Sloan, S. W. (2013). Ge´otechnique 63, No. 7, 531–572 [http://dx.doi.org/10.1680/geot.12.RL.001]
Geotechnical stability analysis
S. W. SLOAN
This paper describes recent advances in stability analysis that combine the limit theorems of classical plasticity with finite elements to give rigorous upper and lower bounds on the failure load. These methods, known as finite-element limit analysis, do not require assumptions to be made about the mode of failure, and use only simple strength parameters that are familiar to geotechnical engineers. The bounding properties of the solutions are invaluable in practice, and enable accurate limit loads to be obtained through the use of an exact error estimate and automatic adaptive meshing procedures. The methods are very general, and can deal with heterogeneous soil profiles, anisotropic strength characteristics, fissured soils, discontinuities, complicated boundary conditions, and complex loading in both two and three dimensions. A new development, which incorporates pore water pressures in finite-element limit analysis, is also described. Following a brief outline of the new techniques, stability solutions are given for several practical problems, including foundations, anchors, slopes, excavations and tunnels.
KEYWORDS: anchors; bearing capacity; excavation; numerical modelling; plasticity; slopes; tunnels
STABILITY ANALYSIS Limit equilibrium In geotechnical engineering, stability analysis is used to Limit equilibrium is the oldest method for performing predict the maximum load that can be supported by a stability analysis, and was first applied in a geotechnical geostructure without inducing failure. This ultimate load, setting by Coulomb (1773). In its most basic form, this which is also known as the limit or collapse load, can be approach presupposes a failure mechanism, and implicitly used to determine the allowable working load by dividing it assumes that the stresses on the failure planes are limited by by a predetermined factor of safety. The precise value of this the traditional strength parameters c and . The chief factor depends on the type of problem, with, for example, advantages of the limit-equilibrium method are its simplicity lower values being appropriate for slopes and higher values and its long history of use, which have resulted in widely being adopted for foundations. Rather than impose a factor available software and extensive collective experience con- of safety on the ultimate load to obtain the allowable work- cerning its reliability. Its main disadvantage, on the other ing load, it is also possible to apply a factor of safety to the hand, is the need to guess the general form of the failure strength parameters prior to performing the stability analysis. surface in advance, with poor choices giving poor estimates In some procedures – finite-element strength reduction of the failure load. In practice, the correct form of the analysis, for example – the actual safety factor on the failure surface is often not intuitively obvious, especially for strength can be found for a given set of applied loads and problems with an irregular geometry, complex loading, or material parameters (which typically comprise the cohesion complicated stratigraphy. There are other shortcomings of and friction angle). the technique, as follows. Once the allowable load is known, the working deform- (a) The resulting stresses do not satisfy equilibrium at every ations are usually determined using some form of settlement point in the domain. analysis. Historically, these deformations have been predicted (b) There is no simple means of checking the accuracy of the using elasticity theory, but they are now often found from a solution. variety of numerical methods, including non-linear finite- (c) It is hard to incorporate anisotropy and inhomogeneity. element analysis. In some cases, particularly those involving (d) It is difficult to generalise the procedure from two to three dense sands, serviceability constraints on the deformations dimensions. may actually control the allowable load rather than the ultimate load-carrying capacity. Despite these limitations, a multitude of limit-equilibrium Broadly speaking, there are four main methods for per- methods have been proposed and implemented, particularly forming geotechnical stability analysis: limit equilibrium, for slope stability analysis. Indeed, early examples of widely limit analysis, slip-line methods, and the displacement finite- used slope stability methods include those of Janbu (1954, element method. In the following, all these techniques will 1973), Bishop (1955), Morgenstern & Price (1965), Spencer be discussed except the slip-line methods. This family of (1967) and Sarma (1973, 1979). A more recent procedure, procedures is omitted, not because they are considered to be described by Donald & Giam (1989b), is also noteworthy, ineffective, but simply because they are not well suited to since it gives a factor of safety that is a strict upper bound the development of general-purpose software which can deal on the true value. with a wide variety of practical problems. The key principles of the limit-equilibrium approach can be illustrated by considering the classical bearing capacity pro- blem for a smooth strip footing, of width B, resting on a deep layer of undrained clay of strength su, as shown in Fig. 1. Manuscript received 14 September 2012; revised manuscript accepted To begin the analysis, the supposition is made that failure 22 January 2013. occurs along a circular surface whose centre lies at some ARC Centre of Excellence for Geotechnical Science and Engineer- point directly above the edge of the footing, as shown in ing, University of Newcastle, NSW, Australia. Fig. 2. In addition, undrained failure along this surface is
531 532 SLOAN Bearing capacityϭϭq ? σ . ∂ ij . p ϭ f σ ελij ∂ σij
B ϭ f()σij 0
Saturated clay ε σ Undrained shear strength ϭ s ij u (a) (b)
Fig. 3. (a) Perfectly plastic material model and (b) associated flow rule Fig. 1. Smooth strip footing on deep layer of undrained clay
0 tions and the yield criterion. For a perfectly plastic material model with an associated flow rule, it can be shown that the θ load supported by a statically admissible stress field is a R lower bound on the true limit load. Although the limit load q for such a material is unique, the optimum stress field is not, and thus it is possible to have a variety of stress fields that furnish the same lower bounds. To illustrate the applica- B tion of lower-bound limit analysis, the smooth rigid footing τ ϭ s u problem shown in Fig. 1 is considered again. The simple stress field shown in Fig. 4, which consists of three distinct zones separated by two vertical stress discontinuities, is σn statically admissible since it satisfies equilibrium, the stress boundary conditions, and the undrained (Tresca) yield criter- Fig. 2. Limit-equilibrium failure surface for strip footing on clay ion 1 3 ¼ 2su everywhere in the domain. Note that each stress discontinuity is statically admissible because the nor- mal and shear stresses are the same on both of its opposing assumed to be governed by the Tresca criterion, with the sides, and that equilibrium is automatically satisfied every- maximum shear strength being fully mobilised at every where in each zone because the unit weight is zero and the point, so that the shear stress is given by ¼ su: stress field is constant. Although the normal and shear Taking moments about the centre of the failure surface, stresses must be continuous across an admissible stress O, the following is obtained. discontinuity, the normal stress on a plane orthogonal to the B discontinuity is permitted to jump. This feature can be ðÞqB 3 ¼ ðÞ2RŁ 3 su 3 R exploited in the construction of stress fields to give useful 2 lower bounds, and is shown in Fig. 4. or Since the stress field supports a