GMKP - Finite Element Method in Geotechnical

Engineering

Theoretical Guide

Computer program
for nonlinear finite element analysis of geotechnical problems

FINE ltd. 1999–2003

April 15, 2003

Contents

1 Preface 3

2 Finte Element Equations 5

2.1 Kinematics discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Finite elements used in GMKP program . . . . . . . . . . . . . . . . . . . 8
2.3.1 2-node rod element . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.2 2-node and 3-node beam elements . . . . . . . . . . . . . . . . . . . 9
2.3.3 Plane 3-node and 6-node triangular elements . . . . . . . . . . . . . 14
2.3.4 4-node and 6-node interface element . . . . . . . . . . . . . . . . . . 19

3 Constitutive models 24
3.1 Constitutive model for anchors . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2 Elastic constitutive model for soil . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 Modified elastic constitutive model for soil . . . . . . . . . . . . . . . . . . 26
3.4 Elasto-plastic constitutive models for soil . . . . . . . . . . . . . . . . . . . 26
3.4.1 Invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4.2 Yield surface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.4.3 Elasto-plastic stiffness matrix . . . . . . . . . . . . . . . . . . . . . 30
3.4.4 Von Mises model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4.5 Drucker-Prager model . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.4.6 Modified Mohr-Coulomb model . . . . . . . . . . . . . . . . . . . . 42
3.5 Interface constitutive model . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5.1 Yield surface and stress update procedure . . . . . . . . . . . . . . 45
3.5.2 Tangent stiffness matrix . . . . . . . . . . . . . . . . . . . . . . . . 47

4 Solution strategies 49
4.1 Stress return mapping: an algorithm for standard plasticity . . . . . . . . . 50
4.1.1 Cutting plane algorithm . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2 Newton–Raphson method . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2.1 Standard Newton–Raphson method . . . . . . . . . . . . . . . . . . 53
4.2.2 Modified Newton–Raphson method . . . . . . . . . . . . . . . . . . 54
4.2.3 Initial stress method . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.2.4 Optimal step-length (line search) . . . . . . . . . . . . . . . . . . . 55

1
CONTENTS 2

4.2.5 Convergence criteria . . . . . . . . . . . . . . . . . . . . . . . . . . 57

4.2.6 Additional remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.3 Arc-length method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.3.1 Spherical ALM (Crisfield) . . . . . . . . . . . . . . . . . . . . . . . 64
4.3.2 Linearized ALM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3.3 Automatic step-length control . . . . . . . . . . . . . . . . . . . . . 69
4.3.4 Additional remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

5 Miscellaneous topics 72
5.1 Stability analysis of earth slopes . . . . . . . . . . . . . . . . . . . . . . . . 72
5.2 K0 procedure to generate initial geostatic stress state . . . . . . . . . . . . 74
5.3 Unloading – reloading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Chapter 1

Preface

Stability analysis of earth structures such as cut slopes, earth and rockfill embankments
has been one of the major disciplines in geotechnical engineering for many years. Tra-
ditionally such analyses have been performed using simplified methods or empirical ap-
proaches [26]. Most of these methods fall into the general category of slice methods of
limit equilibrium. Recall that majority of design codes or commercial engineering soft-
wares [7, 12, 13] are based on these approaches. Examples include, e.g., the Peterson,
Bishop or Sarma methods. All methods essentially assume the soil is at failure at very
point along a certain failure surface. Equilibrium conditions are then considered for failing
soil mass. To arrive at the desired solution, however, a number of simplifying assumptions
must be introduced (predetermined slip surface, direction of slices and related side forces,
etc.).
A suitable alternative to traditional limit equilibrium approaches is the finite element
method in that, that it is more versatile and requires fewer a priory assumptions, especially
regarding the failure mechanism. Evolution of failure zone is gradually dependent on the
deformation behavior of soils described by a suitable constitutive model [19]. Thus no
assumption needs to be made in advance about the shape or location of the failure surface
that arises naturally in the zones where the shear strength of soils is insufficient to resist
the shear load. In modeling failure processes the attention is usually limited to elastic
perfect plastic behavior so that hardening or softening behavior of real soils confirmed
by a number of experimental observations is excluded from the analysis. Such approach
brings us close to aforementioned limit equilibrium methods. An extensive numerical
experimentation on stability of slopes under these assumptions is reported by Griffith
[14].
The use of finite elements in geotechnical engineering, however, is much more versatile
and by no means limited to stability analysis of earth slopes. A variety of applications
in which finite elements are irreplaceable by simple methods is described in an excellent
book by Potts and Zdravkovič [20].
The objective of this manual is to summarize the theoretical grounds needed in the
analysis of geotechnical problems with GEOmkp. The work is organized as follows. Chap-
ter 2, first, briefly reviews the derivation of basic finite element equations and then outlines
in a concise manner formulation of all finite elements implemented within the GEOmkp

3
CHAPTER 1. PREFACE 4

software. Chapter 3 then summarizes several popular constitutive models often used by
geotechnical engineers such as von Mises, Drucker-Prager or Modified Mohr-Coulomb.
See also [19] for additional information on more advanced constitutive models applicable
to analysis of soft soils. The chapter concludes by several examples that show application
of perfect plastic models to the analysis of slopes. Finally, Chapter 4 addresses two most
common solution strategies for the solution of nonlinear system of equations that typically
arises when studying deformation behavior of soils.
Chapter 2

Finte Element Equations

The present section provides derivation of finite element equations governing the respective
boundary value problem. Formulation of the set of elements implemented within the
GMKP program is provided next. Throughout this section, the standard engineering
notation is used (see, e.g., [4]). Limiting our attention to plane strain conditions, see also
y

1m

z
x

Figure 2.1: Body Ω crossed by discontinuity Γd

Fig. 2.1, the stress and strain tensors written in the vector form are
© ª
σ = σxx σyy σxy σzz T , (2.1)
and © ªT
ε= εxx εyy 2εxy εzz . (2.2)
We further introduce the (3 × 4) matrix ∂ defined as
 ∂ ∂ 
∂x
0 ∂y 0
∂ =  0 ∂y ∂ ∂
∂x
0  (2.3)
∂
0 0 0 ∂z
and the (3 × 4) matrix n that stores the components of unit normal vector,
 
nx 0 n y 0
n =  0 n y nx 0  . (2.4)
0 0 0 nz

5
CHAPTER 2. FINTE ELEMENT EQUATIONS 6

2.1 Kinematics discretization

Consider a body Ω bounded by a surface Γ, Fig 2.2. Γu represents a portion of Γ with
prescribed displacements u while tractions t are prescribed on Γt (Γu ∩ Γt = ∅).
_
t
Γ
n Γt

Γu
_
u

Figure 2.2: Body Ω with boundary surface Γ

In the standard finite element method the displacement field can be interpolated over
the body Ω using the nodal shape functions [4] in the form

u(x) = N(x)a, (2.5)

where N is a matrix of standard nodal shape functions to interpolate the nodal degrees
of freedom. Introduction Eq. (2.5) into Eq. (2.8), the strain field is then expressed as

ε(x) = B(x)a, (2.6)

where B = ∂ T N is the familiar strain matrix.

2.2 Governing equations

Consider a linear elastic body Ω. Assuming small strains the linear momentum balance
equation and the kinematic equations result in

∂σ + X = 0, (2.7)

and

ε = ∂ T u. (2.8)

The vector X in Eq. (2.7) represents the vector of body forces. The traction and dis-
placement boundary conditions are given by

nσ = t on Γt , (2.9)
CHAPTER 2. FINTE ELEMENT EQUATIONS 7

and

u=u on Γu . (2.10)

The system of governing equations is usually derived by invoking the principle of

virtual work. In particular, the principle of virtual displacements can be recovered by
forcing the equations of equilibrium, Eq. (2.7), to be satisfied in average sense such that
Z Z
T
δu (∂σ + X) dΩ + δuT (−nσ + t) dΩ = 0, (2.11)
Ω Γt

for all kinematically admissible δu. Next applying Green’s theorem and taking into
account the fact that δu = 0 on Γu gives
Z Z Z
T T
δε σ dΩ = δu X dΩ + δuT t dΓ. (2.12)
Ω Ω Γt

In the context of quasistatic non-linear finite element analysis Eq. (2.12) is usually pre-
sented in its linearized form
Z Z Z
T T
δ∆ε ∆σ dΩ = δ∆u ∆X dΩ + δ∆uT ∆t dΓ, (2.13)
Ω Ω Γt

where ∆ represents an increment of a given quantity over a certain increment of time ∆t.
To proceed, we introduce with the help of Eq. (2.5) an incremental form of constitutive
equations as
∆σ = D∆ε + ∆σ in = DB∆a + ∆σ in , (2.14)
where D is the (4 × 4) instantaneous (tangent) material stiffness matrix and ∆σ in is the
increment of initial stress vector. Contribution to ∆σ in can be attributed to a number of
distinct physical sources (thermal effects, pre-stress of structural elements such as anchors,
pore pressure, etc.). Finally, introducing Eq. (2.14) into Eq. (2.13) yields
Z Z
T T T
δ∆a B DB∆a dΩ = δ∆a NT ∆X dΩ (2.15)
Ω Ω
Z Z
T T T
− δ∆a B ∆σ in dΩ + δ∆a NT ∆t dΓ,
Ω Γt

Noting that Eq. (2.15) must be satisfied for all kinematically admissible δ∆a we arrive
at the traditional form of the discrete system of linear equations

K∆u = ∆f , (2.16)

where K is the instantaneous (tangent) global stiffness matrix and ∆f represents the
generalized load vector. Individual symbols in Eq. (2.16) are provided by
Z
K = BT DB dΩ, (2.17)
ZΩ Z Z
∆f = NT ∆X dΩ − BT ∆σ in dΩ + NT ∆t dΓ. (2.18)
Ω Ω Γt
CHAPTER 2. FINTE ELEMENT EQUATIONS 8

Following the standard finite element procedure, the stiffness matrix is obtained by
the assembly of contribution from individual elements. To that end, the domain Ω is de-
composed into Ne non-intersecting elements Ωe such that Ω = ∪N e=1 Ωe . Formally writing,
e

the global stiffness matrix and the global force vector become

K = AN
e=1 Ke ,
e
(2.19)
∆f = AN
e=1 ∆f e .
e
(2.20)

2.3 Finite elements used in GMKP program

The following section provides a brief overview of individual finite elements used in the
GMKP finite element code. The available elements can be divided into two groups: two-
dimensional plane strain elements (3-node and 6-node triangular elements) and special
elements such as 2-node rod element to model anchors, 2-node and 3-node beam elements
to model supporting walls, tunnel linings or foundation and 4-node and 6-node interface
elements to model relative movement of the structure with respect to the soil. All el-
ements implemented in GMKP are constructed within the framework of isoparametric
formulation, which means that the same interpolation functions are used to approximate
geometry as well as displacement field.

2.3.1 2-node rod element

The 2-node rod element with the linear interpolation of the displacement field is shown
in Fig. 2.3.

u l2

x l ul


yg vg 2 N1 N2
yl





u l1 α −1 1 r




1
Isoparametric mapping

x g ug

Figure 2.3: 2-node rode element

CHAPTER 2. FINTE ELEMENT EQUATIONS 9

Kinematics
The local displacement ul written in terms of the global degrees of freedom
ae = {u1 , v1 , u2 , v2 }T reads

ul = N1 (u1 cos α + v1 sin α) +N2 (u2 cos α + v2 sin α), (2.21)

| {z } | {z }
ul1 ul2

where the isoparametric element shape functions N1 , N2 are given by

1
N1 = (1 − r),
2
1
N2 = (1 + r).
2

Element stiffness matrix

Taking the derivative of Eq. (2.21) with respect to xl gives the axial strain in the form

ε = Ba, (2.22)

where the (1 × 4) matrix B attains the following form

1
B= {− cos α, − sin α, cos α, sin α}. (2.23)
L
where L is the element length. To conclude the derivation of the element stiffness matrix
we introduce the constitutive law in the form
EA
σ = Dε = ε, (2.24)
L
where D = EA/L represent the element axial stiffness; E, A are Young’s modulus and
element cross-sectional area respectively. Finally, making use of Eq. (2.17) on the element
level provides the element stiffness matrix Ke as
 
cos α cos α cos α sin α − cos α cos α − cos α sin α
EA  cos α sin α sin α sin α − cos α sin α − sin α sin α  .
Ke = (2.25)
L  − cos α cos α − cos α sin α cos α cos α cos α sin α 
− cos α sin α − sin α sin α cos α sin α sin α sin α

2.3.2 2-node and 3-node beam elements

The 2-node and 3-node beam elements implemented in GMKP appear in Fig. 2.4.
CHAPTER 2. FINTE ELEMENT EQUATIONS 10

Y, v Global elements

            
y  u 2 v 2 ϕ 2

           
              
  
  
  

             
  
  
  
  
   
  



 
  
 

 

 

 

 




 

  
  
  
  
  x 
u  v  ϕ        

  

 
  
  

 
  
  

 
  
  

 
  
  

 
  
  

 
  
  

  
  

  
  



 
      

  
 y      

  
      

  
      

  
   














































u 


2 
   
v 
   
2 


ϕ 


2 








 
           α   
 

 

 

 

 


2 

 

 

 
     
     
     
     
     
   
 
  
 
  
 
 u  
  3 
 
  v   
  3ϕ
 
 
 3
 
   
 
  
 
  
 
  
 
  
 
    
 
    
 
  
 
  
 
  
 
  
 
 x 

1 1 1
         
 
    
    
    
    
    
 ϕ   
   
   
   
       
   
   
   
   
   
   
   
   
   
   
   
   
   
   
 

 

  u 

  1  
 v 

  1  
 

  1  
 

  

  
 

  3 
  

  

  

  

  

  

  

  2 

  

  

  

  

  
1 



  
  
  
  
  
  
  
  
 α 
  
  
  
  
  
  
  
  
  
  
  
  
  
   

1
ϕ L

Z 
X, u Parent elements
Integration points
     
0 1 r −1 0 1 r

Figure 2.4: 2-node and 3-node beam elements

Kinematics
In general, assuming the local cartesian coordinates x, y to be aligned with the principal
axes of inertia and placed in the origin of the centroid of the beam cross-section provides
the constitutive equations in the form
 
 du 
    


 dx 

 Nx  EA 0 0  
Mz =  0 EIz 0  dϕ z , (2.26)
   − 
Qy 0 0 kGA   dx 

 
 −ϕz + dv 
 
dx
where EA, EIz , kGA represent the axial, bending and shear stiffnesses of the beam, re-
spectively. Individual entries in the resultant stress vector σ = {Nx , Mz , Qy } T represent
the normal force, bending moment and the shear force developed in the beam cross-
section. Finally, the unknown functions in the displacement field u = {u, ϕ, v} T stand
for the longitudinal displacement, rotation about the z-axis and vertical displacement, re-
spectively. They follow from the standard finite element approximation using the element
shape functions and the nodal degrees of freedom

u = Na. (2.27)

2-node beam element: Detailed derivation of the finite element matrices for the 2-node
beam element implemented in GMKP is given in [4]. Here we present only the most
CHAPTER 2. FINTE ELEMENT EQUATIONS 11

Table 2.1: Shape functions for 2-node beam element

Node i Function Ni

1 1−r

2 r

1 £ ¤
3 6r − 6r2
L(1 + 2κ)
1 £ ¤
4 (1 + 2κ) − 2(2 + κ)r + 3r 2
1 + 2κ
1 £ ¤
5 −6r + 6r2
L(1 + 3κ)
1 £ ¤
6 −2(1 − κ)r + 3r 2
1 + 2κ
1 £ ¤
7 (1 + 2κ) − 2κr − 3r 2 + 2r3
1 + 2κ
L £ ¤
8 −(1 + κ)r + (2 + κ)r 2 − r3
1 + 2κ
1 £ ¤
9 2κr + 3r2 − 2r3
1 + 2κ
L £ ¤
10 κr + (1 − κ)r 2 − r3
1 + 2κ

essential part. In particular, the matrix N in Eq. (2.27) assumes the form
 
N1 0 0 N2 0 0
N =  0 −N3 N4 0 −N5 N6  , (2.28)
0 N7 −N8 0 N9 −N10
where individual shape functions are listed in Table 2.1. The variable κ that appears in
individual terms of the shape functions is given by
6EIz
κ= ,
kGAL2
where k is the shear correction factor and L is the length of the beam. The finite element
representation of the strain field
½ ¾
du dϕz dv
ε= ,− , −ϕz + , (2.29)
dx dx dx
CHAPTER 2. FINTE ELEMENT EQUATIONS 12

Table 2.2: Shape functions for 3-node beam element

Node i Regular function Ni Substitute function N i

1 1 1
1 2
r(r − 1) (
2 3
− r)

1 1 1
2 2
r(r − 1) (
2 3
+ r)

2
3 (1 − r2 ) 3

calls for introduction of the strain matrix B. Using Eq. (2.28) and taking into account
the transformation of coordinates from the local to the global coordinate system it is easy
to see that
 
CN10 SN10 0 CN20 SN20 0
B= −SN30 CN30 −N40 −SN50 CN50 −N60 ,
0 0 0 0 0 0
−S(N3 + N7 ) C(N3 + N7 ) −N4 − N8 −S(N5 + N9 ) C(N5 + N9 ) −N6 − N10
(2.30)
where

C = cos(α),
S = sin(α), (2.31)
1 dNi
Ni0 = ,
J dr
The angle α in the above equation is defined in Fig. 2.4 and the Jacobian J follows from
Eq. (2.35).

3-node beam element: Assuming the standard isoparametric shape functions listed in
Table 2.2 to approximate the displacement field gives the matrix N in the form
 
N1 0 0 N2 0 0 N3 0 0
N =  0 N1 0 0 N2 0 0 N3 0  , (2.32)
0 0 N1 0 0 N2 0 0 N3

Next, recall the representation of the strain field (2.29) and use Eq. (2.32) to arrive at
 
CN10 SN10 0 CN20 SN20 0 CN30 SN30 0
B= 0 −N10 0 0 − N20 0 −N30 0 . (2.33)
0 0 0 0 0 0
−SN1 CN1 N 1 −SN2 CN2 N 2 −SN3 CN3 N 3

The standard B matrix was again augmented to account for the transformation of coor-
dinates. Parameters C, S, Ni0 receive the same meaning as in Eq. (2.31) with the Jacobian
CHAPTER 2. FINTE ELEMENT EQUATIONS 13

J found from Eq. (2.36). In addition, substitute shape functions N i were used to define
dv
the variation of ϕ in the definition of shear strain −ϕz + to avoid shear force locking.
dx
Note that the substitute shape functions coincide with the regular shape functions at the
reduced Gaussian integration points. Details can be found in [10, 19].

Element stiffness matrix

Derivation of the stiffness matrix follows Eq. (2.17). The result is
N
X
Ke = wj BT (rj )DB(rj )J, (2.34)
j=1

where the Jacobian J reads

J = L for 2 − node element, (2.35)

L
J = for 3 − node element. (2.36)
2
Locations of integration points within parent elements are stored in Table 2.3.

Table 2.3: Integration points for 2-node and 3-node beam elements

Integration 2-node beam 3-node beam

point coordinate r weight coordinate r weight

1 0.211324865 1.0 -0.774596669241483 5/9

2 0.788675131 1.0 0.0 8/9
3 0.774596669241483 5/9
CHAPTER 2. FINTE ELEMENT EQUATIONS 14

3




3


y, v Global elements








1 1


 

4
 

2 2
x, u Parent elements
s s
3 (0,1) 3 (0,1)






integration 3
points 5


6 6




1 7 5
!

!
#

"#
r 


1 4 


2 


r


 

"

1(0,0) 2(1,0) 1(0,0) 4 2(1,0)

Figure 2.5: 4-node and 6-node interface elements

2.3.3 Plane 3-node and 6-node triangular elements

Plane 3-node and 6-node isoparametric triangular elements are implemented in GMKP.
Geometry of both elements is evident from Fig. 2.5.

Kinematics
The displacement interpolation functions are listed in Table 2.4. The element degrees of
freedom are
a = {u1 , v1 , u2 , v2 , u3 , v3 } T 3 − node elem, (2.37)
T
a = {u1 , v1 , u2 , v2 , u3 , v3 , u4 , v4 , u5 , v5 , u6 , v6 } 6 − node elem. (2.38)
The displacement field inside the element is uniquely described by the above nodal pa-
rameters n n
X X
u= N i ui , v= Ni vi , (2.39)
i=1 i=1
where n is the number of element nodes.

Element stiffness matrix

The components of the strain tensor follow from Eq. (2.6). The element stiffness matrix is
then defined by Eq. (2.17). Here, the integral is again evaluated by Gaussian quadrature
CHAPTER 2. FINTE ELEMENT EQUATIONS 15

Table 2.4: Interpolation functions for 3-node and 6-node triangular elements

Node Function Included only if node i

i Ni is defined

i=4 i=5 i=6

1 1
1 1−r−s − N4 − N6
2 2
1 1
2 r − N4 − N5
2 2
1 1
3 s − N5 − N6
2 2

4 4r(1 − r − s)

5 4rs

6 4s(1 − r − s)
CHAPTER 2. FINTE ELEMENT EQUATIONS 16

Table 2.5: Integration points for 3-node triangular element

Integration Coordinate Coordinate Weight

point r s

1 1/3 1/3 1/2

Table 2.6: Integration points for 6-node triangular element

Integration Coordinate Coordinate Weight

point r s

1 0.1012865073235 0.1012865073235 0.1259391805448

2 0.7974269853531 0.1012865073235 0.1259391805448
3 0.1012865073235 0.7974269853531 0.1259391805448
4 0.4701420641051 0.0597158717898 0.1323941527885
5 0.4701420641051 0.4701420641051 0.1323941527885
6 0.0597158717898 0.4701420641051 0.1323941527885
7 0.3333333333333 0.3333333333333 0.2250000000000
CHAPTER 2. FINTE ELEMENT EQUATIONS 17

so that
N
X
Ke = wj BT (rj , sj )DB(rj , sj )J(rj , sj ), (2.40)
j=1

where wi is the weight for a given integration point i, N is the number of integration
points and J is the Jacobian of the transformation given by
∂x ∂y ∂x ∂y
J(r, s) = − . (2.41)
∂r ∂s ∂s ∂r
The linear 3-node element is integrated at 1 integration points, while N = 7 is assumed
for the quadratic 6-node element, see Fig. 2.5. Locations of integration points within
parent elements are stored in Tables 2.5 and 2.6. Further details on the evaluation of the
element stiffness matrix can be found in [4].

Element load vector

In this section, we turn our attention to the evaluation of element load vector attributed
to the gravity loading, pore pressure loading and loading that arise during excavation.
Gravity loading: The forces generated by the self weight of the soil follow from the first
term on the right hand side of Eq. (2.18) and are given by
Z
gr
fe = NT X γe dVe , (2.42)
Ve

where X γe = {0, γe } T and γe is the element self-weight per unit volume. The integral in
Eq. (2.42) thus redistributes the net vertical force to all element nodes.
Pore pressure: To arrive at the element load vector due to the prescribed pore pressure
we first recall the concept of effective stresses. The total stress vector then assumes the
form
σ = Dε − 3mp, (2.43)
where p > 0 represents the prescribed liquid pore pressure in the fully saturated soil.
Matrix D now stands for the stiffness matrix of the porous skeleton and vector m is
introduced later in Section 3.4.1 Eq. (3.4). Also not that the matrix of the solid phase of
the porous skeleton is taken as rigid (undeformable). Eq. (2.14) now becomes

σ = BDa − Np, (2.44)

where vector p stores the nodal values of the prescribed pore pressure. Introducing
Eq. (2.44) into Eq. (2.15) then gives (recall the second term on the right hand side of
Eq. (2.18)) Z
f pp
e = − BT Ne p dVe . (2.45)
Ve

Excavation problem: When a portion of material is excavated (open excavation, tunneling)

forces must be applied along the excavated surface such that the remaining material
CHAPTER 2. FINTE ELEMENT EQUATIONS 18

experiences the correct unloading effect and the new free surface is stress free [25, 5].
The excavation procedure is schematically displayed in Fig. 2.6. Suppose that prior to
excavation the material in the original body is loaded to attain the initial stresses σ A0 in
zone A and σB0 in zone B, respectively. This initial stress state can be recovered as a
superposition of two loading stages. To that end, suppose that the material from zone A is
removed. To maintain the initial stress state σB0 developed in zone B the new free surface
must be loaded by forces FAB exerted by body A on to body B. Similarly, the forces FBA
having the same magnitude but opposite direction as forces FAB must be applied to body
A to comply with the equilibrium requirements. It now becomes evident that in order to
complete the excavation procedure the unwanted layer of forces FAB must be removed by
applying force FBA to body B thus arriving at the required stress free surface, Fig. 2.6.

t
FBA
FAB
Xγ
A σA
0

B
σB
0

FBA

σB + σB
0

Figure 2.6: Excavation process and excavation forces

In mathematical terms the excavation forces FBA follow from the principal of virtual
work written as
Z Z Z
T T γ
δε σ A0 dΩ = δu X dΩ + δuT t dΓ. (2.46)
ΩA ΩA ΓA

After discretization Eq. (2.46) becomes

µ Z Z ¶
NeA T T T γ
Ae=1 δae B σ A0 e dVe − δae N X e dVe = δaAB T F BA .
T
(2.47)
Ve Ve

The final step requires to relate the element nodal degrees of freedom to the degrees of
freedom associated with the nodes on the new free surface. This can be done with the
help of the localization matrix L such that

ae = Le aBA . (2.48)
CHAPTER 2. FINTE ELEMENT EQUATIONS 19

zoom
A 1 1
Fba Fba
1a 1b

1b
zoom
2 2
1a Fba Fba

Figure 2.7: Localization

Plugging back from Eq. (2.48) into Eq. (2.47) finally gives the vector of excavation forces
in the form
µZ Z ¶
NeA T T T T γ
F BA = Ae=1 Le B σ A0 e dVe − Le N X e dVe . (2.49)
Ve Ve

The localization procedure as shown in Fig. 2.7 essentially corresponds to the selection
of elements attached to the excavation surface. Thus the remaining elements present in
body A do not have to be taken into account when computing the excavation forces F BA
in Eq. (2.49).
The theoretical grounds set in the above paragraphs are demonstrated on a simple
problem of open excavation. Figs. 2.8-2.10 illustrate a sequence of computational tasks
presented in Fig. 2.6. The initial state prior to excavation is represented by Fig. 2.8
showing a variation of the horizontal stress due to pure gravity loading. Fig. 2.9 then
corresponds to the state found in body B on Fig. 2.6 after removing the soil from open
cut but keeping the original stresses by applying forces FAB along new free boundaries.
Finally, Fig. 2.10 displays the final deformation and stress state after removing the un-
wanted forces FAB with the help forces FBA that have the same magnitude but opposite
direction, see Fig. 2.6 showing body B after completing the excavation sequence.

2.3.4 4-node and 6-node interface element

This section presents derivation of the element stiffness matrix for the 4-node and 6-
node interface elements that are compatible with 3-node and 6-node triangular elements,
respectively, also implemented within the GMKP finite element code. Both elements are
displayed in Fig. 2.11.

Kinematics
In the finite element framework the global displacements are approximated using the stan-
dard element shape functions listed in Table 2.7. Referring to Fig 2.11 the displacement
CHAPTER 2. FINTE ELEMENT EQUATIONS 20

Figure 2.8: Excavation procedure – initial state

Figure 2.9: Excavation procedure – intermediate state

Figure 2.10: Excavation procedure – final state

CHAPTER 2. FINTE ELEMENT EQUATIONS 21

y l vl 4 y l vl 5

  xl u l


 xl u l

α
2

6 α
2

Top Top 


3 isoparametric 4 3 isoparametric
y g vg  Bottom mapping  Bottom mapping
 
1 1
N1 N3 N2
N1 N2
1  

2 2
 1
−1 1 r −1 3 1 r

xg ug

Figure 2.11: 4-node and 6-node interface elements

Table 2.7: Interpolation functions for 4-node and 6-node interface elements

Node Function Included only if

i Ni node 3 is defined

1 1
1 (1 − r) − N3
2 2
1 1
2 (1 + r) − N3
2 2

3 (1 − r2 )

field for the 4-node interface element receives the form

utop = N 1 u3 + N 2 u4 , (2.50)
ubot = N 1 u1 + N 2 u2 ,
v top = N 1 v3 + N 2 v4 ,
v bot = N 1 v1 + N 2 v2 ,

In the compact form the global nodal degrees of freedom ui , vi are

ag = {u1 , v1 , u2 , v2 , u3 , v3 , u4 , v4 } T . (2.51)
CHAPTER 2. FINTE ELEMENT EQUATIONS 22

Similarly for the 6-node interface element we get

utop = N 1 u4 + N 2 u5 + N 3 u6 , (2.52)
ubot = N 1 u1 + N 2 u2 + N 3 u3 ,
v top = N 1 v4 + N 2 v5 + N 3 v6 ,
v bot = N 1 v1 + N 2 v2 + N 3 v3 ,

and
ag = {u1 , v1 , u2 , v2 , u3 , v3 , u4 , v4 , u5 , v5 , u6 , v6 } T . (2.53)

Element stiffness matrix

The stress-displacement relationship of the interface model assumes the form
½ ¾ ½ ¾
τ [[u]]l
= D , (2.54)
σ [[v]]l

where [[u]]l and [[v]]l represent the relative displacements of the top and bottom of the
interface element in the local coordinate system, Fig. 2.11. For isotropic linear elastic
behavior the interface material stiffness matrix D takes the form
· ¸
Ks 0
D= , (2.55)
0 Kn

where Ks and Kn are the elastic shear and normal stiffnesses, respectively. They can be
related to the interface shear and Young’s moduli Gint , Eint as
Gint
Ks = ,
t
Eint
Kn = ,
t
where t represents the interface stiffness. It should be noted here that setting the inter-
face stiffnesses Ks , Kn to low values may lead to excessively large elastic displacements.
However, if the elastic parameters are too large (attempt to model a perfect bond), then
the numerical ill-conditioning may occur. This is usually manifested by the oscillation
interface stresses. It has been argued that such unwanted oscillatory behavior can be
reduced by using the Newton-Cotes integration scheme (integration points coincide with
the element nodes) when computing the element stiffness matrix [9, 15]. This integration
scheme is employed in GMKP. On the contrary, the results presented in [11] suggest that
the use of Newton-Cotes integration scheme has no benefit over the Gaussian quadrature.
The global degrees of freedom in Eqs. (2.51) and (2.53) are related to local displacement
jumps in the form ( )
[[u]]l
= Bag , (2.56)
[[v]]l
CHAPTER 2. FINTE ELEMENT EQUATIONS 23

where the matrix B assumes the form

B = [−TB1 − TB2 TB1 TB2 ] 4 − node elem, (2.57)

B = [−TB1 − TB2 − TB3 TB1 TB2 TB3 ] 6 − node elem. (2.58)

and · ¸ · ¸
cos α sin α Ni 0
T= Bi = . (2.59)
− sin α cos α 0 Ni
The element stiffness Ke then follows from
Z
L 1 T
Ke = B DB dr, (2.60)
2 −1

where L is the element length.

Element load vector

We limit our attention to the element load due to pore pressure. Assuming drained
boundary conditions and using the concept of effective stresses gives the vector of total
stresses in the form ½ ¾ ½ ¾ ½ ¾
τ £ ¤ [[u]]l 0
= D − p, (2.61)
σ [[v]]l 1
where the average value of pore pressure p is given by
1
p = [(p1 + p3 )N1 + (p2 + p4 )N2 ] 4 − node elem, (2.62)
2
1
p = [(p1 + p4 )N1 + (p2 + p5 )N2 + (p3 + p6 )N3 ] 6 − node elem. (2.63)
2
Finally, recall the second term in Eq. (2.18) to get the element load vector f e due to
initial pore pressure as Z ½ ¾
L 1 T 0
fe = B p d r. (2.64)
2 −1 1
Chapter 3

Constitutive models

3.1 Constitutive model for anchors

An important topic which needs to be addressed is concerned with the methods of sup-
porting walls (props, ties, anchors) or reinforcing the soil body (geotextiles). A two-node
rod element with an axial stiffness but with no bending stiffness can be used to model hor-
izontal struts, ranking struts, geotextiles or node-to-node anchors. In general the element
is used to model ties between two points in space. In the current implementation, such
an element is considered to be a free element not necessarily connected to the underlying
finite element mesh. Its deformation is realized by tying the free element nodes to the
active nodes in the element mesh. This step is done automatically. This element can
be subjected to both tensile forces (anchors) and compressive forces (struts). Allowable
limits on the tensile as well as compressive force can be set to simulate element failure,
e.g., anchor tensile failure or compressive failure of props due to buckling.

EA/L
σ 1 σ
Rt
σ in

ε ε

Rc

Figure 3.1: Constitutive model for anchors

The material behavior is limited to linear elasticity up to failure as shown in Fig. 3.1.
The corresponding incremental stress-strain relationship assumes the form
EA
∆σ = ∆ε + ∆σin , (3.1)
L

24
CHAPTER 3. CONSTITUTIVE MODELS 25

where E is Young’s modulus, A represents the element cross-sectional area and L is the
element length; ∆σin stands for the initial stress increment associated with the initial
anchor pre-stress given by
∆Fpre
∆σin = , (3.2)
A
where Fpre is the applied pre-stress force. Values Rt and Rc in Fig 3.1 correspond to
allowable tensile and compressive strengths, respectively. Note that care must be taken
when computing the axial element stiffness EA/L in plane strain applications as it should
represent an equivalent axial stiffness per unit length that takes into account the element
spacing in the out of plane direction.
As a default setting the element is assumed to sustain zero compressive stress. Thus,
it is deactivated when in compression. But it can be reactivated again when subjected
to tension. However, when the allowable strength limits are exceeded the element is
automatically removed from the analysis.

3.2 Elastic constitutive model for soil

Although not realistic for soils the program makes possible to analyze a purely elastic
isotropic material. A typical stress-strain curve for a linear elastic material is plotted in
Fig. 3.2. Note that such a model assumes that the loading and unloading branch coincide.

σ
E = Eur
1
loading/unloading

Figure 3.2: Linear elastic constitutive model for soil

The assumption of isotropy (all material constants are independent of the orientation
of coordinate axis) is common for all material models implemented in GEOmkp. In such a
case there are only two independent material constants necessary to represent the material
behavior, e.g., the Young modulus E and Poisson ratio ν. Limiting our attention to the
plane-strain description, the incremental constitutive equation for linear elastic isotropic
CHAPTER 3. CONSTITUTIVE MODELS 26

material read
     

 σxx 
 1−ν ν 0 ν 
 εxx 

  E  ν   
σyy  1−ν 0 ν  εyy
=  (3.3)

 τxy 
 (1 + ν)(1 − 2ν) 0 0 1−2ν
2
0   γ xy 

   
σzz ν ν 0 1−ν εzz

3.3 Modified elastic constitutive model for soil

A slightly more realistic prediction of the material behavior can be achieved when as-
suming different material response in loading and unloading. Such an approach calls for
a third independent material parameter governing the unloading-reloading branch of the
stress strain curve as shown in Fig. 3.3. An experimental evidence suggests to set the value
of the unloading-reloading Young’s modulus Eur approximately three times the value of
E. This is the default setting in GEOmkp.

σ
E
1
loading
Eur

1
unloading
ε

Figure 3.3: Modified linear elastic constitutive model for soil

3.4 Elasto-plastic constitutive models for soil

One of the key topics in the geomechanical engineering is an assessment of stability and
ultimate load bearing capacity of soils. Although a number of simple approaches based
on limit equilibrium is available for the solution of this problem such as the Petterson,
Bishop or Sarma methods, there is an increasing need for more accurate and reliable
approaches that take the actual behavior of soils into account, especially in applications
involving soils that show softening behavior, e.g., dense sands or overconsolidated clays,
see Fig. 3.4.
The purpose of this chapter is to review several constitutive models generally known
to geotechnical engineers. The chapter starts with the classical von Mises model [4, 19]
often called when assuming the total stress approach. Although this model is not currently
implemented within the GEOmkp software, we take advantage of its simplicity and use
CHAPTER 3. CONSTITUTIVE MODELS 27

σ ET
E
1
1
loading
Eur

unloading
ε
Elastic Hardening Softening

Figure 3.4: Typical soil behavior involving hardening and softening

it as a stepping stone for more complex constitutive models such as the Drucker-Prager
and modified Mohr-Coulomb models discussed next. Note that unlike the von Mises
constitutive model these models draw on the use of effective parameters.

3.4.1 Invariants
Before proceeding with the actual formulation of individual constitutive models we first
define the following matrices and vectors extensively used in this chapter:

T
m = {1/3, 1/3, 0, 1/3} ,
   
2/3 −1/3 0 −1/3 1 0 0 0
   
   
−1/3 2/3 0 −1/3 0 1 0 0 
   
P = 

,
 Q = 

,
 (3.4)
 0 0 2 0  0 0 1/2 0
   
   
−1/3 −1/3 0 2/3 0 0 0 1

with the following useful connections

T T 1
m Qm = m m = , PQP = P, PQm = 0. (3.5)
3
In addition, standard engineering representation of stress and strain is used throughout
the text. Assuming the Cartesian coordinate system with axes x1 , x2 and x3 , and the
plain strain state of stress the symmetric second order Cauchy stress tensor σ is then
represented as (4 × 41) vector,
T
σ = {σ11 , σ22 , σ12 , σ33 } . (3.6)
CHAPTER 3. CONSTITUTIVE MODELS 28

Similarly we write the symmetric second-order tensor of small strains ε in the form
T
ε = {ε11 , ε22 , γ12 , ε33 } . (3.7)

The deviatoric counterparts of stress σ and strain ε are then given by

s = PQσ, e = PQε. (3.8)

The state of stress at a given material point provided by Eq. (3.6) can be also written
in terms of three basic invariants. Assuming standard elasticity notation (pressure is
negative) these invariants can be written with the help of Eq. (3.4) as
T
σm = m σ, (3.9)
r
1 T
J = σ Pσ, (3.10)
2 Ã !
√
1 3 3 I3s
θ = − arcsin , (3.11)
3 2 J3

where σm is the effective mean stress, J is defined as a square root of the second invariant
of the deviatoric stress and θ is the Lode’s angle; I3s in Eq. (3.11) stands for the third
invariant of the deviatoric stress. Assuming plane strain conditions the quantity I 3s takes
the following form
I3s = s11 s22 s33 − s33 s12 s12 , (3.12)
where sij are the components of the deviatoric stress tensor given by Eq. (3.8). It becomes
also advantageous, later in this chapter, to define certain equivalent measures of strain
vectors ε and e as
r
2 T
εeq = ε Qε , (3.13)
3
r
2 T
γeq = ε QPQε . (3.14)
3

3.4.2 Yield surface

It is the well known fact that the plastic behavior of solids in general is characterized
by a non-unique stress-strain relationship. Evidence of such a behavior is the presence
of irrecoverable (plastic) strains (ε) upon unloading. Fig. 3.5 provides an illustrative
example of uniaxial behavior of a material loaded beyond the elastic limit. The plastic
strains attributed to yielding, however, can occur only if the stresses σ satisfy a certain
yield criterion
F (σ, κ) = 0, (3.15)
T
where the components of vector κ = {κ1 , κ2 . . .} are called the hardening/softening pa-
rameters. If the material exhibits hardening/softening, the surface described by Eq. (3.15)
CHAPTER 3. CONSTITUTIVE MODELS 29

σ
Loading

Unloading

ε ε
pl

Figure 3.5: Uniaxial plastic behavior

expands/contracts depending on the loading history and the hardening/softening param-

eters κ = κ(t). The condition F (σ, κ) < 0 then corresponds either to initial elastic
loading, or elastic unloading from a previously reached plastic state. In a time interval
during which the material remains in plastic state, the yield condition Eq. (3.15) is satis-
fied (recall that F (σ, κ) > 0 is not allowed). After differentiating the yield condition we
arrive at the so called consistency condition [4]
³ ∂F ´T ³ ∂F ´T
dF = dσ + dκ = 0, (3.16)
∂σ ∂κ
where, under plane strain state of stress,
½ ¾T
∂F ∂F ∂F ∂F ∂F
= , , , , (3.17)
∂σ ∂σ11 ∂σ22 ∂σ12 ∂σ33
½ ¾T
∂F ∂F ∂F ∂F
= , ,..., .
∂κ ∂κ1 ∂κ2 ∂κk
³ ´T
The material constants can be chosen such that for loading ∂∂F κ dκ < 0. The consis-
tency condition then provides the following loading criterion:

³ ∂F ´T 
> 0 plastic loading,
T ∂F
dσ = dσ = 0 neutral loading, (3.18)
∂σ ∂σ  
< 0 elastic loading.

Apart from the yield condition, Eq. (3.15), the description of the plastic deformation
requires a certain assumption about the direction of the plastic flow. Such hypothesis is
called the flow rule. It is quite generally postulated that the plastic flow will occur in the
direction normal to a plastic potential surface
G = G(σ, κ). (3.19)
CHAPTER 3. CONSTITUTIVE MODELS 30

The plastic strain increment can be then defined as

∂G
dεpl = dλ , (3.20)
∂σ
where dλ is a proportionality constant, as yet undetermined. The special case of G = F
is known as associated plasticity and the flow rule Eq. (3.19) is termed the normality rule.
In general case, however, when G 6= F the plasticity is non-associated. Parameter dλ can
be eliminated using consistency condition Eq. (3.16) written in Melan’s form as
³ ∂F ´T
dF = dσ − Hdλ = 0, (3.21)
∂σ
where ³ ∂F ´T ∂κ ³ ∂F ´T
H=− valid for dσ > 0, (3.22)
∂κ ∂λ ∂σ
is the modulus of plastic hardening/softening.

3.4.3 Elasto-plastic stiffness matrix

This section completes the summary of general theoretical grounds needed for the de-
scription of plastic behavior of solids by formulating the elasto-plastic material stiffness
matrix. Two approaches are examined to arrive at the desired result. The first approach
draws on standard incremental form of constitutive equations combined with consistency
condition given by Eq. (3.16), while the second approach confirms with the actual al-
gorithmic procedure for the stress update. The interested reader may also consult an
excellent paper on this subject by Simo and Taylor [24].

Standard tangent stiffness matrix

Assuming additive decomposition of small strains the total strain vector admits the fol-
lowing representation
dε = dεel + dεpl . (3.23)
Relating the increment of stress dσ during plastic loading to the elastic part of the total
strain dεel we get
dσ = Del ( dε − dεpl ). (3.24)
where Del represents, keeping up with the plane strain conditions, the (4 × 4) elastic
stiffness matrix. Substituting for dεpl from Eq. (3.20) into Eq. (3.24) gives
µ ¶
el ∂G
dσ = D dε − dλ . (3.25)
∂σ

After substituting for dσ from Eq. (3.25) into consistency condition Eq. (3.21) we get
µ ¶T µ ¶
∂F el ∂G
D dε − dλ − Hdλ = 0. (3.26)
∂σ ∂σ
CHAPTER 3. CONSTITUTIVE MODELS 31

Solving for dλ from Eq. (3.26) yields

³ ´T
∂F
Del dε ∂σ
dλ = ³ ´T ³ ´. (3.27)
H + ∂ σ D ∂∂G
∂F el
σ

Back substitution of dλ into Eq. (3.24) then results in

 ³  ´T
D el µ ¶ ∂F
 ∂G  ∂σ
dσ = Del − ³ ´T ³ ´ ∂σ  dε. (3.28)
H + ∂∂F el ∂G
σ D ∂σ

Eq. (3.28) represents an explicit expansion which determines the stress change in terms of
strain change with the instantaneous (elasto-plastic) tangent stiffness matrix D ep written
as

dσ = Dep dε (3.29)
³ ´³ ´T
D elDel ∂G
∂σ
∂F
D el ng nT Del
∂σ
ep el el
D = D − ³ ´T ³ ´ = D − , (3.30)
∂F el ∂G H + nT Del ng
H + ∂σ D ∂σ

where
∂F ∂G
n=, ng = , (3.31)
∂σ ∂σ
represent normals to the yield and potential surfaces in the stress space, respectively.

Algorithmic tangent stiffness matrix

It has been confirmed that the use of the elasto-plastic stiffness matrix given by Eq. (3.30)
spoils the quadratic convergence of the Newton-Raphson iterative solver. Simo and Tay-
lor [24] showed that in order to cure this, the elasto-plastic stiffness matrix must be
consistent with the algorithmic procedure used for the stress update. Thus, following the
standard predictor-corrector stress update procedure [24, 18, 6, 23, to cite a few] we write
the incremental form of the constitutive equation follows

σ i − σ i−1 = Del (εi − εi−1 ) − Del (λi − λi−1 )nig , (3.32)

Taking the time derivative of Eq. (3.32) yields

dσ = Del dε − dλDel ng − ∆λDel dng , (3.33)

with
∂ng
dng = dσ. (3.34)
∂σ
CHAPTER 3. CONSTITUTIVE MODELS 32

Introducing Eq. (3.34) into Eq. (3.33) we get after some manipulation
∂ng
(Del )−1 dσ = dε − dλng − ∆λ dσ, (3.35)
∂σ
· ¸
el −1 ∂ng
(D ) + ∆λ dσ = dε − dλng , (3.36)
∂σ
· ¸−1
el −1 ∂ng
dσ = (D ) + ∆λ ( dε − dλng ), (3.37)
∂σ
| {z }
D
· ¸−1
el −1 ∂ng
D = (D ) + ∆λ . (3.38)
∂σ
Writing the consistency condition Eq. (3.16) in the rate form
³ ∂F ´T
dF = dσ − H dλ = 0, (3.39)
∂σ
provides
1 ³ ∂F ´T
T
n dσ
dλ = dσ = , (3.40)
H | ∂σ{z } H
nT
With the help of Eq. (3.40) Eq. (3.37) can be inverted to get
" T
#
n n g
dε = D−1 + dσ. (3.41)
H

Finally, applying the Sherman-Morrison formula written in its general form as

B = I + rN T , (3.42)

rN T
B−1 = I − , (3.43)
1 + rTN
to Eq. (3.41) results in the desired algorithmic tangent stiffness matrix
" T
#
Dng n D
dσ = D − dε, (3.44)
H + nT Dng
| {z }
Dcons
T
Dng n D
Dcons = D − . (3.45)
H + nT Dng
Note that using the above form of the instantaneous elasto-plastic stiffness matrix (D cons )
instead of Dep in the iterative solver maintains the theoretically proved quadratic conver-
gence of the Newton-Raphson method.
CHAPTER 3. CONSTITUTIVE MODELS 33

Tresca σ1
J
0
Mises θ = + 30

F=G
θ
0
θ = − 30

σ3 σ2
σm

Figure 3.6: Von Mises failure surface

3.4.4 Von Mises model

We begin our exposition to selected constitutive models with the Mises model primarily
due to its simplicity. In geotechnical applications this model often serves as a substitute
for the Tresca model [19]. Recall that the Tresca model is plotted as hexagonal cylinder
in the principal stress space. The corners of the Tresca failure surface imply singularities
in the yield function which have in the past caused havoc when attempting to use the
yield function for obtaining analytical solutions to simple boundary values problems. The
Mises failure surface on the other hand is smooth and plots as an infinite cylinder in
the principal stress space. Its projections into the deviatoric and meridian planes are
displayed in Fig. 3.6. This figure further justifies the use of the Mises failure surface in
place of the Tresca model. It is also worth noting that in case of Mises or Tresca model the
onset of the yielding does not depend on the volumetric part of the stress tensor (mean
stress σm ), see Fig. 3.6.

Yield surface
Two yield surfaces are examined with the von Mises model. Assuming the undrained
conditions the von Mises yield function can be written in terms undrained shear strength
Su as
Su (κ)
F (σ, κ) = J − , (3.46)
cos θ
where J follows from Eq. (3.10). The vector of hardening parameters now reduces to to
pl
a single parameter κ related to an equivalent shear strain such that κ = γeq . See also
Fig. 3.7 showing the bilinear form hardening law written in terms of shear strength S u
and equivalent shear strain γeq specified by Eq. (3.14); θ is a given value of the Lode’s
angle. In particular, setting θ = ±300 results in a yield surface circumscribed to the
Tresca model, see Fig. 3.6.
Dependance of the shear strength on the hardening/softening parameter κ calls for the
derivation of the hardening/softening parameter H defined by Eq. (3.22). To that end,
we start from the bilinear form of the stress-strain curve plotted in Fig 3.7. In a general
CHAPTER 3. CONSTITUTIVE MODELS 34

Su
h1
0
S u
pl
geq

geq

Figure 3.7: Uniaxial hardening law in shear.

case of three-dimensional state of stress it is acceptable to write an evolution of undrained

pl
shear strength Su in terms of equivalent plastic shear strain γeq (recall Eq. (3.14) to set
T 1
pl pl pl 2
γeq = (2/3(ε ) QPQε )) ) as
pl
Su (κ) = Su (γeq ) = Su0 + h1 γeq
pl
, (3.47)

>
where h1 0 is a hardening/softening parameter, respectively, and Su0 is the initial shear
<
strength. Using Eq. (3.22) together with Eqs. (3.46) and (3.47) readily provides
dF dSu dκ
H=− , (3.48)
dSu dκ dλ
where
dF 1
= − , (3.49)
∂Su cos θ
dSu dSu
= pl
= h1 . (3.50)
dκ dγeq
Adopting strain hardening approach makes possible to write an increment of hardening
parameter dκ in terms of plastic multiplier dλ as
dκ = ηdλ. (3.51)
pl
Next, recall that dκ = dγeq , assume associated plasticity (F = G) and use Eqs. (3.20)
and (3.5) to get
∂F 1
dεpl = dλ = dλ Pσ, (3.52)
∂σ r 2J
pl 2 pl T 1
dκ = dγeq = (ε ) QPQεpl = √ dλ, (3.53)
3 3
dκ 1
= η = √ . (3.54)
dλ 3
CHAPTER 3. CONSTITUTIVE MODELS 35

sy
h2
0 pl
s y
eeq

eeq

Figure 3.8: Hardening modulus depends on normal strength σy .

After substituting for η from Eq. (3.54) and using Eqs. (3.49) and (3.50) we finally arrive
at the desired form of the hardening modulus H written as
· ¸
1 h1
H = −ηh1 − √ = √ . (3.55)
3 3 cos θ
Just for the sake of completeness we now present more common form of the von Mises
yield criterion written in terms of the uniaxial tensile strength as

σy (κ)
F =J− √ . (3.56)
3
where σy is tensile yield strength. In analogy with Eq. (3.47) we assume again a bilinear
form of strain hardening plasticity and write an evolution of the tensile yield strength as
function of a certain hardening parameter κ, see Fig 3.8, as

σy (κ) = σy (εpl 0 pl
eq ) = σu + h2 εeq . (3.57)
q
2
where we set κ = εpl
ep = 3
(εpl )T Qεpl , h2 is hardening parameter, σu0 is the initial normal
strength and the equivalent plastic strain εpl
eq represents a uniaxial strain measure. With
the above definition of yield function and hardening law Eqs. (3.48)-(3.50) and (3.54) now
become
∂F ∂σ y ∂κ
H = − , (3.58)
∂σ y ∂κ ∂λ
∂F 1
= −√ , (3.59)
∂σy 3
∂σy
= h2 , (3.60)
∂κ
∂κ 1
= η = √ . (3.61)
∂λ 3
CHAPTER 3. CONSTITUTIVE MODELS 36

Final substitution of Eqs. (3.59) - (3.61) into Eq. (3.58) gives the constant hardening
modulus H in the form · ¸
1 h2
H = −ηh2 − √ = . (3.62)
3 3

Tangent stiffness matrix

To conclude this section we present a closed form of the tangent stiffness-matrix for the
associated von Mises model given by Eq. (3.56). Assuming isotropic material and using
the relation
∂F 1
= Pσ. (3.63)
∂σ 2J
gives
³ ∂F ´T ³ ∂F ´
el
D = µ, (3.64)
∂σ ∂σ
³ ´³ ∂F ´T T
Del Pσσ PDel
el ∂F el
D D = . (3.65)
∂σ ∂σ 4J 2
where µ in Eq. (3.64) is the shear modulus. Finally, substituting the above relations into
Eq. (3.30) provides
T T
ep Del Pσσ PDel
el el 4µ2 Pσσ
D =D − = D − . (3.66)
4J 2 (µ + h2 /3) J 2 (µ + h2 /3)

3.4.5 Drucker-Prager model

If the results of laboratory tests are plotted in terms effective rather than total stress the
failure criterion becomes dependent on the hydrostatic or mean stress. Such dependence
can be accounted for by using the Drucker-Prager plasticity model.

Yield surface
The Drucker-Prager model can be thought of as an extension of the von Mises model by
including the first invariant of the stress tensor (mean stress) into formulation of the yield
surface. The Drucker-Prager yield surface then plots as a cylindrical cone in the principal
stress space. Corresponding projections into deviatoric and meridian planes appear in
Figs 3.9 and 3.10. Following [19] the Drucker-Prager yield criterion then assumes the
form
F (σ, κ) = J + (σm − c(κ1 ) cot ϕ(κ2 ))M (ϕ(κ2 ))JP = 0, (3.67)
where J and σm are given by Eqs. (3.10) and (3.9), respectively. Recall that vector
T
κ = (κ1 , κ2 ) stores the hardening/softening parameters.
As with the von Mises and Tresca models the radius of the Drucker-Prager circle in
the deviatoric plane can be defined by matching the Drucker-Prager and Mohr-Coulomb
models at a particular value of Lode’s angle θ. This is illustrated in Fig 3.10 where the
irregular hexagon of the Mohr-Coulomb surfaces is compared with the circular shape of the
CHAPTER 3. CONSTITUTIVE MODELS 37

MJP
F 1
MPP
JP
1
G

Jc

−σ m
c cotg ϕ c cotg ϕ MJP
a pp

Figure 3.9: Projections of the yield function and plastic potential functions of Drucker-
Prager model into meridian plane.
.

MC
Modified MC
DP: fit to MC at θ = −30
0
σ1
0
DP: fit to MC at θ = +30
DP: inscribed to MC

0
θ = +30

0
θ = −30

σ3 σ2

Figure 3.10: Drucker-Prager and Mohr Coulomb yield surfaces in the deviatoric plane.
CHAPTER 3. CONSTITUTIVE MODELS 38

Drucker-Prager surface in the deviatoric plane. Three alternative Drucker-Prager circles

are shown. Assuming that both surfaces match at θ = −300 (triaxial compression) we
arrive at the Drucker-Prager circle circumscribed to the Mohr-Coulomb function (green
circle). The corresponding value of MJP reads
√
θ=−30◦ 2 3 sin ϕ
MJP = . (3.68)
3 − sin ϕ
If we desire the Drucker-Prager circle touches the Mohr-Coulomb hexagon at θ = +30 0
(triaxial extension - blue circle) we set the value of MJP to
√
θ=+30◦ 2 3 sin ϕ
MJP = . (3.69)
3 + sin ϕ
Finally, the inscribed circle is found, see [19] for more details, when setting

ins sin ϕ
MJP = sin θ ins
. (3.70)
cos θins + √ sin ϕ
3

with
sin ϕ
θins = arctan √ . (3.71)
3
The model is completed by adopting a plastic potential function of the form
PP
G = J + [σm − app ] MJP = 0, (3.72)

where app follows from Fig. 3.9 when matching F and G for the current value of stress σ.
This gives
c MJP
app = −σm + (σm − c cot ϕ) P P .
MJP
After substituting app into Eq. (3.19) the plastic potential can be written in the form
· ¸
c c MJP PP
G = J + σm − σm + (σm − c cot ϕ) P P MJP = 0, (3.73)
MJP
PP
where MJP is the gradient of the plastic potential function in J − σm space (see Fig 3.9).
PP
If MJP = MJP the yield and plastic potential functions are the same and the model
PP
becomes associated. MJP can be related to the angle of dilation ψ, by substituting ψ
for ϕ in Eqs. (3.68)-(3.70). As with the von Mises plasticity model we now proceed
with the derivation of the hardening/softening modulus H. To that end, consider the
bilinear form of the hardening/softening law for the cohesion c and the angle of internal
friction ϕ plotted in Fig. 3.11. In Fig. 3.11 cin , ϕin and cres , ϕres represent the initial and
residual values of c and ϕ, respectively. It is also evident from Fig. 3.11 that the vector
of hardening/softening parameters κ now becomes

κ(κ1 , κ2 ) = κ(εpl pl
eq , γeq ).
CHAPTER 3. CONSTITUTIVE MODELS 39

c
cres
hc > 0 hardening
1
c in
1
hc < 0 softening
cres

1/2
ε pl
eq
= (2/3εTpl ε pl)

ϕ
ϕres
hϕ > 0 hardening
1
ϕ in
1
hϕ < 0 softening
ϕres

1/2
γ pl
eq
= (2/3eTpl e pl )

deviatoric part of the strain tensor

Figure 3.11: Drucker-Prager and Mohr Coulomb yield surfaces in the deviatoric plane.
CHAPTER 3. CONSTITUTIVE MODELS 40

Referring to Fig. 3.11 and using Eq. (3.22) the hardening/softening modulus H assumes
the form
∂F dc dκ1 ∂F dϕ dκ2
H=− − , (3.74)
∂c dκ1 dλ ∂ϕ dκ2 dλ
where
∂F c dMJP
= 2 MJP + (σm − c cot ϕ) , (3.75)
∂ϕ sin ϕ dϕ
dF
= − cot ϕ MJP , (3.76)
dc
dc dc
= pl = h1 , (3.77)
dκ1 εeq
dϕ dϕ
= pl = h2 . (3.78)
dκ2 eeq
Derivatives of MJP with respect to ϕ for selected values of θ are
ins
√
dMJP 3 3 cos ϕ
= 3 , (3.79)
dϕ (3 + sin2 ϕ) 2
θ=−30◦
√
dMJP 6 3 cos ϕ
= , (3.80)
dϕ (1 − sin ϕ)2
θ=+30◦
√
dMJP 6 3 cos ϕ
= . (3.81)
dϕ (1 + sin ϕ)2
As in the previous section we accept the strain hardening approach to write increments
of the hardening parameters in the form
r r
2 1 2
dκ1 = dεpl eq = η1 dλ = (dεpl )T Qdεpl = + (M P P )2 dλ, (3.82)
3 3 9 JP
r
pl 2 1
dκ2 = deeq = η2 dλ = (dεpl )T QPQdεpl = √ dλ, (3.83)
3 3
where
r
dκ1 1 2
η1 = = + (M P P )2 , (3.84)
dλ 3 9 JP
dκ2 1
η2 = = √ . (3.85)
dλ 3
T T
The terms (dεpl ) QPQdεpl and (dεpl ) Qdεpl are
T 1 T T 1
(dεpl ) QPQdεpl = ( σ P + m MJP PP
) QPQ ( Pσ + MJP PP
m)
2J 2J
1 T 1 T PP
= 2
σ PQPQP σ + σ PQPQMJP m
4J 2J
T PP 1 T PP PP
+ m MJP QPQ + m MJP QPQ MJP m
2J
1 T 2J 2 1
= 2
σ Pσ = 2
= , (3.86)
4J 4J 2
CHAPTER 3. CONSTITUTIVE MODELS 41

T 1 T T 1 1 T
(dεpl ) Qdεpl = ( σ P + m MJP PP
) Q ( Pσ + MJP PP
m) = σ PQP σ
2J 2J 4J 2
1 T PP T PP 1 T PP PP
+ σ PQMJP m + m MJP Q Pσ + m MJP QMJP m
2J 2J
1 1
= + (M P P )2 , (3.87)
2 3 JP
where connections (3.5) were used to arrive at the final results. Finally, substitution
of Eqs. (3.75) - (3.85) back into Eq. (3.74) readily provides the searched form of the
hardening/softening modulus as
· ¸
c dMJP
H = η1 h1 cot ϕMJP − η2 h2 MJP + (σm − c cot ϕ) . (3.88)
sin2 ϕ dϕ

Tangent stiffness matrix

The instantaneous tangent stiffness matrix follows from Eqs. (3.30) or (3.45). The partial
derivatives of the yield function and plastic potential function with respect to stress can
be found using the chain rule
∂F ∂F ∂σm ∂F ∂J
n = = + , (3.89)
∂σ ∂σm ∂σ ∂J ∂σ
∂G ∂G ