Available online at www.sciencedirect.

com

Computers and Structures 85 (2007) 1795–1807

www.elsevier.com/locate/compstruc

An efficient return algorithm for non-associated plasticity

with linear yield criteria in principal stress space
Johan Clausen a, Lars Damkilde a,*
, Lars Andersen b

a
Esbjerg Institute of Technology, Aalborg University, Niels Bohrs Vej 8, 6700 Esbjerg, Denmark
b
Department of Civil Engineering, Aalborg University, Sohngårdsholmvej 57, 9000 Aalborg, Denmark

Received 8 December 2006; accepted 4 April 2007

Available online 23 May 2007

Abstract

An efficient return algorithm for stress update in numerical plasticity computations is presented. The yield criterion must be linear in
principal stress space and can be composed of any number of yield planes. Each of these yield planes may have an associated or non-
associated flow rule. The stress return and the formation of the constitutive matrix is carried out in principal stress space. Here the
manipulations simplify and rely on geometrical arguments. The singularities arising at the intersection of yield planes are dealt with
in a straightforward way also based on geometrical considerations. The method is exemplified on non-associated Mohr–Coulomb plas-
ticity throughout the paper.
 2007 Elsevier Ltd. All rights reserved.

Keywords: Plastic stress update; Return mapping; Non-associated plasticity; Geotechnics; Mohr–Coulomb yield criterion; Non-linear FEM

1. Introduction and some recent contributions are given in the work by

Asensio and Moreno [4] and Rosati and Valoroso [5].
Stress update is a key part of numerical computations Several classical yield criteria are linear in principal
involving material plasticity. As the stress update is per- stress space. This includes the Mohr–Coulomb criterion
formed many times within each load step it is important often applied to soil and other granular materials. A special
that the process is fast and accurate in order to ensure an case of this is the Tresca criterion for metals and undrained
efficient numerical solution. Given a strain increment from soils. Another special case of the Mohr–Coulomb criterion
the solution of the global equilibrium equations the stress is the Rankine, or maximal principal stress criterion, which
must be updated at each integration point. The traditional is often used in conjunction with other criteria to mend
stress update schemes may be divided into two categories: deficiencies concerning yield strength in tension. A depic-
Explicit integration and return mapping. The method pre- tion of the Tresca and Mohr–Coulomb criteria in principal
sented in this paper comes under the latter category and is stress space can be seen in Fig. 1.
an updated and revised version of the conference paper [1]. With non-linear yield criteria, the return mapping pro-
The return mapping method was first promoted by Krieg cess is usually iterative; but the return to a linear yield
and Krieg [2] and this type of stress update method seems plane with a linear plastic potential can be performed in
to have been the most popular in recent years, judging from one step as the gradients in stress space are constant along
the number of papers on the subject. An overview of return the return path. The challenge arises when returning the
mapping algorithms is given in the book by Crisfield [3], stress to lines or points of intersection between yield planes,
which is not seldom in the case of linear criteria. When the
stress point is located on such a line or point, more than
*
Corresponding author. Tel.: +45 79 12 76 48. one yield plane is active, and the gradients are undefined.
E-mail address: [email protected] (L. Damkilde). Several approaches to this problem have been proposed

0045-7949/$ - see front matter  2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruc.2007.04.002
1796 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807

Fig. 1. Examples of linear yield criteria in principal stress space: (a) The Tresca criterion. (b) The Mohr–Coulomb Criterion.

by various authors in relation to the Mohr–Coulomb or the de ¼ dee þ dep ð1Þ

Tresca criterion. Abbo and Sloan [6] propose an approxi-
In perfect plasticity, plastic strains occur during yielding
mative rounding of the intersections, thereby avoiding
when
any singularities.
 T
When singularities are present several authors have pro- of
posed solutions based on Koiter’s theorem [7]. Crisfield [3] f ðrÞ ¼ 0 and dr ¼ 0 ð2Þ
or
takes a direct approach and performs the derivations nec-
essary for the stress return and the formation of constitu- where f is the yield function and r is the stress vector. The
tive matrices in the general six-dimensional stress space. matrix transpose is denoted with superscript T. The stress
This approach is valid for any yield criterion, but for iso- and strain vectors are ordered according to
tropic yield criteria the method presented in this paper is r ¼ ½ rx ry rz sxy sxz syz T
advantageous as it reduces the dimension of the problem ð3Þ
from six to three, no matter whether the criterion is linear e ¼ ½ ex ey ez 2exy 2exz 2eyz T
or not. The approach of De Borst [8] is similar to that of Eq. (2a) describes a closed hypersurface in stress space, and
Crisfield, but includes the non-associated case, hardening a stress state located inside this surface (f < 0) is elastic. As
and a method of detecting singularities similar to the one an elastic stress increment is related to an elastic strain
applied in the present paper. Pankaj and Bićanic [9] elabo- increment by Hooke’s law, use of (1) provides
rate on the detection of singular regions in the principal
stress space for Mohr–Coulomb plasticity. Perić and Neto dr ¼ D dee ¼ Dðde  dep Þ ¼ D de  D dep ð4Þ
[10] and Larsson and Runesson [11] carry out the manipu- where D is the elastic constitutive matrix. The present anal-
lations and establish the constitutive matrices in principal yses are confined to linear, isotropic elasticity. Here D is gi-
stress space, using rather complicated tensorial analysis. ven in terms of the Young’s modulus, E, and Poisson’s
The method of the present paper basically yields the ratio, m. For a finite strain increment, integration of (4)
same result as the methods of the references mentioned yields a finite stress increment
above. However, the derivation is carried out with geomet-
ric argumentation in the three-dimensional principal stress Dr ¼ DDe  DDep ¼ Dre  Drp ð5Þ
space and is thereby much simpler. The same goes for the which implies the assumption that a finite stress increment
resulting formulae for the stress update and formation of is composed of an elastic part followed by a plastic part, see
the constitutive matrices, which are all carried out in prin- Fig. 2.
cipal stress space and subsequently transformed back into Eq. (5) can also be written as
the original six-dimensional stress space. Additionally the
method is valid for all linear elastic–perfectly plastic mate- rC ¼ rB  Drp ð6Þ
rial models, where the yield functions and the plastic poten-
tials are linear in principal stress space. All formulae are
expressed using matrix notation and are therefore well sui-
ted for implementation in computer code. The simplicity
results in computational efficiency, which is demonstrated
with numerical examples.

