point is not feasible, then, keeping the value of A fixed, an attempt is made to find a feasible value of F. If this is not possible, then A is set to zero in which case it is always possible to find a feasible F. 389

Downloaded from http://comjnl.oxfordjournals.org/ at Pontif-cia Universidade Catlica do Rio de Janeiro on March 21, 2014

For any values of A and F, equation (2) may be integrated to determine E, starting with E = 0 at x0, then X may be determined from equation (3) and y, from equation (13). Thus the values of E(xn) and M(xn) may be regarded as functions of A and F, and we wish to determine those values of A and F which satisfy equations (11) and (12). 3. Solution of the equations The basis of the technique for solving the equations is the Newton-Raphson method together with certain controls on the values of A and F used. Suppose that En{\, F) and Mn{\, F) denote the values of E and M when x = xn. Starting with values Ao and Fo, suppose the required solution is Ao -f SAand.F0 + 8F. If 8A and 8F are small, then, approximately, En{\0 + 8X, Fo + 8F) = En{\, Fo)+ S A ^ " and + SA, Fo + 8F) = Mn(\0, Fo) + SA-^<^-" = 0 (14)

^F=0

05)

where the derivatives are evaluated at (Ao, Fo).

Slip surfaces

From equations (4), (5) and (8) we have:

. , t a n <b'

XfA.

will be reduced, i.e., a value of v(< 1) can be found such that: (D(A0 + vSA, F o + vSF) < O(A0, Fo). (30)

This is done by comparing the values of O(Ai, Ft) with C>(A0, Fo) and if O has increased then Xt and Fx are replaced by KA0 + A,) and K^o+Fi), respectively, and the process repeated. Fmax>F>Fmin (18) The quantity C is introduced so that the contributions where El and CM% to < ] > are of the same order of magnitude. ((A - A/) tan <f>' E and M are of different dimensions and if, as is usual, 1 + XfA > 0 ^ (19) Fmin = max units of lb and ft are used, then on practical slopes M2 1 +XfA is much greater than E2. In these circumstances, if C and is not introduced, i.e. C = 1, then the number of itera A/) tan tions would be significantly increased as follows. (20) 1 +XfA Normally within a few iterations the point (Ao, Fo) would be close to the curve for which Mn = 0, and Usually it is the lower limit to F which is most critical, although the point (Ao + SA, Fo + 8F) would be closer since for most values of x, 1 + XfA > 0. Hence if the to the final point where both En and Mn = 0, the value upper limit does not exist or if Fmax > Fmin + 0 - 1 , then of M\ would be larger at (Ao + SA, Fo + SF) than at we set: (Ao, Fo) and hence because M\ would be the dominant F=Fm/n+0-1. (21) contribution to O the values of SA and SF would be repeatedly halved until condition (30) is satisfied. At KFmax < Fmin + 0-1, then we set: this stage (Ao + vSA, F o + vSF) is very close to (Ao, Fo) F = i(Fmin + Fmax). (22) and so several iterations would have been made giving only a slight improvement. Thus the accepted values If F mi/I > Fmax, then it is not possible to find a feasible F (A, F) would follow very closely with small steps the with this value of A, so we set A = 0, in which case there curve M = 0 until finally En = 0. is no upper limit to F and we set: Thus we need to use a value of C which is positive F = max {A tan <f>') + 0 1. (23) and with dimensions of E2/M2. When comparing <5(A,, F,) with <E>(A0, Fo) then C is given by (6) A restriction is applied to the magnitude of the steps SA and SF obtained from equations (16) and (17). If either

Hence, in order that L + Kx > 0 for all x, then F must lie in the range

Downloaded from http://comjnl.oxfordjournals.org/ at Pontif-cia Universidade Catlica do Rio de Janeiro on March 21, 2014

|SA| > 0-5 or |SF| > 0 - 5 then the subsequent values A( and F] are given by A, = Ao + vSA and where Ft = Fo + v8F v = 0-5/max (|SA|, |SF|). (24) (25) (26)