vertical principal stress of 4s Ł 1 ¼ 4su in the zone beneath the footing, this defines a lower q ¼ u (1) bound on the bearing capacity of q ¼ 4s : 2 Ł low u sin In contrast to the lower-bound theorem, the upper-bound The lowest value of q, and hence the geometry of the critical theorem requires the computation of a kinematically admis- surface, can be found by setting dq/dŁ ¼ 0. This leads to the sible velocity field that satisfies the velocity boundary condi- simple non-linear equation tan Ł 2Ł ¼ 0, which can be tions and the plastic flow rule. For such a field, an upper solved to yield the critical angle Łc ¼ 66.88. Inserting this bound on the collapse load is obtained by equating the critical angle in equation (1) gives the approximate bearing power expended by the external loads to the power dissi- capacity as pated internally by plastic deformation. Note that although the true limit load from such a calculation is unique, the q ¼ 5:52s u actual failure mechanism is not. This implies that multiple which is approximately 7% above the exact solution mechanisms may give the same limit load, and it is neces- q ¼ (2 + )su derived by Prandtl (1920). sary to seek the mechanism that gives the lowest upper bound. A simple upper-bound mechanism for this strip footing example, shown in Fig. 5, assumes that failure Limit analysis occurs by the rigid-body rotation of a circular segment, with Limit analysis is based on the plastic bounding theorems developed by Drucker et al. (1951, 1952), and assumes qsϭ 4 small deformations, a perfectly plastic material (Fig. 3(a)), low u and an associated flow rule (Fig. 3(b)). The last assumption, which is often termed the normality rule, implies that the p ϭ σ ϭ 4s σ ϭ 0 plastic strain rates _ij are normal to the yield surface, f( ij), σ3 0 1u 3 p º_ º_ so that _ij ¼ @ f =@ ij, where is a non-negative plastic ϭ ϭ σ ϭ 2s multiplier. For this type of plasticity model it is necessary to σ1u2s σ3u2s 1u work with velocities and strain rates, rather than displace- ments and strains, as the latter become undefined at col- lapse. The lower-bound theorem is based on the principle of a Stress discontinuity statically admissible stress field. Such a stress field is defined as one that satisfies equilibrium, the stress boundary condi- Fig. 4. Lower-bound stress field for strip footing on clay GEOTECHNICAL STABILITY ANALYSIS 533 0 RBϭ sinθ procedures will be the focus of attention in this paper, and ϭ RBc /2sin(θ /2) θ inevitably lead to some form of optimisation problem, the R θ/2 θ/2 solution of which defines either a statically admissible stress q upp Rc field or a kinematically admissible velocity field. Finite- element formulations of the limit theorems inherit all the advantages of the finite-element method, and can model B v Rigid complex geometries, layered soils, anisotropy, soil–structure θ/2 Circular failure . interaction, interface effects, discontinuities, complicated ω surface loadings, and a wide variety of boundary conditions. The Rigid zone, zero velocity success of this approach, however, hinges on the develop- ment of formulations and solution algorithms that are robust, Fig. 5. Upper-bound failure mechanism for strip footing on clay efficient and extendable to three dimensions. Moreover, some means of refining the mesh is needed to ensure that the ‘gap’ between the upper- and lower-bound limit loads is all the internal energy being dissipated along the velocity sufficiently small. discontinuity. From the geometry of Fig. 5, the rate of internal energy (i.e. power) dissipation is ð Displacement finite-element analysis _ W int ¼ Pint ¼ ˜ussudL As a result of the rapid evolution of powerful user-friendly software, displacement finite-element analysis is now widely used in geotechnical practice – not only for the prediction ¼ ðÞRø_ 3 su 3 ðÞ2RŁ of deformations, but also for the prediction of stability. This where ø_ is the angular velocity of the segment about point method is very general, and can accommodate advanced 0, and ˜us is the tangential velocity jump across the constitutive models that incorporate non-associated flow, discontinuity. Equating this quantity to the rate of work (i.e. heterogeneity, anisotropy, and work/strain-hardening and power) expended by the external forces softening. In addition, robust procedures are available for _ modelling interface behaviour, soil–structure interaction and W ext ¼ Pext ¼ quppB 3 v large deformations, as well as fully coupled consolidation Ł and dynamics. When it is used to predict stability under ø static loading, displacement finite-element analysis can be ¼ quppB 3 Rc _ sin 2 used in two different modes. B (a) The loads are applied in increments until the deformation ¼ q B 3 ø_ upp 2 response indicates that a state of collapse has been reached. The approximate ultimate load so obtained and substituting for R gives furnishes a safety factor in terms of force, not strength, 4suŁ and requires a complete simulation of the load– q ¼ upp sin2 Ł deformation response (e.g. Sloan, 1979, 1981; Toh & Sloan, 1980; Sloan & Randolph, 1982; De Borst & Setting dq /dŁ ¼ 0 furnishes the critical angle Ł ¼ 66.88, upp c Vermeer, 1984). Unless advanced procedures (such as which in turn gives the lowest (optimal) upper bound for arc-length methods) are used, this approach leads to this mechanism as instability in the calculations at collapse if the problem is : qupp ¼ 5 52su loaded by prescribed forces rather than by prescribed displacements. (b) Successive analyses with reduced strengths are conducted Combining this result with the previous lower-bound esti- until equilibrium can no longer be maintained (e.g. mate, the exact bearing capacity for the footing on the ideal Zienkiewicz et al., 1975; Dawson et al., 1999; Griffiths & material in this analysis must lie within the range Lane, 1999). This approach, known as strength reduction 4su < q < 5:52su analysis, involves monitoring deformations at specified control points in the soil, and gives the safety factor in Although the limit-equilibrium and upper-bound calculations terms of strength – much like the safety factor that is give the same estimate of the bearing capacity for this case, computed in traditional slope stability calculations using their results are generally different for more complex failure limit equilibrium. Since the method relies on non- mechanisms where the limit-equilibrium solution may not be convergence of the finite-element simulations to indicate kinematically admissible. Notwithstanding the limitations failure, considerable care must be exercised to ensure that that stem from the assumption of a simple perfectly plastic the non-convergence is caused by genuine collapse, and material model, the ability of the limit theorems to provide not some other numerical effect. rigorous bounds on the collapse load is one of their great attractions. Indeed, for complex practical problems where the Figure 6 shows the load–deformation response for a smooth, failure load is difficult to estimate by other methods, this is rigid strip footing on clay, computed using the displacement a compelling advantage, and one of the few instances in finite-element program SNAC (Abbo & Sloan, 2000), for a non-linear mechanics where the error in an approximate soil with a rigidity index G/su ¼ 100 and undrained Poisson’s solution can be bounded exactly. ratio u ¼ 0.49. In this example the rigid foundation is Although the limit theorems can be applied in an analy- simulated by the application of uniform vertical displace- tical setting to give useful bounds for simple problems, ments to nodes underneath the footing, and 15-noded (quar- discrete numerical formulations provide a more general tic) triangles are used to ensure that the soil deformations means of harnessing their power. In particular, finite-element are modelled accurately under incompressible conditions limit analysis formulations have evolved rapidly in recent (Sloan, 1979, 1981; Sloan & Randolph, 1982). The finite- years, and are now sufficiently developed for large-scale element program SNAC, developed at The University of practical applications in geotechnical engineering. These Newcastle over the past two decades, employs adaptive 534 SLOAN 6
Prescribed displacement analysis (2ϩ π )s 5 u
Smooth footing
4
u
s
3
Pressure/ ϭ G/ su 100
Smooth 2 Smooth ϭ νu 0·49 ϭ φu 0
1 48 quartic triangles 825 degrees of freedom Smooth 0 0 1234 5 6 7 8 9 10 11 12 13 14 15 100 (footing displacement)/B
Fig. 6. Displacement finite-element analysis of strip footing on clay explicit methods to integrate the stress–strain relations and four-noded quadrilateral). Locking can also occur for dis- load–deformation response to within a specified accuracy, placement finite-element analysis with the Mohr–Coulomb and is thus well suited to collapse predictions (Sloan, 1987; model, which involves dilatational plastic shearing and is Abbo & Sloan, 1996; Sheng & Sloan, 2001; Sloan et al., widely used for drained stability predictions (Sloan, 1981). 2001). For the mesh shown, the displacement finite-element To ensure that an element is suitable for accurate collapse analysis indicates a clear collapse pressure of 5.19su, which load predictions, under both undrained and drained condi- is within 1% of Prandtl’s exact result of (2 + )su: Unlike tions, three different strategies have been proposed. the methods discussed previously, stability analysis with the displacement finite-element method requires not only the (a) The use of ‘reduced’ integration in forming the element conventional strength parameters, but also the deformation stiffness matrices (e.g. Zienkiewicz et al., 1975; Zienkie- parameters (Poisson’s ratio and shear modulus in this case). wicz, 1979; Griffiths, 1982). This approach, which has Displacement finite-element analysis computes the form of been widely used with the quadratic eight-noded quad- the failure mechanism automatically, and can model a variety rilateral, reduces the number of constraints on the nodal of complicated loadings and boundary conditions. The degrees of freedom at collapse, and introduces additional method is not for the naı¨ve user, however, and even with the ‘flexibility’ into the displacement field by approximate advent of sophisticated geotechnical software considerable numerical integration of the element stiffness matrices. In care and experience are required to use the procedure with general, the method gives good estimates of the collapse confidence in geotechnical practice (Potts, 2003). Since a load, but may generate unrealistic deformation patterns displacement finite-element solution satisfies equilibrium and for some problems (Sloan, 1983; Sloan & Randolph, the flow rule only in a ‘weak’ sense over the domain, the 1983). Selective integration methods, which under- quality of the resulting collapse load prediction is often integrate the volumetric stiffness terms while fully critically dependent on the mesh adopted. Sensitivity studies, integrating the deviatoric stiffness terms, may also be using successively finer meshes, are generally advisable to used in some cases to alleviate the problem of locking for confirm the accuracy of the computed limit load, since no low-order elements (Malkus & Hughes, 1978). reliable error estimate is available for the elasto-plastic (b) The use of high-order triangular elements, with full models commonly used in geotechnical analysis. In addition, integration of the stiffness matrices. This approach, first the accuracy of the limit load can be affected by the number advocated by the author (Sloan, 1979, 1981; Sloan & of load steps used in the analysis (Sloan, 1981; Abbo & Randolph, 1982), follows from the observation that, as Sloan, 1996; Sheng & Sloan, 2001), the numerical integra- meshes of high-order triangles are refined, the new tion scheme used to evaluate the elasto-plastic stresses (Potts degrees of freedom are added at a faster rate than the & Gens, 1985; Sloan, 1987; Sloan et al., 2001), the toler- nodal constraints imposed by the incompressibility ances used to check convergence of the global equilibrium condition, thus avoiding the problem of locking. Since iterations, and the type of element employed (Nagtegaal et these elements use full integration, no difficulties are al., 1974; Sloan, 1979, 1981; Toh & Sloan, 1980; Sloan & encountered with spurious deformation patterns. Randolph, 1982). Of these factors, the correct choice of Although a variety of triangular elements can be shown element is particularly crucial for stability analysis, since the to be suitable for geotechnical stability analysis, the 15- incompressibility constraint imposed by undrained analysis noded triangle, with a quartic displacement expansion, may lead to ‘locking’ where the load–deformation response gives good collapse load predictions under both plane- rises continuously with increasing deformation, regardless of strain and axisymmetric loading. This element is also the mesh discretisation adopted. This phenomenon is due to highly effective for drained stability applications invol- constraints on the nodal displacements, generated by the ving dilatational plasticity models, and can be imple- incompressibility condition, multiplying at a faster rate than mented to give efficient run times (Sloan, 1979, 1981; the degrees of freedom as the mesh is refined, and it is Sloan & Randolph, 1982). For plane-strain deformation, especially pronounced for axisymmetric loading with low- which generates fewer constraints than axisymmetric order elements (such as the linear three-noded triangle and deformation, the six-noded quadratic triangle with full GEOTECHNICAL STABILITY ANALYSIS 535 integration is a viable alternative to the 15-noded triangle, of its development will be given. This review serves to and gives reliable estimates of the collapse load. highlight some of the advantages and drawbacks of the (c) The use of mixed pressure–displacement formulations. approach, as well as its application to practical examples. To avoid numerical oscillations in the solutions, these elements traditionally use a pressure expansion that is one order lower than the displacement expansion (e.g. a six- Historical development of finite-element lower-bound analysis noded quadratic displacement triangle with a linear Lysmer (1970) was an early pioneer in applying finite pressure variation interpolated at the corner nodes), but elements and optimisation theory to compute rigorous lower they can also be used in a ‘stabilised’ form where the bounds for plane-strain geotechnical problems. Lysmer’s pressure and displacement expansions are of equal order formulation was based on a linear three-noded triangle, with (Pastor et al., 1997, 1999). Although they appear to give the unknowns being the normal stresses at the end of each good results, these formulations are more complicated side, plus another ‘internal’ normal stress, and he employed than the previous two options, and have not been widely linear programming to solve the resulting optimisation pro- adopted for geotechnical stability analysis. blem. To satisfy the Mohr–Coulomb yield function in its native form, the Cartesian stresses at every point in an In geotechnical applications, undrained and drained stability element must satisfy a non-linear (quadratic) inequality analyses can be performed as limiting cases of fully coupled constraint. To avoid this type of constraint, and thus generate Biot consolidation, with the former case corresponding to a a linear programming problem, Lysmer (1970) linearised the very fast loading rate and the latter case corresponding to a yield surface using an internal polyhedral approximation that very slow loading rate. Interestingly, when using this ap- replaced each non-linear yield inequality