2. Fundamentals of plasticity and return mapping

The basic relation in small-strain plasticity is that a

strain increment is composed of an elastic and a plastic part
Fig. 2. The principle of return mapping.
 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807 1797

The term Drp is usually referred to as the plastic corrector Eq. (14) is valid for any elastic–perfectly plastic continuum.
stress, rC = rA + Dr is the updated stress state and rB =
rA + Dre is the elastic predictor stress state. Eqs. (5) and 2.2. Consistent constitutive matrix
(6) are basically the return mapping scheme, which is also
illustrated in Fig. 2. In general, plastic strain increments If Dep is used in the global iterations the convergence
are derived from a plastic potential, g, as will be slow, as the stress and strain increments are finite
og rather than infinitesimal. Therefore a relation is needed
dep ¼ dk ð7Þ between changes in finite stress and strain increments,
or
dDr ¼ Depc dDe ð15Þ
where k is a positive multiplier. Eq. (7) is termed the flow
epc
rule. If g = f the flow rule is associated, but in soil mechan- where D is the so-called consistent constitutive matrix,
ics most often g 5 f. In principle the plastic corrector is first derived by Simo and Taylor [12]. Insertion of (9) in
found by inserting (7) into (1) and integrating (5), while remembering that Dre = DDe, yields
Z kþDk 
og og 
Drp ¼ D dk ð8Þ Dr ¼ DDe  DkD  ð16Þ
k or or C
Eq. (8) is evaluated as A small perturbation of (16) gives

p og  og o2 g
Dr ¼ DkD  or ð9Þ dDr ¼ DdDe  dDkD  DkD 2 dDr ð17Þ
or C or or

og  and after rearranging
Drp ¼ DkD  ð10Þ
or B  1  
o2 g og
where jC refers to evaluation at the updated stress point, dDr ¼ I þ DkD 2 D IdDe  dDk ð18Þ
or or
rC, and jB at the predictor point, rB. Eq. (9) corresponds
to fully implicit integration and usually requires an itera- By introduction of the matrices
tive procedure for general yield criteria, as rC is unknown.  1
o2 g
For linear criteria and potentials, (9) and (10) yield the T ¼ I þ DkD 2 and Dc ¼ TD ð19Þ
same result. Eq. (10) is named the radial return after Krieg or
and Krieg [2] and is exact for linear yield criteria, but in Eq. (18) can be written as
general not as robust as the implicit version.
og
dDr ¼ Dc dDe  dDkDc ð20Þ
or
2.1. Infinitesimal constitutive matrix
Comparing Eqs. (20) and (12) and following the same
For use in the global equilibrium iterations a constitu- approach as in obtaining (14), the relation between changes
tive matrix must be calculated. This is composed of an in finite stress and strain increments is found to be
infinitesimal constitutive matrix, Dep, which is then modi- dDr ¼ Depc dDe
fied to be consistent with global equilibrium iterations of  T
the Newton–Raphson type. Dep relates infinitesimal strain Dc og of Dc ð21Þ
Depc ¼ Dc   orTor
and stress increments of og
Dc or
or
dr ¼ Dep de ð11Þ where T and hereby Dc is evaluated at rC. For linear crite-
Eqs. (4) and (7) are combined into ria Crisfield [3] showed that the consistent constitutive ma-
trix, Depc, can be calculated in a much simpler fashion at
og
dr ¼ Dde  dkD ð12Þ the stress predictor point, rB
or
Depc ¼ T Dep
By insertion of (12) into (2b), dk is found to be 
of T o2 g  ð22Þ
Dde T ¼ I  DkD 2 
dk ¼ or T ð13Þ or B
of og
or
D or with Dep given by (14). Hereby the matrix inversion is
The relation between infinitesimal stresses and strains is avoided.
then obtained by back–substitution into Eq. (12),
3. Stress update in principal stress space
dr ¼ Dep de where
 
og of T The stress update and formation of the consistent con-
D D ð14Þ
Dep ¼ D  orT@r stitutive matrix requires the derivative of the yield function
of og
or
D or and the first and second derivatives of the plastic potential.
1798 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807

This is a cumbersome task when carried out in the general The Mohr–Coulomb criterion comprises six planes in
six-dimensional stress space for linear criteria as shown by principal stress space forming an irregular pyramid as
Crisfield [3]. As only isotropic material models are consid- can be seen in Fig. 1b. If the principal stresses are ordered
ered the manipulations can be carried out with respect to according to
any set of coordinate axes. Therefore the predictor stress r1 = r2 = r3 ; ð23Þ
is transformed into principal stress space and returned to
the yield surface. Considering the fact that the stress return the stresses are returned to only one of the six yield planes,
preserves the principal directions, the updated stress can as the other five correspond to an interchange of the order-
then be transformed back into the original coordinate sys- ing in Eq. (23). This plane is referred to as the primary yield
tem. The constitutive matrices are also formed in principal plane and it is shown in Fig. 4. The figure shows the pri-
stress space and then subsequently transformed. All trans- mary yield plane from two different points of view and also
formations rely on standard coordinate transformation. It the cross sections in the planes r1 = r2 and r2 = r3. The ro-
will be shown in the following that this approach simplifies man numerals refer to different stress predictor regions,
the manipulations of Section 2 remarkably. There are two which will be defined subsequently.
reasons for this. Firstly the dimension of the problem In the following the components of vectors and matrices
reduces from six to three, and secondly, in the three-dimen- are expressed with respect to the principal axes unless
sional stress space the stress states can be visualised graph- otherwise stated. This means that the last three compo-
ically, making it possible to apply geometric arguments. nents of vectors are always zero and may not be shown
The approach is applicable for general isotropic yield crite- as a matter of convenience. Even so, all matrices and vec-
ria, but in the following only criteria which are linear in tors are six-dimensional.
principal stress space will be considered. In this case,
closed-form solutions are found. The formulae are exempli- 3.1. Return to a plane, general formulation
fied on the Mohr–Coulomb material model with a non-
associated flow rule. The equation of a yield plane in the principal stress
Linear yield criteria in the principal stresses are visual- space can be written as
ised as planes in principal stress space. These planes inter-
f ðrÞ ¼ aT ðr  rf Þ ¼ 0 ð24Þ
sect in lines and points, making three types of stress returns
f
and constitutive matrices necessary: where r is a point on the plane and a is the gradient,
of
– Return to a yield plane. a¼ ð25Þ
or
– Return to a line, i.e. intersection of two yield planes.
– Return to a point, i.e. intersection of three or more yield The plastic potential is also taken to be linear in principal
planes. stress space, i.e.
og
The three types of return are visualised in Fig. 3. The gðrÞ ¼ bT r with b ¼ ð26Þ
or
formulae for the different returns and corresponding con-
Both a and b are constant. A first-order Taylor expansion
stitutive matrices will be established in the following. The
of (5), using (9), yields the well established solution for Drp,
conditions for determining which return is needed will also
see, for example, reference [13],
be established by dividing the stress space into different
stress regions. f ðrB Þ
Drp ¼ Db ¼ f ðrB Þrp ð27aÞ
bT Da
Db
rp ¼ T ð27bÞ
b Da
where rp is the direction of the plastic corrector in principal
stress space, i.e. rp is at an angle with the plastic strain
direction, b.

3.1.1. Return to a plane, Mohr–Coulomb plasticity

The Mohr–Coulomb criterion and plastic potential are
usually written as
f ðrÞ ¼ ðr1  r3 Þ þ ðr1 þ r3 Þ sin u  2c cos u ¼ 0 ð28Þ
gðrÞ ¼ ðr1  r3 Þ þ ðr1 þ r3 Þ sin w ð29Þ
where u is the angle of internal friction, c is the cohesion
Fig. 3. Three intersecting yield planes in principal stress space with three and w is the dilation angle. Rewriting Eqs. (28) and (29)
types of return shown. to the format of Eq. (24) one obtains
 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807 1799

Fig. 4. Different views of Mohr–Coulomb yield plane in principal stress space: (a) isometric view, (b) Trace in p-plane, (c) intersection of the planes
r1 = r2 and f = 0 (compressive meridian) and (d) crossing of planes r2 = r3 and f = 0 (tensile meridian). P is the hydrostatic axis.

pffiffiffi
f ðrÞ ¼ aT1 ðr  ra Þ ¼ kr1  r3  2c k ¼ 0 ð30Þ
gðrÞ ¼ bT1 r ¼ mr1  r3 ð31Þ
where
T 1 þ sin u
a1 ¼ ½ k 0 1  ; k¼ ð32Þ
1  sin u
T 1 þ sin w
b1 ¼ ½ m 0 1  ; m¼ ð33Þ
1  sin w
As the point on the plane, the apex point, ra, with the prin-
cipal coordinates
pffiffiffi
2c k T ð34Þ
ra ¼ ½1 1 1 Fig. 5. Return to intersection line, ‘.
k1
is chosen. Thepuniaxial
ffiffiffi compressive yield strength of the
material is 2c k ¼ aT1 ra . Together with the ordering of
the principal stresses in (23), (30) describes the triangular r ‘ / a1  a2 ð36Þ
plane in principal stress space shown in Fig. 4. The scaled
direction of the plastic corrector is obtained by insertion of where ‘‘·’’ is the cross product between the first three com-
(32) and (33) in (27). ponents of the vectors, so that r‘ is perpendicular to both a1
and a2. The length of r‘ is not important, hence the use of
3.2. Return to a line, general formulation ‘‘/’’ instead of ‘‘=’’ in Eq. (36).
Analogously the direction of the plastic potential line, r‘g
The intersection between two yield planes f1 = 0 and is defined by
f2 = 0 defines a line, ‘ (see Fig. 5), with the equation
r‘g / b1  b2 ð37Þ
‘ : r ¼ tr‘ þ r‘ ð35Þ
where t is a parameter with the unit of stress and r‘ is a The plastic strain increment must be perpendicular to
point on the line. The direction vector of the line is r‘, the direction of the plastic potential line, r‘g , see Fig. 6.
1800 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807

This stress return also conforms to the solution of Koiter

[7], in the sense that the resulting strain increment can be
expressed as a linear combination between the gradients
of all the active potential planes.

3.4. Stress regions

In the previous sections formulae for the returned stress

state have been given. In this section it will be clarified how
to determine to which plane, line or point the stress should
Fig. 6. The plastic strain increment is perpendicular to the potential line be returned. In order to do this the concept of stress regions
when returning to an intersection line, ‘.
is introduced, and the boundary planes that separate them
are defined. Each yield plane, line and point is associated
T T with a particular stress region. When the predictor stress
ðDep Þ r‘g ¼ 0 () ðD1 Drp Þ r‘g ¼ 0
is located in a given region it must be returned to the cor-
() ðrB  rC ÞT D1 r‘g ¼ 0 ð38Þ responding plane, line or point. Two stress regions, I and
II, separated by a boundary plane, pI-II = 0 are illustrated
As the updated stress, rC, belongs to the line, Eq. (35) can
in Fig. 7. When the yield functions and plastic potentials
be substituted into Eq. (38) and give a solution for t ex-
are linear in the principal stresses, the boundary planes
pressed in the direction vectors of the intersection line
are also linear. The direction of the plastic corrector, rp,
and the plastic potential line
cf. (27), and the direction vector of the line, r‘, define the
T
ðr‘g Þ D1 ðrB  r‘ Þ orientation of the plane, and so the equation of a boundary
t¼ ð39Þ plane can be found as:
ðr‘g ÞT D1 r‘
T
This return also corresponds to Koiter’s theorem [7], pIII ðrÞ ¼ ðrp  r‘ Þ ðr  r‘ Þ ¼ nTIII ðr  r‘ Þ ¼ 0 ð43Þ
that states that the plastic strain increment is composed where nI-II is the normal of the plane. The indices indicate
of a linear combination of the strain directions of the active which stress regions the plane separates. The point on the
potential planes. plane is r‘, which can be taken as a point that also belongs
to ‘, see Fig. 7 and Eq. (35). If two stress regions are lo-
3.2.1. Return to a line, Mohr–Coulomb plasticity cated as seen in Fig. 7, the following is valid for a given pre-
The Mohr–Coulomb plane in Fig. 4 is delimited by two dictor stress, rB located outside the yield locus, i.e.
lines, ‘1 and ‘2 with the parametric equations f(rB) = 0:
‘1 : r ¼ t1 r‘1 þ ra ; r‘1 ¼ ½ 1 1 k T pIII ðrB Þ 5 0 () Reg: I () Return to f ¼ 0
a T ð40Þ ð44Þ
‘2 : r ¼ t2 r‘2 þr ; r‘2 ¼ ½1 k k pIII ðrB Þ > 0 () Reg: II () Return to ‘
where ra is the apex point defined in (34). The line denoted
‘1 corresponds to triaxial compression, whereas the line de-
noted ‘2 corresponds to triaxial tension. 3.4.1. Stress regions, Mohr–Coulomb plasticity
The corresponding potential direction vectors are given Four distinct returns exist for a given predictor stress:
by Return to the yield plane f = 0, to lines ‘1 or ‘2 and to
8 9 8 9 the apex point ra. Therefore four stress regions, I–IV, are
<1 >
> = <1 >
> = needed, see Figs. 4 and 8. Four boundary planes separate
‘ ‘
rg;1 ¼ 1 and rg;2 ¼ m ð41Þ
: >
> ; : >
> ;
m m
where m is defined in (32). With these direction vectors the
parameters t1 and t2 in (40) are found from (39) and the up-
dated stress is then given by ((40)a) or ((40)b) as
appropriate.

3.3. Return to a point

If the stress is to be returned to a singularity point, ra,

e.g. an apex point, see Fig. 3, there is no need for calcula-
tions, as the returned stress is simply
Fig. 7. Boundary plane pI–II = 0 with normal nI–II, which separates the
rC ¼ ra ð42Þ stress regions I and II.
 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807 1801

Fig. 8. Boundary planes and stress regions in Mohr–Coulomb plasticity from two different points of view. (a) Isometric view of yield and boundary
planes. (b) Overview of the stress regions seen from the direction of rp1. The roman numerals represent stress regions.

Table 1 Dep b ¼ 0 ð48Þ

Conditions for Mohr–Coulomb stress return
Condition Region Return to Along the yield surface the stress increments are elastic,
pI–II = 0 ^ pI-III 5 0 I f=0 dr ¼ Dep ðdkb þ dee Þ ¼ Dep dee ¼ Ddee ð49Þ
pI–II < 0 ^ pI-III < 0 II ‘1
pI–II > 0 ^ pI-III > 0 III ‘2 where (1), (4) and (7) have been utilised.
t1 > 0 ^ t2 > 0 IV apex
Valid when f(rB) = 0.
4.1. Dep on a plane

the stress regions, as seen in Fig. 8. The equations of the When the updated stress state is located on a yield plane,
boundary planes pI-II and pI-III can be found from (43), the infinitesimal constitutive matrix is given by (14), here
 T repeated as
pIII ðrÞ ¼ rp1  r‘1 ðr  ra Þ ¼ 0 ð45Þ
p  DbaT D
T
pIIII ðrÞ ¼ r1  r‘2 ðr  ra Þ ¼ 0 ð46Þ Dep
f ¼ D ð50Þ
aT Db
In principle the equations of the other two boundary
planes, pII-IV and pIII-IV, are needed, but a computationally 4.2. Dep on a line
more efficient means of determining whether the predictor
stress is located in Region IV exists. The parametric equa- When the updated stress is located on a line the only
tions of ‘1 and ‘2, cf. Eq. (40), are defined such that t1 = t2 possible direction of the stress increment is in the direction
= 0 at the apex. If t1 > 0 ^ t2 > 0, the predictor stress is lo- of the line, r‘, see Fig. 9. The infinitesimal constitutive
cated in Region IV. This way of evaluating the stress region matrix on the line, Dep‘ , must be singular with respect to
is efficient because t1 or t2 might be needed when updating the strain directions associated with both of the yield
the stresses.
The conditions for determining the region, and hereby
the return, are then deduced from Fig. 8 and can be seen
in Table 1.

4. Infinitesimal constitutive matrix

The infinitesimal constitutive matrix, see (11), relates

infinitesimal stress and strain increments as
dr ¼ Dep de ð47Þ

For perfect plasticity (2b) states that the strain increment

Fig. 9. An elastic strain direction vector, re and a direction vector r‘ of an
must be tangential to the yield surface. This means that intersection line between to yield planes. The vectors b1 and b2 are the
Dep is singular in the direction of the plastic strain incre- gradients of the two potential planes, see (26), associated with the yield
ment, b = og/or, planes.
1802 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807

planes that define the line, b1 = og1/or and b2 = og2/or, DkD(o2 g/or2) can be formed by geometrical arguments
and to any linear combination of the two and achieve a very simple form,
2 3
Dep
‘ ðl1 b1 þ l2 b2 Þ ¼ 0 ð51Þ 0
6 0 7
where l1 and l2 are plastic multipliers. As r‘ is the only 6 7
6 7
6 0 7
possible direction of the stress increment, the elastic strain 6 7
6 p
Dr1  Dr2p 7
increment must, according to Eq. (49), have the direction o2 g 66
7
7
DkD 2 ¼ 6 r1  r2 7
re ¼ D1 r‘ ð52Þ or 6 p p 7
6 Dr1  Dr3 7
6 7
Any strain increment in principal stress space can be writ- 6 r1  r3 7
6 7
ten as a linear combination of three non-parallel directions 4 Dr2  Dr3 5
p p

r2  r3
de ¼ dl1 b1 þ dl2 b2 þ dcre ð53Þ
ð59Þ
Then from Eqs. (49) and (51) the following system of equa-
The lower right 3 · 3 partition holds for any kind of plas-
tions is defined
ticity but the upper left 3 · 3 partition consist of zeros for
b ep re ¼ r‘
D linear potentials only.
‘
b ep b1 ¼ 0 The components of the plastic corrector in principal
D ð54Þ
‘ stress space, Drp1 , Drp2 , Drp3 , are given by Eq. (6). The prin-
b ep b2
D ¼0 cipal stresses r1, r2, r3, are either the values at the predictor
‘
point, rB, or at the updated stress point, rC. If the general
where D b ep only contains elements related to normal stres-
‘ definition of Depc, Eq. (21) is used, then the denominators
ses, i.e. the elements of the upper left quadrant. The solu- are evaluated at rC. Otherwise, if the plastic potential and
tion to Eq. (54) reads the yield function are linear, leading to a Depc defined by
r‘ ðr‘g Þ
T
ða1  a2 Þðb1  b2 Þ
T (22), the denominators are evaluated at rB.
b ep ¼
D ¼ ð55Þ If the denominator of any of the fractions in (59) vanish,
‘ T T
ðr‘ Þ D1 r‘g ða1  a2 Þ D1 ðb1  b2 Þ the fraction is reduced to unity, which is the limit for the
when (52) and (36) are utilised. The full solution in six- denominator ! 0. To elaborate on this, consider the plas-
dimensional stress space includes the shear stiffness, G, tic corrector
8 B 9
 C
E 033 033 < r1  r1 >
> =
ep b ep
D‘ ¼ G þ D ‘ ; G ¼ ð56Þ
2ð1 þ mÞ 033 I33 Drp ¼ rB  rC ¼ rB2  rC2 ð60Þ
>
: B >
;
r3  rC3

4.3. Dep on a point as can be seen from Eq. (6). As an example Eq. (60) is in-
serted in the (4,4)-term of (59)
When the updated stress is located at an apex point, see  
o2 g Drp  Drp2 rB1  rC1  ðrB2  rC2 Þ
Fig. 3, the infinitesimal matrix must be singular with DkD 2 ¼ B1 ¼
or 4;4 r1  rB2 rB1  rB2
respect to any direction in the principal stress space, i.e.
the direction of the normal stresses rC1  rC2
¼1 ð61Þ
b ep ¼ 0 rB1  rB2
D point ) Dep
point ¼ G ð57Þ
If the predictor stress is located in region II, the stress is
If the yield plane contains a point on the hydrostatic axis, returned to ‘1 where rC1 ¼ rC2 , see Fig. 10. This implies that
this will always be an apex point for isotropic material, and (61) reduces to unity. This will be the case for all stress
hence an intersection point for six yield planes in six- states in the limit rB1 ! rB2 , cf. Fig. 10, which leads to the
dimensional stress space. This means that Dep point is singular conclusion that unity is indeed the limit for rB1 ¼ rB2 .
with respect to any direction and therefore Both the infinitesimal and the consistent constitutive
Dep
point ¼ 0 ð58Þ matrices are thus formed in principal stress space. For lin-
ear yield criteria the infinitesimal constitutive matrix Dep is
formed from either (50), (56), (57) or (58), as appropriate.
4.4. Consistent constitutive matrix The modification matrix, T, is formed by inserting (59) in
(22). Then the consistent constitutive matrix in principal
The consistent constitutive matrix is defined in Section stress space is formed by Depc = TDep, and finally trans-
2.1. In Ref. [14] an alternative form is derived. The idea formed back into the original stress space using coordinate
is that the consistent constitutive matrix, Depc, and hereby transformation, see Appendix A.
the modification matrix T of (21) and (22) is formed It should be noted that the consistent constitutive matrix
in principal space. In principal stress space the term found from the equations above is identical to the consis-
 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807 1803

When the method is used on plane problems, special care

should be taken in the calculation of this transformation
matrix and the calculation of the modification matrix T,
see Appendix B.

6. Examples of implementation

An example of the performance of the method is pre-

sented in this section. First the limit load is computed for
a circular footing resting on a Mohr–Coulomb material,
in the associated as well as in the non-associated case. Next
a comparison of computation time between the present
method and the direct implementation of return mapping
by Crisfield [3] is shown.

6.1. Computational example

Fig. 10. All predictor stress states in the limit rB1 ! rB2 are returned to the
line ‘1 where rC1 ¼ rC2 .
A smooth rigid footing is placed on a domain of fric-
tional soil, as shown in Fig. 11. Two cases will be consid-
tent constitutive matrix derived by Crisfield [3] by a direct ered: A strip footing, i.e. plane strain, and a circular
application of the return mapping formulae in Section 2. footing, i.e. axisymmetry. The geometry and the boundary
But the expression for the double singular constitutive conditions are seen in Fig. 11.
matrix on a line, Eq. (55), is much simpler, and hence faster The frictional soil has a weight of 20 kN/m3 and has the
to compute, than the corresponding expression in reference yield parameters u = 20, c = 0, and the deformation
[3]. parameters E = 2 · 107 Pa, m = 0.26. Both an associated
and a non-associated material model is employed. In the
5. Summary of the method non-associated model the dilation angle is set to w = 0.
A forced displacement, u, is applied to the footing. These
A brief summary of the method is given in Table 2. Con- parameters correspond to a calculation of the bearing
cerning the calculation of the principal stresses in Step 2 capacity factor Nc. The average footing pressure, p, is com-
this can be done either analytically or with a built-in eigen- puted as the sum of footing reactions divided by the foot-
solver, depending on the efficiency. The calculation of the ing area. The footing nodes are free to move in the
infinitesimal constitutive matrices in Step 6 can be per- horizontal direction in order to simulate the smooth
formed outside the function and supplied as an input, as interface.
these matrices are constant in the principal stress space. The results will be compared with the exact values given
This is not the case for the consistent constitutive matrix, by Martin [15,16], Nc = 1.57862 for plane strain and
and therefore the modification matrix, T, must be calcu- Nc = 1.271 for axisymmetry with the present parameters.
lated for each stress update. The exact bearing capacity is then given by pex = rNc,
The coordinate transformation matrix, A for general where r is the footing halfwidth or radius. The exact values
six-dimensional stress states is derived in Appendix A. are for an associated material only, but they will also be

Table 2
Summary of the method
0 0
INPUT: r ADe D yield parameters
0 0
1.Predictor stress, r B = r A + DDe 0
2.Calculate principal predictor stress, rB
0 0
3.Calculate fi(rB). If all fi < 0, r C = r B, Depc = D and EXIT
4.Determine predictor stress region with boundary planes
5.Calculate rC by returning the stress
6.Calculate infinitesimal constitutive matrix, Dep
7.Calculate modification matrix, T
8.Calculate consistent constitutive matrix, Depc
9.Calculate principal directions and transformation matrix, A
(see Appendix A)
10. Transform rC and Depc back into the original space
0 0 Fig. 11. Geometry and boundary conditions in the computational
OUTPUT: r CD epc example. The system is (axi-)symmetric with respect to the left boundary.
1804 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807

compared to the numerical results for an associated

material.
The domain is modelled with 6-noded triangular linear
strain elements. The left, right and lower boundaries are
supported against displacements perpendicular to their
directions. An example of the element mesh with 1962
degrees of freedom can be seen in Fig. 12. The global con-
vergence tolerance is set to
pffiffiffiffiffiffiffiffiffiffi
TOL ¼ 105 RT R ð62Þ
where R is the vector of global reactions at the last con-
verged load step.
In Fig. 13 examples of load-displacement histories can
be seen. The load has been normalised with respect to the
exact bearing capacity and the displacement with respect Fig. 14. Relative difference from exact solution of the associated material
to the footing radius/halfwidth. The curves are computed versus the number of degrees of freedom, ndof.
with the mesh shown in Fig. 12, which is rather coarse,
which leads to an overshoot compared to the exact bearing The relative difference in maximum load for both mate-
capacity. This overshoot decreases as the mesh is refined, rials is shown in Fig. 14. This relative difference is given by
see Fig. 14. As expected the associated material can sustain  
a higher load before collapse. pmax
Relative difference ¼  1  100% ð63Þ
pex
It is seen that the results converge for both materials
although it could be expected that in the limit the failure
load for the non-associated case will be lower than pex.
With the denser meshes the calculation with the non-
associated material needs an increased number of load
steps to avoid numerical instability. This problem is more
outspoken with higher friction angles, if the dilation angle
is kept as w = 0. This problem is related to the non-unique-
ness of the solutions for a non-associated material and is
therefore not related to the stress return in each Gauss
point.

6.2. Comparison with classical implementation

For an estimation of the efficiency of the presented

Fig. 12. Example of element mesh with 468 elements and 1962 degrees of method a comparison with the direct implementation of
freedom. the return mapping scheme shown in Crisfield [3] is carried
out. The results of both the returned stress and the consti-
tutive matrices are identical within machine precision, so
the comparison is on computation time only. The material
is the non-associated material used in the previous section.
Comparison is made based on computation time for 10 000
stress returns and formation of the corresponding consis-
tent constitutive matrices. The algorithms are implemented
in MATLAB, and the computations are carried out on a lap-
top computer with Pentium (R) M 1.4 GHz processor and
512 MB RAM. The material parameters are the same as in
the previous example in the associated case. The results are
shown in Table 3.
It is seen that the present method is substantially faster,
especially when returning to a line, which is the case for
roughly 75% of the stress returns with the axisymmetry cal-
Fig. 13. Normalized load–displacement curves generated with the element culation in the example of the previous section. One reason
mesh shown in Fig. 12. for the increased speed is mentioned in Section 5, namely
 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807 1805

Table 3 The average numbers of equilibrium iterations for all the

Comparison of computation time for 10 000 stress returns load steps shown in Fig. 13 are 3.73 and 3.63 for the non-
T classic
10 000 returns to Tclassic Tpresent
T present associated material in axisymmetry and plane strain respec-
Plane 3.194 s 2.644 s 1.208
tively. An identical calculation carried out using the infini-
Line 4.867 s 3.154 s 1.543 tesimal constitutive matrix, Dep, instead of Depc makes
Point 3.395 s 2.633 s 1.289 these numbers increase with at least a factor 10.

7. Conclusion
that Dep must be calculated in each stress update in the
classical method, while this is not the case in the present A method is presented for stress update in the principal
method, where Dep is a constant in principal stress space. stress space for isotropic material models. The formulation
The price to be paid, however, is the coordinate transfor- in principal stress space results in simple and efficient for-
mation. In the above example the built-in eigensolver of mulae for the stress update, that are easily implemented
MATLAB has been utilised.
in finite element software as matrix notation is employed.
The method is elaborated for linear yield criteria with
6.3. Rate of convergence linear plastic potentials, and it is exemplified on a Mohr–
Coulomb material, assuming both associated and non-
To examine the properties the proposed consistent con- associated plasticity. All types of singularities are handled,
stitutive matrix, two examples of the rate of convergence and it is also explained how to determine if the predictor
will be given. The examples are taken from the calculation stress is located in a singular region, in a simple and unam-
of the curves shown in Fig. 13 for the non-associated mate- biguous manner. The method also includes calculation of
rials. The residual is given by constitutive matrices in the principal stress space, for all
qffiffiffiffiffiffiffiffiffiffiffi
types of stress return. In case of stress returns to lines
Residual ¼ QT Q ð64Þ and points the formulae simplify considerably compared
where Q is the global vector of residual forces. to the direct implementation of the return mapping
More specifically the development of the residual for the formulae.
equilibrium iterations of load step 9 in Fig. 13 is shown in It is shown that the method performs correctly and effi-
Table 4. It is seen that the convergence rate is quadratic or cient in comparison with classical methods, and that the
nealy quadratic. This is to expected as the presented consti- quadratic convergence rate, which should be expected for
tutive matrices are identical to the direct derivation found the consistent constitutive matrix, is achieved.
in reference [3]. The presented method is implemented in MATLAB and
In the particular load step shown in Table 4, the distri- FORTRAN and the code can be obtained from the corre-

bution of stress points in the different stress regions are as sponding author.
shown in Table 5. The table shows that the constitutive
matrix on the lines is activated in a large number of the Appendix A. Coordinate transformation matrix
stress returns.
The principal stresses and directions are found by solv-
ing the well-known eigenvalue problem
Table 4
Development of global residual for load step 9 in plane strain and ðr0ij  ldij Þnj ¼ 0; i ¼ 1; 2; 3 ðA:1Þ
axisymmetry
Iteration number Axisymmetry Plane strain where r0ij is the stress tensor, l is the eigenvalue, dij is the
1 2.018 · 103 1.533 · 103 Kronecker delta and nj is the eigenvector. The three eigen-
2 1.116 · 102 8.131 · 102 vectors form a coordinate transformation tensor, Kij
3 3.769 2.045 · 102 2 x0 3
4 4.823 · 103 2.063 · 10 cx cx0y cx0z
5 7.859 · 102 6 7
Kij ¼ ½ n1j n2j n3j  ¼ 4 cy0x cy0y cy0z 5 ðA:2Þ
6 4.296 · 106
z0 z0 z0
cx cy cz

where the components are direction cosines between the

Table 5
Distributions of stress points in different stress regions, see Fig. 8 and
two sets of axes, e.g. cy0x ¼ cos wy0x , where wx0y is the angle be-
Table 1 tween the y 0 -axis and the x-axis. With the elements of Kij
Regions Return to Axisymmetry Plane strain
the transformation of the strain and stress vector (see Eq.
(3)) can be written as
I Plane, f = 0 18% 77%
II Line, ‘1 78% 23%
e ¼ Ae0 or e0 ¼ A1 e ðA:3Þ
III Line, ‘2 0% 0%
T 0 0 T
IV Apex rC = ra 4% 0% r¼A r or r ¼ A r ðA:4Þ
1806 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807

The transformation matrix, A, is given as

2 3
cx0x cx0x cy0x cy0x cz0x cz0x cx0x cy0x cz0x cx0x cy0x cz0x
6 cx0 cx0 cy0 cy0 cz0 cz0 cx0y cy0y cz0y cx0y cy0y cz0y 7
6 y y y y y y 7
6 x0 x0 y0 y0 z0 z0 x0 y0 z0 x0 y0 z0 7
6 cz cz cz cz cz cz cz cz cz c z cz cz 7
A¼6
6 2cx0 cx0 2cy0 cy0 2cz0 cz0 cx0 cy0 þ cx0 cy0 cz0 cx0 þ cz0 cx0 cy0 cz0 þ cy0 cz0 7
7
6 x y x y x y x y y x x y y x x y y x 7
6 x0 x0 7
4 2cz cx 2cy0z cy0x 2cz0z cz0x cx0z cy0x þ cx0x cy0z cz0z cx0x þ cz0x cx0z cy0z cz0x þ cy0x cz0z 5
2cx0y cx0z 2cy0y cy0z 2cz0y cz0z cx0y cy0z þ cx0z cy0y cz0y cx0z þ cz0z cx0y cy0y cz0z þ cy0z cz0y
ðA:5Þ
ep
A constitutive matrix, D is transformed according to
Dep ¼ AT D0ep A or D0ep ¼ ADep AT ðA:6Þ Fig. B.1. The x 0 y 0 z 0 and x y z coordinate systems in a plane problem.
2 3
cos2 a 0 sin2 a cos a sin a
Appendix B. Remarks about plane calculations 6 0 1 0 07
6 7
A¼6 7 ðB:3Þ
4 sin2 a 0 cos2 a  cos a sin a 5
The derivations presented in this paper have been con-
cerned with general three-dimensional stress states. Several  sin 2a 0 sin 2a cos2 a  sin2 a
problems, including the ones in Section 6.1, can be treated And finally if rz = r3
as plane problems, either as plane strain or axisymmetry. 2 3
In this appendix a few remarks will be tied to the imple- cos2 a sin2 a 0 cos a sin a
mentation of the method in two dimensions (2D). The 6 sin2 a cos2 a 0  cos a sin a 7
6 7
stress and strain vectors in 2D are taken to be A¼6 7 ðB:4Þ
4 0 0 1 0 5
8 9 8 9
>
> rx >> >
> ex >>  sin 2a sin 2a 0 cos2 a  sin2 a
>
> >
> >
> >
>
< ry >
> = < ey >
> = The transformation matrices shown in Eqs. (B.2)–(B.4)
r¼ and e ¼ ðB:1Þ are based on the assumption that the full stress vector in
>
> rz >> >
> ez >>
>
> >
> >
> >
> principal stress space has the form
: >
> ; >
: >
;
sxy 2exy
r ¼ ½ r1 r2 r3 0 T ðB:5Þ
For axisymmetry x, y and z is taken to be radial, axial and
circumferential directions, respectively. with r1 = r2 = r3. The coordinate transformations are
then carried out at shown in Eqs. (A.4) and (A.6).
B.1. Coordinate transformation matrix in 2D
B.2. Modification matrix T in 2D
Compared to the coordinate transformation matrix in
The modification matrix, T is needed for the calculation
Eq. (A.5), the plane coordinate transformation matrix is
of the consistent constitutive matrix, cf. Eq. (22). In plane
a lot simpler. This is due to the fact that the z direction
calculations the term DkD (o2g/or2) of Eq. (59) reduces to
is always a principal direction. However, an important 2 3
point to consider is, that the out-of-plane stress, rz is not 0
always the intermediate principal stress, i.e. it is possible 6 0 7
o2 g 6 6
7
7
to have r1 = rz or r3 = rz, and therefore rz plays a part DkD 2 ¼ 6 7 ðB:6Þ
or 4 0
in evaluating the yield criterion, as this is expressed in r1 p p 5
Drf Drg
and r3, cf. Eq. (30). rf rg
This problem can be addressed by a proper ordering of
where the indices f and g take the following values
the rows in the stress transformation matrix, A. The ele-
ments of A depend on the angle between the two coordi-
– if rBz ¼ rB1 then f = 2 and g = 3
nate systems, a, shown in Fig. B.1.
– else if rBz ¼ rB2 then f = 1 and g = 3
If rz = r1
– or finally if rBz ¼ rB3 then f = 1 and g = 2
2 3
1 0 0 0
6 7 If the two active principal stresses are identical, rf = rg,
60 cos2 a sin2 a cos a sin a7 the term reduces to unity on account of the elaboration
6 7
A¼6 7 ðB:2Þ given in Section 4.4.
60 sin2 a cos2 a  cos a sin a 7
4 5
2 2
0  sin 2a sin 2a cos a  sin a References

Else if rz = r2 [1] Clausen J, Damkilde L, Andersen L. An efficient return algorithm for

non-associated Mohr–Coulomb plasticity. In: Topping BHV, editor.
 J. Clausen et al. / Computers and Structures 85 (2007) 1795–1807 1807

Proceedings of the tenth international conference on civil, structural [9] Pankaj, Bićanić N. Detection of multiple active yield conditions
and environmental engineering computing. Stirling, United King- for Mohr–Coulomb elasto-plasticity. Comput Struct 1997;62(1):
dom: Civil-Comp Press. p. 144. 51–61.
[2] Krieg R, Krieg D. Accuracies of numerical solution methods for the [10] Perić D, de Souza Neto E. A new computational model for tresca
elastic-perfectly plastic model. ASME J Pressure Vessel Technol plasticity at finite strains with an optimal parametrization in the
1977:510–5. principal space. Comput Methods Appl Eng 1999;171:463–89.
[3] Crisfield M. Non-linear finite element analysis of solids and struc- [11] Larsson R, Runesson K. Implicit integration and consistent linear-
tures. Advanced topics, vol. 2. John Wiley & Sons; 1997. ization for yield criteria of the Mohr–Coulomb type. Mech Cohesive-
[4] Asensio G, Moreno C. Linearization and return mapping algorithms Frictional Mater 1996;1:367–83.
for elastoplasticity models. Int J Numer Methods Eng 2003;57: [12] Simo J, Taylor R. Consistent tangent operators for rate-indepen-
991–1014. dent elastoplasticity. Comput Methods Appl Mech Eng 1985;48:
[5] Rosati L, Valoroso N. A return map algorithm for general isotropic 101–18.
elasto/visco-plastic materials in principal space. Int J Numer Meth- [13] Crisfield M. Non-linear finite element analysis of solids and struc-
ods Eng 2004;60:461–98. tures. Essentials, vol. 1. John Wiley & Sons; 1991.
[6] Abbo A, Sloan S. A smooth hyperbolic approximation to the Mohr– [14] Clausen J, Damkilde L, Andersen L. Efficient return algorithms for
Coulomb yield criterion. Comput Struct 1995;54(3):427–41. associated plasticity with multiple yield planes. Int J Numer Meth
[7] Koiter W. Stress-strain relations, uniqueness and variational theo- Eng 2006;66(6):1036–59.
rems for elastic-plastic materials with a singular yield surface. Quart [15] Martin CM. User guide for ABC – analysis of bearing capacity,
Appl Math 1953;11:350–4. version 1.0, OUEL Report No. 2261/03, University of Oxford; 2004.
[8] De Borst R. Integration of plasticity equations for singular yield [16] Martin CM. Exact bearing capacity for strip footings. <http://
functions. Comput Struct 1987;26(5):823–9. www.civil.eng.ox.ac.uk/people/cmm/ncnqngamma.xls> 2005.