In this way very large increments are prevented, which would otherwise occur usually in the early iterations, when the matrix of first derivatives is ill-conditioned. (c) The function O given by O(A, F) = E2n + CM2 (27)

where C is any positive constant, satisfies the inequality

(28)

with the derivatives evaluated at (Ao, Fo). The value of C is kept constant until a point (Ai,F,) is found for which < J > has decreased. When this occurs the value of C and O are re-evaluated using the derivatives at the newly accepted point. With these three additional controls on the values of A and F, in approximately 80% of practical cases the process converges in less than 10 iterations, giving values of A and F correct to 3 decimal places. An upper limit is imposed in the program of 20 iterations of the type so far described. If this limit is reached, then further iterations are performed, restarting with the initial estimates for F and A, in which only F is varied until the equation Mn = 0 is satisfied. This is performed using the usual single variable Newton-Raphson process given by
(32)

where SA and 8F satisfy equations (14) and (15) and the derivatives are evaluated at (Ao, Fo). Hence if (29) <D(A0 + SA, F0 + SF)> <(Ao, Fo), then it is possible to reduce the size of the step so that O
390

If this process converges within 20 iterations, then the value of F obtained and the initial estimate for A is used as the starting point for further iterations of the original type, with a limit again of 20 iterations imposed. The results of a case in which the original 20 iterations did not converge, but which subsequently converged

Slip surfaces

after finding a value of F for which Mn = 0 is given in Table 1. From the table it may be seen that the first four iterations took the point (F, A) near to a local minimum of O at about (1-15, 1-2). The other sixteen iterations of the first batch produced only two more acceptable points, viz. (11513, 1-1922) and (11537, 11883) with lower values of < 1 > . After the first batch had been unsuccessful, the values of F and A were returned to their initial values and F varied to make Mn zero. This converged in four iterations to the point (1-203, 0-3) from which the Newton process applied to both variables converged in a further four iterations to the point (1-134, 0-491). As will be explained in 5, the last two iterations were needed to refine these values of A and F using accurate formulae for the evaluation of M and its derivatives. In a few cases the triple system of iterations described above has not converged to give solutions. Several of these have been investigated in detail and it is almost certain that in these cases a solution in the feasible region does not exist. This conclusion was reached by considering a set of slip surfaces gradually changing from one for which a physically acceptable solution had been found to the required surface for which a solution could not be found. In this way the solution was traced and was found to approach the boundary of the feasible region andfinallyattempted to cross it.
4. The calculation of E, M and their derivatives

Table 1 Values of F, X, En and Mn for successive iterations in case which did not converge within the first batch of 20 iterations F

A -0 3000 02000 07000 12000 07000 09500 10750 11375 11687 11844 11922 0 6922 0 9422 1 0672 1 1297 1 1609 1 1766 1 1844 1 1883 0 6883 -0 -0 -0 -0 3000 3000 3000 3000

En x 10-s

Mn X 10-6

L-5000 1-4009 1-2389

t 1453

1-5293 1-3373 1-2413 1-1933

I1693

I1573 11513 4598 3056 2284 1 1899 1 1706 ] 1610 1 1561 11537 14301 15000 11343 11994 12033 12033 11450 11354 11337 11336 11347

-1-0805 -10311 -0-8328 -0-7833 -1-2486 -10101 -0-8843 -0-8221 -0-7951 -0-7854 -0-7829 -1-1634 -0-9650 -0-8630 -0-8136 -0-7925 -0-7849 -0-7828 -0-7825 -1-1243 -10805 - 0 1903 -0-3844 -0-3954

-3 0052 -3 2138 -2 0886 -2 2306 -7 6868 -5 5238 -3 8812 -2 9413 -2 4928 2 3127 2 2520 -6 4712 -4 7484 -3 4841 -2 7754 2 4429 -2 3115 2 2677 2 2553 9221 -3 0052 0 9161 00492 00002 00000 -1 1350 -0 1339 -0 0022 0 0327 0 0000

Downloaded from http://comjnl.oxfordjournals.org/ at Pontif-cia Universidade Catlica do Rio de Janeiro on March 21, 2014

We wish to obtain the formulae for determining En, -r-p, -z-j-, Mn, -J^TT, and -y~^ for any values of A andf. We will denote by Kn L,, JV, and P, the constant values of K, L, N and P in the slice *,_, < x < x ; for i= 1,2,..., n. If in the /th slice x is measured from x ; _, then the solution of equation (2) is E= Hence L,
r

-0 3000 -0-3954 -0 6238 0-1297 -0- 5083 00140 -0- 4911 0-0002 -0- 4908 -0- 4942 0 0000 0 0000

. a , , - , + P,x + iNjX2}.

(33)

Pfi,
where 6, = x, *,_,.

(34)

(35)

The derivatives of K, L, N and P can easily be obtained from equations (4), (5), (6) and (7). By starting with
E

Differentiating equation (34) with respect to F gives

YF = 3A =

(38)

equations (34), (36) and (37) may be used to determine From equation (13) we have and differentiating with respect to A gives

M, = M,_x + J (X - A,E)dx,
Using equations (3), (8) and (33) we have
IL, 391
x

(39)

_ \{ktx + ntj) Li + K,x

'-'

(40)

SY//> surfaces

hence
A,}

(,,_, + P,x + iN,x2)dx.

(41)

mate values, based on the trapezium rule of integration, can be obtained much more quickly. From equations (39), (3) and (8) we have, approximately, Mj = Mi_x + ibi{(\i(kib, + m,) Ai)Et + (A,w,- Ai)Ei_l}, (48)
ib

Differentiating equation (41) with respect to F and A, ' + (A,m ; and

/)

J>E,

;)
i

^F
- l

~i>F

tF

Jo

L-, +

Ai)^:-

(49)
Downloaded from http://comjnl.oxfordjournals.org/ at Pontif-cia Universidade Catlica do Rio de Janeiro on March 21, 2014

-C
and

IM, (L, + KiX)

2

(}h 4 _ ?',

(42)

(Am,(50)

Jo

~T1v^ {L,E, _ [
Lij +- A,-.X

{AfA'/X + 7,-) /4,}

(L,

u,^-l
(43)

These approximate formulae are used until the iterative process converges, and then the exact formulae are used to refine the values of F and A. Normally, about 6 of the simpler iterations are followed by 1 or 2 of the iterations in which M and its derivatives are evaluated accurately. Thus there is an overall saving of approximately 50% of the computing time by using the approximate formulae first. In order to evaluate Mn and its derivatives accurately, it was shown in 4 that it is necessary to evaluate integrals of the following form: 1 + Hx
dx (51) (52)

By changing the scale of x in these integrals so that the range of integration is from 0 to 1, and letting H; = K,b,IL, then we obtain the following recurrence relations: f1 1
3

and

(44)

9(\+Hxf

" ' = "'-'+J 0 (TH

and

(45) (46)

for values of H > 1 and for; = 0, 1, . . ., 4. For non-zero values of H there are analytical formulae for these integrals which may easily be generated by the following recurrence relations:
Bn = ~TT l g (1 + ") Cn = V53)

'jx'dx,

(47)

1 +H

(54)

where the coefficients 7}, C/y and Fy are independent of x. Since only values of A and F are used for which Ls and Lt + Kfi; are positive then H-, > 1 and all the integrals are convergent. The evaluation of the basic -1 x^dx r x^dx

(55)
r C -y+i / f>

(56)

integrals j Q ^ - ^ and JQ f p p r ^ a will be described

in the next section. 5. The evaluation of the integrals The calculation of M and its derivatives takes approximately \ of the time of each iteration, but their approxi392

When H = 0, it is obvious that these formulae cannot be used. Also, for small values of H, their use introduces serious rounding-off errors since each integral is expressed as the relatively small difference between two large numbers. For small values of H, however, it can be seen that the integrands may be expanded as polynomials in x, and so Gaussian numerical integration formulae may be satisfactorily used.

Slip surfaces

By evaluating the most complicated integral C4, for a sequence of values of H starting with 0-95, and using both the recurrence relations (53),. . ., (56) and Gaussian integration formulae of various orders, it was found that, using floating-point arithmetic with a precision of about 9 decimal places, the integrals could be obtained correct to at least 6 significant decimal places by using the analytical formulae for H < 0-4 and H > 0-5, and for 0-4 < # < 0-5 using the Gauss formula of order 5. 6. Discussion The methods described in this paper have been incorporated in a computer program wh;ch has been used by a number of civil engineers. A description of the practical aspects of the program and of some of the problems associated with it will be reported elsewhere. The method of specifying the data for an analysis is similar to that for the slip-circle program described by Little and Price (1958). It is well known that the Newton-Raphson technique may be used for solving non-linear simultaneous equations, and it is known that extra controls on the variables are often essential to make the process converge. The control (c) described in 3, that the sum of the squares of the residuals should be reduced by each iteration is References

of general application for one or more non-linear equations. It is less arbitrary than the control (b) of 3 which is advocated in N.P.L. (1961) which requires an arbitrary fixed limit to the change, or proportional change, of any of the variables. However, it should be noted that squares of the residuals should be scaled if necessary so that they are of the same order of magnitude, otherwise the restriction can increase the number of iterations considerably. The scaling may conveniently be performed by dividing each residual by the sum of the squares of its derivatives with respect to the unknowns. 7. Acknowledgement The authors gratefully acknowledge the assistance of members of the staff of the English Electric London Computing Service, and in particular, Mrs. J. Skinner, with the programming of the preliminary experiments of these calculations. The authors are grateful to Messrs. Binnie and Partners for permission to publish the results in Table 1, and for many helpful discussions and encouragement with the project. This study was supported in its early stages by a research grant awarded by the Department of Scientific and Industrial Research.

Downloaded from http://comjnl.oxfordjournals.org/ at Pontif-cia Universidade Catlica do Rio de Janeiro on March 21, 2014

A. L., and PRICE, V. E. (1958). "The Use of an Electronic Computer for Stability Analysis," Ceotechnique, Vol. 8, 3 p. 113. MORGENSTERN, N. R., and PRICE, V. E. (1965). "The Analysis of the Stability of General Slip Surfaces," Geotechnique Vol 15 1, p. 79. N.P.L. (1961). Modern Computing Methods, 2nd Edition, chapter 6, 23, London, H.M.S.O.
LITTLE,

Correspondence
To the Editor, The Computer Journal. Sir, Many algorithms have been published in which values of 7T, 277, 7T2 etc., are written as constants, commonly with about 9 significant figures. Sometimes comments are included that these constants are intended as approximations to 77 etc., and that the number of significant figures should be adjusted to the capacity of the computer. But it is not necessary to use any explicit approximation to 77 or related values, since in ALGOL 60 the appropriate value can be generated by the statement pi: = 4 X arctan (1) and the variable pi may then be used as a component of arithmetic expressions. The accuracy with which pi approximates to n is limited only by the accuracy of the arctan function (and possible roundoff). For instance, in FORTRAN II-D on the IBM 1620, the corresponding statement PI = 4 0*ATANF (10) (with 28 decimal digits in the mantissae of floating-point numbers) does indeed give 77 correctly to 28 significant figures. Likewise, of course, if e is required then the statement e : exp (1) can be used. Yours faithfully, G. J. TEE University of Lancaster, 108 St. Leonardgate, Lancaster 24 November 1966

393