constraint by a proach for stability calculations with a Mohr–Coulomb yield series of linear inequalities. The accuracy of the resulting criterion, a non-associated flow rule with a zero (or small) linearisation can be controlled by varying the number of dilation angle should be used to obtain realistic estimates of sides in the polyhedral approximation, with the highest the collapse load (Small et al., 1976; Small, 1977; Sloan & accuracy being obtained at the cost of additional constraints Abbo, 1999). If a finite dilation angle is adopted, the load– and increased solution times. Because the stress field inside deformation response will display a hardening characteristic each element is assumed to vary linearly, it is sufficient to and fail to asymptote towards a clear collapse state. impose these inequalities at each node to ensure that the linearised yield condition is satisfied throughout the domain, thereby satisfying a key condition of the lower-bound theo- Comparison of methods for stability analysis rem. In addition to the triangular elements used for model- Table 1 summarises the key features of the limit equili- ling the soil, Lysmer’s formulation also included statically brium, limit analysis and displacement finite-element ap- admissible stress discontinuities along the edges between proaches for assessing geotechnical stability. Clearly, the adjacent elements. These greatly enhance the accuracy of a limit-equilibrium method has shortcomings, some of which finite-element lower-bound formulation, especially when sin- will be explored further in a later section of this paper, while gularities are present in the stress field (such as at the edge the displacement finite-element method is the most general. of a rigid footing), and feature prominently in most subse- Conventional limit analysis has the intrinsic advantage of quent implementations of the method. Application of the providing solutions that bound the collapse load from above element equilibrium equations, the discontinuity equilibrium and below, but it is restricted to the use of simple soil models equations and the stress boundary conditions leads to a set and is often difficult to apply in practice. The results in Table of equality constraints on the unknown stresses, while, as 1 suggest that finite-element limit analysis, which combines described above, the linearised yield criterion generates a the generality of the finite-element approach with the rigour large set of linear inequality constraints. The objective func- of limit analysis, is an appealing alternative to traditional tion, which corresponds to the collapse load, is a linear stability prediction techniques. The potential of this type of function of the stresses. After assembling all the element method will be explored fully in this paper, with a particular and nodal contributions for the mesh, the collapse load, focus on its practical utility and scope for future development. denoted by the quantity cT , is maximised by solving a linear programming problem of the form Maximise cT collapse load FINITE-ELEMENT LIMIT ANALYSIS subject to A ¼ b continuum and discontinuity The theory of finite-element limit analysis is quite differ- 1 1 equilibrium, stress boundary ent from that of displacement-based finite-element analysis, even though both methods are rooted in the concept of a conditions A < b linearised yield conditions discrete formulation. Before discussing the fundamental de- 2 2 tails of finite-element limit analysis, a brief historical review (2)
Table 1. Properties of traditional methods used for geotechnical stability analysis
Property Limit Upper-bound limit Lower-bound limit Displacement finite-element equilibrium analysis analysis analysis
Assumed failure mechanism? Yes Yes – No Equilibrium satisfied everywhere? No (globally) – Yes No (nodes only) Flow rule satisfied everywhere? No Yes – No (integration points only) Complex loading and boundary conditions No Yes Yes Yes possible? Complex soil models possible? No No No Yes Coupled analysis possible? No No No Yes Error estimate? No Yes (with lower bound) Yes (with upper bound) No 536 SLOAN where c, b1 and b2 are vectors of constants; A1 and A2 are present no special difficulties, other than adding geometrical matrices of constants; and is a global vector of unknown complexity and increasing the number of unknowns. An normal and ‘internal’ stresses acting on the element edges. early discrete lower-bound formulation based on non-linear Although Lysmer’s finite-element approach for computing programming was described in Belytschko & Hodge (1970). lower bounds was a pivotal conceptual advance, it has three This procedure used piecewise-quadratic equilibrium stress significant limitations that prevented it from being used fields, and maximised the collapse load, subject to the non- widely in practice. The first of these stems from the choice linear yield constraints, by means of a sequential uncon- of variables used in the formulation, which leads to a poorly strained minimisation technique. Although it furnishes rigor- conditioned system of constraint equations that is highly ous lower bounds, the method proved to be slow for large- sensitive to the shape of the elements in the mesh. The scale problems. In a subsequent modification of Lysmer’s second shortcoming of the method is its computational formulation, Basudhar et al. (1979) incorporated the non- inefficiency, which follows from the use of the simplex linear yield constraints directly, converted the constrained algorithm to solve the linear programming problem defined optimisation problem to an unconstrained one using the by equation (2). Since the iterations required by this algo- extended penalty method of Kavlie & Moe (1971), and rithm grow rapidly with the size of the optimisation problem computed the optimal solution (best lower bound) using a being tackled, the number of elements that can be used in a variant of the sequential unconstrained minimisation tech- mesh is severely restricted. The third limitation of the nique (Powell, 1964). Following this work, Arai & Tagyo formulation is that it does not include a strategy for ‘extend- (1985) used constant-stress elements, and the sequential ing’ the stress field over a semi-infinite domain so that the unconstrained minimisation technique with the conjugate equilibrium, stress boundary and yield conditions are satis- gradient algorithm of Fletcher & Reeves (1964), to obtain a fied everywhere. This process, also known as ‘completing’ statically admissible stress field for geotechnical problems. the stress field, is necessary for the solution to be classed as Although both these non-linear formulations require only a a rigorous lower bound. modest number of inequality constraints to ensure that the Following Lysmer’s seminal work, Anderheggen & Kno¨p- stresses satisfy the yield criterion, they still proved unsuita- fel (1972), Pastor (1978) and Bottero et al. (1980) proposed ble for large-scale geotechnical problems, owing to the various discrete methods for two-dimensional lower-bound computational inefficiency of the methods employed to solve limit analysis that were all based on linear triangles and the corresponding optimisation problem. linear programming. These procedures introduced a number Lyamin (1999) and Lyamin & Sloan (2002a) dramatically of key improvements, including the use of Cartesian stresses improved the practical utility of the discrete lower-bound as problem variables to simplify the formulation, and the method by employing linear stress elements, imposing the development of special extension elements for generating non-linear yield conditions in their native form, and solving complete solutions in semi-infinite media. Soon after, Pastor the resulting non-linear optimisation problem using a variant & Turgeman (1982) proposed a lower-bound technique for of an algorithm developed for mixed limit analysis formula- modelling the important case of axisymmetric loading. tions (Zouain et al., 1993). After assembling all the element Although potentially powerful, these early methods were and nodal contributions, the load carried by the unknown T T limited by the computational performance of the linear stresses and body forces, denoted by c1 and c2 h respec- programming codes at the time, and could solve only rel- tively, is maximised by solving the following non-linear atively small problems. Indeed, the practical utility of dis- programming problem crete limit analysis techniques has been strongly linked to Maximise the development of efficient algorithms for solving the T T associated optimisation problems. These problems have spe- c1 þ c2 h collapse load or body force cial features, including extremely sparse and unsymmetric constraint equations, which must be exploited fully in order subject to to solve large cases efficiently. In an effort to address this A11 þ A12h ¼ b1 continuum equilibrium issue, Sloan (1988a, 1988b) proposed a fast linear program- A ¼ b discontinuity equilibrium, ming formulation that can solve small- to medium-scale 2 2 two-dimensional problems on a standard desktop machine. stress boundary conditions This procedure is based on a novel active set algorithm, i which employs a steepest-edge search in the optimisation f ( ) < 0 yield conditions for each node i iterations, and fully exploits the highly sparse nature of the (3) lower-bound constraint matrix. The method has been used successfully to predict the stability of a wide variety of two- where c1, c2, b1 and b2 are vectors of constants; A11, A12 dimensional problems, including tunnels (Assadi & Sloan, and A2 are matrices of constants; f is the non-linear yield 1991; Sloan & Assadi, 1991, 1992), slopes (Yu et al., 1998), criterion; i is a local vector of Cartesian stresses at node i; foundations (Ukritchon et al., 1998; Merifield et al., 1999), is a global vector of unknown Cartesian stresses; and h is anchors (Merifield et al., 2001, 2006a), braced excavations a global vector of unknown body forces acting on each (Ukritchon et al., 2003), and longwall mine workings (Sloan element. Including the body forces in the formulation per- & Assadi, 1994). mits stability numbers based on the unit weight to be Although lower-bound methods based on linear program- optimised, and is especially useful in predicting the load ming are capable of providing useful solutions for two- capacity of slopes, tunnels and excavations. The solution dimensional problems of moderate size, they are poorly method used by Lyamin & Sloan (2002a) is an interior suited to three-dimensional analysis, as huge numbers of point, two-stage, quasi-Newton scheme that exploits the inequalities arise when the yield criterion is linearised. underlying structure of the lower-bound optimisation pro- Moreover, it is not always clear how to linearise a three- blem. Since its iteration count is largely independent of the dimensional yield surface in an optimal manner. Both of grid refinement for a given problem, the method is able to these issues can be avoided by leaving the yield constraints handle large-scale two-dimensional meshes with several in their native form and adopting non-linear programming thousand elements in a few seconds, and is many times algorithms to solve the resulting optimisation problem. In- faster than traditional linear programming formulations. The deed, with this approach, three-dimensional formulations detailed timing comparisons presented by Lyamin & Sloan GEOTECHNICAL STABILITY ANALYSIS 537 (2002a) suggest that, compared with the linear programming linear velocity field using three-noded triangles, with each approach of Sloan (1988a), their technique typically gives at node having two unknown velocities, and each element least a 50-fold reduction in CPU time for large two-dimen- being associated with a fixed number of unknown plastic sional problems. Further advantages include the ability to multiplier rates. To ensure the solution is kinematically model three-dimensional problems, where the number of admissible, the velocities and plastic multiplier rates must unknowns can be huge, as well as any type of convex yield satisfy a set of linear constraints arising from the flow rule, criterion. Thanks to its efficiency and robustness, the lower- with the former unknowns also being subject to the appro- bound method of Lyamin & Sloan (2002a) has been used to priate boundary conditions. For a given set of prescribed predict the stability of a wide range of geotechnical pro- velocities, the finite-element formulation optimises the velo- blems, including tunnels (Lyamin & Sloan, 2000), sinkholes cities and plastic multiplier rates to minimise the power and cavities (Augarde et al., 2003a, 2003b), two- and three- dissipated internally minus the rate of work done by fixed dimensional foundations on clay and/or sand (Shiau et al., external forces. Once this quantity is known, it can be 2003; Hjiaj et al., 2004, 2005; Salgado et al., 2004), anchors equated to the power expended by the external loads to in clay or sand (Merifield et al., 2003, 2005, 2006a), furnish a strict upper bound on the true limit load. To foundations on rock (Merifield et al., 2006b), and slopes in generate a linear programming problem with an upper-bound soil or rock (Li et al., 2008, 2009a, 2009b, 2010). Following finite-element formulation, it is again necessary to linearise the work of Lyamin & Sloan (2002a), Krabbenhøft & the yield criterion. The polyhedral approximation must be Damkilde (2003) proposed another efficient lower-bound external to the parent yield surface to ensure a rigorous method, aimed primarily at solving structural engineering upper bound, and each face of the linearised surface is problems, based on non-linear programming. associated with a single plastic multiplier. After assembling Owing to the presence of singularities in their yield all the element and nodal contributions, the power dissipa- surfaces, where the gradients with respect to the stresses tion in the triangles and the discontinuities, denoted by the T Tº_ become undefined, the Tresca and Mohr–Coulomb criteria quantities c1 u and c2 , minus the rate of work done by any T pose special difficulties in finite-element limit analysis. fixed external forces, denoted by c3 u, is minimised by Lyamin & Sloan (2002a) overcame this difficulty by local solving a linear programming problem of the form smoothing of the yield surface vertices, with an accompany- Minimise ing modification to the search direction to preserve feasibil- T Tº_ T ity during the optimisation iterations. An attractive c1 u þ c2 c3 u power dissipation minus rate of alternative method for solving lower-bound limit analysis work done by fixed external forces problems, which does not require differentiability of the yield surface in the optimisation process, is to use second- subject to _ order cone programming (Ciria, 2004; Makrodimopoulos & A11u þ A12º ¼ 0 continuum flow rule Martin, 2006). This solution method can be applied to a variety of yield criteria in two dimensions, including the A2u ¼ 0 discontinuity flow rule Tresca and Mohr–Coulomb models, and has proved to be A3u ¼ b3 velocity boundary conditions robust and efficient for large-scale geotechnical problems A u < 0 discontinuity signs (Krabbenhøft et al., 2007). In three-dimensional cases, sec- 4 ond-order cone programming can be used for Von Mises and º_ > 0 plastic multiplier Drucker–Prager yield criteria, but not for Tresca or Mohr– Coulomb models. For the latter, which are of particular where c1, c2, c3 and b3 are vectors of constants; A11, A12, interest in geotechnical applications, it is possible to use a A2, A3 and A4 are matrices of constants; u is a global different cone-based solution algorithm that is known as vector of nodal velocities; and º_ is a global vector of semi-definite programming (Krabbenhøft et al., 2008). Like element plastic multipliers. the second-order cone programming method, this approach Following these early procedures that focused on plane does not require smoothing of any yield surface vertices, problems, Turgeman & Pastor (1982) extended the upper- and it has proved to be both robust and efficient for large- bound formulation of Bottero et al. (1980) to handle scale applications (Krabbenhøft et al., 2008). In summary, axisymmetric geometries, but only for Von Mises and Tresca the second-order cone programming and semi-definite pro- materials. gramming methods are, respectively, the solution methods of While the above upper-bound methods inherit all the key choice for the Tresca/Mohr–Coulomb models under two- advantages of the finite-element technique, and hence can and three-dimensional conditions. For yield criteria that are model complex problems in two dimensions, they were not curved in the meridional plane, however, such as the Hoek– widely applied in practice because of the CPU time required Brown model for rock, these procedures are inapplicable, to solve their associated linear programming problems. In an and the more general interior point solution algorithm effort to rectify this handicap, Sloan (1989) proposed an proposed by Lyamin & Sloan (2002a) is appropriate. upper-bound method based on the steepest-edge active set solution scheme (Sloan, 1988b), which had proved successful for lower-bound limit analysis. Although it still suffered from Historical development of finite-element upper-bound analysis the shortcoming of having to specify the direction of shearing Early discrete formulations of the upper-bound theorem, along the velocity discontinuities a priori, the resulting meth- based on finite elements and linear programming, were od was subsequently used to generate useful upper bounds for proposed by Anderheggen & Kno¨pfel (1972) and Maier et a variety of underground structures including trapdoors (Sloan al. (1972). Although quite general, these methods were et al., 1990) and tunnels (Assadi & Sloan, 1991; Sloan & concerned primarily with structural applications. The subse- Assadi, 1991, 1992, 1994). Owing to the nature of the quent plane-strain procedures of Pastor & Turgeman (1976) algorithm used to solve the associated linear optimisation and Bottero et al. (1980), which focused on geotechnical problem, however, the procedure proved to be inefficient for applications with Tresca and Mohr–Coulomb yield criteria, large-scale examples involving thousands of elements. permit a limited number of velocity discontinuities to occur Most early discrete formulations of the upper-bound theo- between elements, but require the direction of shearing to be rem employed the three-noded triangle with a linearised yield specified a priori. These formulations assume a piecewise function, since this leads to an optimisation problem with 538 SLOAN linear constraints where the power dissipation can be ex- (1965). Hodge & Belytschko (1968) reported slow conver- pressed solely in terms of the element plastic multipliers. By gence of the procedure, owing to the complex nature of the using an element with a constant-strain field, it is sufficient objective function. Following this initial work, various other to enforce the flow rule over each triangle to define a non-linear programming formulations were proposed for com- kinematically admissible velocity field. Additional flow rule puting upper bounds on the load capacity of plates, shells and constraints, of course, are needed to define kinematically structures (Biron & Charleux, 1972; Nguyen et al., 1978). admissible velocity jumps across each discontinuity. If dis- Huh & Yang (1991) developed a general upper-bound continuities are not included in a mesh of three-noded procedure for plane stress problems using triangular elements triangles, the elements should be arranged so that four with a linear velocity field. Their method focused on a so- triangles form a quadrilateral, with the central node lying at called ‘ -norm’ family of yield criteria, which includes the the intersection of the diagonals. Failing to observe this rule von Mises model as a special case, and the results suggest for undrained (incompressible) problems may lead to ‘lock- that it is accurate and efficient for relatively large two- ing’, where the elements cannot provide enough degrees of dimensional problems. In a further development, Capsoni & freedom to satisfy the constant-volume condition (Nagtegaal Corradi (1997) proposed another discrete upper-bound ap- et al., 1974). In response to this shortcoming, Yu et al. proach where the straining modes are modelled indepen- (1994) developed a six-noded linear strain triangle for upper- dently of rigid-body motions. This allows finite elements that bound limit analysis. This element can model a velocity field are not involved in the collapse mechanism to be omitted accurately with fewer elements than the constant-strain trian- from the dissipated power summation, and avoids problems gle and, in the absence of discontinuities, no special grid with non-differentiability of the upper-bound functional. arrangement is required for incompressible deformation. In a different non-linear approach, Jiang (1994) proposed The need to specify both the location and the direction of an upper-bound formulation, based on a regularised model shearing for each discontinuity in an upper-bound analysis is of limit analysis (Friaaˆ, 1979), which assumes the material a significant drawback, since it requires a good guess of the is visco-plastic, and uses two parameters to characterise its likely collapse mechanism in advance. This shortcoming was creep behaviour. By fixing the first of these parameters to addressed by Sloan & Kleeman (1995), who generalised the unity, and letting the second one tend to infinity, it can be upper-bound formulation of Sloan (1989) to include velocity shown that the visco-plastic power dissipation converges to discontinuities at all edges shared by adjacent triangles. In the plastic power dissipation, and a rigorous upper bound is their formulation, the direction of shearing is found as part obtained. Although this is an indirect method, the visco- of the optimisation process, and discontinuities are either plastic functional is always convex, even for three-dimen- active or inactive, depending on which deformation pattern sional Mohr–Coulomb and von Mises yield criteria, and gives the least amount of dissipated power. Each discontinu- there is always a unique solution that minimises it. To solve ity is defined by four nodes, and requires four additional the resulting non-linear optimisation problem, Jiang (1994) plastic multipliers to describe the normal and tangential employed the augmented Lagrangian method in conjunction velocity jumps along its length. The upper-bound procedure with the algorithm of Uzawa (Fortin & Glowinski, 1983). In of Sloan & Kleeman (1995) assumes a linearised yield a later paper, Jiang (1995) established that the same non- criterion, and gives rise to a linear programming problem linear programming scheme can be applied to perform that can be solved using the active set solution algorithm of upper-bound limit analysis directly. Jiang’s formulations per- Sloan (1988b). It has proved to be computationally efficient form well for a variety of two-dimensional examples, but for small- to medium-scale problems in two dimensions, have not been extended to deal with discontinuities in the and, because of the presence of velocity discontinuities at all velocity field or three-dimensional geometries. Parallel to shared element edges, gives good estimates of the limit load this development, Liu et al. (1995) proposed a direct itera- without the need for special grid arrangements. Examples tive method for performing three-dimensional upper-bound where the method has provided useful upper bounds include limit analysis. This scheme treats the rigid zones separately slopes (Yu et al., 1998), foundations (Ukritchon et al., 1998; from the plastic zones during each iteration, and neatly Merifield et al., 1999), anchors (Merifield et al., 2001, avoids the numerical difficulties that stem from a non- 2006a), and braced excavations (Ukritchon et al., 2003). differentiable objective function in the former. Their paper It is not straightforward to develop discontinuous upper- suggests that the process is efficient and numerically stable, bound formulations that can model an arbitrary yield condi- and can be implemented easily in an existing displacement tion. This is because the internal power dissipation depends finite-element code. It has not, however, been widely used to on the state of stress as well as on the strain rates, so that in generate rigorous upper bounds for geotechnical problems. addition to finding the velocity and plastic multiplier fields Following in the footsteps of their successful lower-bound that satisfy the flow rule, it is also necessary to compute a formulation, Lyamin & Sloan (2002b) developed an upper- stress field that satisfies the yield criterion. Moreover, kine- bound finite-element method that was also based on matically admissible discontinuities are difficult to incorpo- non-linear programming. This procedure assumes that the rate at all inter-element edges in three dimensions. velocities vary linearly over each element, and that each The plate formulation described by Hodge & Belytschko element is associated with a constant-stress field and a single (1968) was one of the first attempts to develop a finite- plastic multiplier rate. Flow rule constraints are imposed on element upper-bound method based on non-linear program- the nodal velocities, element plastic multipliers and element ming. Their analysis used classical theory to specify the stresses to ensure that the solution is kinematically admis- deformation field solely by the velocity normal to the original sible. In addition, to satisfy the consistency condition, the middle surface of the plate. This normal velocity was element stresses are constrained to obey the yield criterion, approximated within each element by a second-order poly- and the plastic multipliers are constrained to be non-negative. nomial, composed of independent nodal parameters, and the Using the approach developed in Sloan & Kleeman (1995), total internal power included contributions from plastic defor- the formulation of Lyamin & Sloan (2002b) allows velocity mation through the elements, across hinge lines between discontinuities along shared element edges, with the velocity elements and along clamped boundaries. The resulting uncon- jumps across each discontinuity being defined by additional strained optimisation problem requires the ratio of the inter- non-negative unknowns (plastic multipliers). Their procedure nal and external energy dissipation rates to be minimised, and appears to be the first rigorous upper-bound method that was solved using the simplex method of Nelder & Mead incorporates both continuum and discontinuity deformation GEOTECHNICAL STABILITY ANALYSIS 539 in two and three dimensions. Although the yield behaviour in tion of Lyamin & Sloan (2002a), with some important the discontinuities is restricted to models with a linear yield modifications to handle stress discontinuities, and highlights envelope (e.g. Tresca and Mohr–Coulomb), it is otherwise the fundamental differences between finite-element limit quite general. The resulting optimisation problem can be analysis and displacement finite-element analysis. It also solved in terms of the nodal velocities and element stresses illustrates the power of using a discrete formulation of the alone by applying a two-stage, quasi-Newton algorithm classical limit theorems. directly to the Kuhn–Tucker optimality conditions (Lyamin Figure 7 shows a soil mass, with volume V and boundary & Sloan, 2002b). Consequently, the element plastic multi- area A, subject to a set of fixed surface stresses (tractions) t pliers do not need to be included explicitly as variables. This acting on the boundary At, as well as an unknown set of formulation has been used to compute accurate upper bounds tractions q acting on the boundary Aq: In practice t might for a wide range of important geotechnical problems, includ- correspond, for example, to a prescribed surcharge while q ing sinkholes (Augarde et al., 2003a), tunnels (Lyamin & might correspond to an unknown bearing capacity. Also Sloan, 2000), mines (Augarde et al., 2003b), foundation shown in Fig. 7 is a system of fixed body forces g and bearing capacity (Shiau et al., 2003; Hjiaj et al., 2004, 2005; unknown body forces h acting over the volume V. The Salgado et al., 2004) and anchors (Merifield et al., 2003). former is typically a prescribed unit weight, while the latter, Krabbenhøft et al. (2005) modified the upper-bound for- which corresponds to an unknown body force capacity, will mulation of Lyamin & Sloan (2002b) by proposing a new be shown to be very useful in computing the stability of stress-based method that uses patches of continuum elements slopes, tunnels and excavations. to incorporate velocity discontinuities in two and three Recalling the problem solved earlier in the section ‘Limit dimensions. The elements in these patches have zero thick- analysis’, a lower-bound calculation seeks to find a statically T ness, with opposing nodal pairs having the same coordinates, admissible stress field ¼ fgxx, yy, zz, xy, yz, xz and the scheme results in a simple and efficient structure for that satisfies equilibrium throughout V, balances the pre- programming. Moreover, the procedure can accommodate scribed tractions t on At, nowhere violates the yield criterion yield criteria that have curved envelopes, such as the Hoek– f so that f( ) < 0, and maximises the collapse load Brown model. Interestingly, the same idea can also be used ð ð to incorporate stress discontinuities in discrete formulations Q ¼ Q1ðÞq dA þ Q2ðÞh dV of lower-bound limit analysis (Lyamin et al., 2005a). A V To avoid the problems associated with non-smooth yield q surfaces in the optimisation process, second-order cone programming can be used to solve discrete formulations of In the above, the functions Q and Q depend on the case at the upper-bound theorem (Ciria, 2004; Makrodimopoulos & 1 2 hand. For example, in a bearing capacity problem, Q2 ¼ 0, Martin, 2007). This class of solution scheme is highly effec- and one typically wants to maximise the load carried by the tive for lower-bound formulations, as discussed previously, tractions normal to a boundary edge, q , so that Q ¼ q and and has proved to be equally effective for upper-bound n 1 n ð formulations. Indeed, second-order cone programming is the method of choice for solving the optimisation problems that Q ¼ qndA (4) arise from finite-element upper-bound formulations, provided Aq the yield function can be expressed in a conic quadratic form (such as the von Mises/Drucker–Prager model in plane strain or three dimensions, and the Tresca/Mohr–Coulomb For slope, tunnel and excavation problems, on the other model in plane strain). For the Tresca/Mohr–Coulomb mod- hand, one often wants to maximise a dimensionless stability ª el in three dimensions, the resulting upper-bound optimisa- number that is a function of the soil unit weight . In this ª tion problems can be solved efficiently using semi-definite case Q1 ¼ 0 and Q2 ¼ , giving programming, just as in the lower-bound case. ð Q ¼ ª dV V FINITE-ELEMENT LOWER-BOUND FORMULATION An efficient formulation of the lower-bound method will now be briefly described. This section follows the formula- where ª is a variable which can be optimised.
w q σyy
A w h τyx Aq τyz τxy
τzy σxx V τ At xz g τzx t σzz
y
u ghϩ x y yy z
ϩ ghxx ux ϩ uz ghzz
Fig. 7. Surface and body forces acting on soil mass 540 SLOAN Lower-bound finite elements equation (4), can be expressed in terms of the vertical nodal Following Lyamin & Sloan (2002a), linear elements are stresses at the nodes underneath the footing. used to discretise the domain. These elements, shown in Fig. The key steps in formulating an efficient lower-bound 8, enable a statically admissible stress field to be found in a method using finite elements are now outlined for the two- rigorous manner, and have proved to be highly effective in dimensional case. A similar approach holds for three dimen- large-scale applications. sions. The lower bound is found by formulating and solving a non-linear optimisation problem, where the nodal stresses and/or element body forces are the unknowns, and the Objective function objective function to be maximised corresponds to the In many applications, such as bearing capacity calcula- collapse load. The unknowns are subject to equilibrium tions, the objective function to be maximised corresponds to equality constraints for each continuum element, equilibrium a force acting along the boundary of the domain (the equality constraints for each discontinuity, stress boundary collapse load). The common case of optimising the external conditions, and a yield condition inequality constraint for traction q along a boundary segment in two dimensions is each node. shown in Fig. 10. Figure 9 shows a very simple lower-bound mesh for the Since the stresses vary linearly throughout an element, the strip footing problem considered previously in Fig. 1. In this normal and shear loads acting on an edge of length L and mesh, each node i is associated with a vector of three unit thickness are given by unknown stresses, and each element e is associated with a 1 2 vector of two unknown body forces (which are not used in Qn L qn qn ¼ 1 þ 2 (5) this example, but are included for the sake of generality). Qs 2 qs qs Owing to the presence of stress discontinuities between all inter-element edges, multiple nodes may share the same where the local surface stresses qn and qs can be expressed coordinates, and each node is unique to an element. Across in terms of the Cartesian stresses at node i using the each stress discontinuity, the normal and shear stresses are standard transformation equations continuous. To satisfy the stress boundary conditions indi- 8 9 "# cated, the nodal stresses along the corresponding edges in <> i => the grid are subject to appropriate equality constraints (using qi cos2 sin2 sin 2 xx n ¼ i (6) the standard stress transformation relations), and the stresses qi 1 sin 2 1 sin 2 cos 2 > yy > i s 2 2 : i ; at each node i in the grid, , are subject to the yield xy condition f( i) < 0. The load to be maximised, given by When summed over each loaded boundary edge, the con- tributions Qn and Qs give the total force acting on the soil
ϭ⌠ Qs ⌡qAsd ⌠ 2 Q ϭ qAd q n n ⌡ n y 2 q s s n 2 q
iiϭ i i T iiϭ i i i i i T x 1 Nodeσ {σ xx, στyy,}xy Nodeσ {σ xx, σσττyy,,zz, xy yz,}τ xz q n 1 Segment length ϭ L Elementheeϭ {hh ,e } T Elementheeϭ {hhh ,e ,e } T 1 x y x y z q s Fig. 8. Linear elements for lower-bound limit analysis Fig. 10. Optimising the load along a boundary
qn ϭϭ⌠ MaximiseQqA⌡ n d collapse load
στϭϭ0 iiϭ i i T n Nodesσ {σστxx ,yy ,xy }
eeϭ e T Triangleshh { x ,}h y
0
ϭ
τ Stresses in triangles satisfy equilibrium
∂ ∂ σxx τxy ϩ ϩϩϭhg0 ∂x ∂y xx n ∂ ∂ 0 σyy τxy ϩ ϩϩϭhg0 ϭ ∂y ∂x yy τ s
Stresses in discontinuities satisfy equilibrium
στnnand ns continuous for adjacent elements y Stresses at nodes satisfy yield conditionf (σ i )р 0 x
Fig. 9. Illustrative lower-bound mesh for strip footing problem GEOTECHNICAL STABILITY ANALYSIS 541 mass per unit thickness. Typically, one of these quantities is T 1 bi 0 ci Bi ¼ e (10) maximised, but it is also possible to maximise a force A 0 ci bi resultant in a specified direction. Using equations (5) and (6), and summing over all the loaded edges, the objective and bi and ci are constants that depend on the nodal function (collapse load) can be expressed as coordinates. Rather than impose the equilibrium constraints in their native form, it is convenient to multiply both sides T c1 of equation (9) by the element area Ae: This permits an elegant implementation of stress discontinuities, as described where c1 is a vector of constants, and is the global vector in the next section, and leads to the modified equilibrium relations of unknown nodal stresses. At first sight, it would appear that the above approach is restricted to problems with linear B T B T B T e ¼ ðÞhe þ ge Ae (11) geometry and linear loading. This limitation can be relaxed, 1 2 3 however, by using the approach of Lyamin & Sloan (2002a), which uses a local coordinate system. where e e e T For body loads h ¼ hx, hy acting on an element of e T e T bi 0 ci area A and unit thickness, the corresponding resultant Bi ¼ A Bi ¼ (12) e e e e 0 ci bi forces, Qx ¼ A hx and Qy ¼ A hy, are shown in Fig. 11. These forces may be assembled over the grid to give the total load produced as Imposing the constraints in equation (11) ensures that the stresses satisfy the equilibrium conditions at every point in T c2 h an element, thus satisfying a key requirement of the lower- bound theorem. where c2 is a vector of constants (element areas) and h is a global vector of element body loads. In practice, the most common case of body force optimisation involves a variable Discontinuity equilibrium unit weight ª so that he ¼ {0, ª}T: For pseudo-dynamic Stress discontinuities can dramatically improve the accu- stability analysis, however, it is also useful to be able to racy of the collapse load obtained from lower-bound calcula- optimise the lateral body force component he: tions, and are introduced along all inter-element edges. x Following the formulation of Lyamin et al. (2005a), each discontinuity is modelled by a patch of continuum elements Continuum equilibrium of zero thickness, with opposing nodal pairs having the same In order to be statically admissible, the stresses in each coordinates, as shown in Fig. 12. element must satisfy the equilibrium equations To satisfy equilibrium, and thus be statically admissible, the normal and shear stresses must be the same on both @ xx @ xy sides of the discontinuity according to the relations þ þ hx þ gx ¼ 0 @x @y 1 2 3 4 (7) nn ¼ nn , nn ¼ nn (13) @ @ 1 2 3 4 yy þ xy þ h þ g ¼ 0 ns ns ns ns @y @x y y where for node i Over each triangle, the stresses vary linearly according to 8 9 "# the relations <> i => i 2 2 xx nn cos sin sin 2 i X3 i ¼ 1 1 > yy > (14) i ns 2 sin 2 2 sin 2 cos 2 : i ; ¼ N i (8) xy i¼1 Equations (13) and (14) imply that each pair of nodes on a where Ni are linear shape functions that are dependent on x stress discontinuity must obey two equality constraints on and y and the element nodal coordinates. Inserting equation their associated Cartesian stresses. Summing these con- (8) into equation (7) yields the pair of equilibrium equations straints over all nodal pairs on the discontinuities gives the BT BT BT e ¼ (he þ ge) (9) global set of conditions that must be satisfied for disconti- 1 2 3 nuity equilibrium. Since equation (11) holds true for any value of the where the terms B are the standard strain–displacement i element area, it is possible to set ! 0 for the triangles D1 (compatibility) matrices, defined by and D2 in Fig. 12, so that (x1, y1) ¼ (x2, y2) and (x3, y3) ¼ (x4, y4). Considering triangle D1, it can be shown that the equilibrium relations in equation (11) then become
3 ϭ (,xy11 ) (, xy 22 ) ϭ Qy (,xy33 ) (, xy 44 ) n D2 4 1 Q D s y x y 1 n δ → 0 2 Element area ϭ Ae x x s Fig. 11. Optimising body forces over an element Fig. 12. Statically admissible stress discontinuity 542 SLOAN 1 T T e T T 1 2 3 T f()σ р 0 B1 B1 0 ¼ B1 B1 0 ¼ 0 which implies 1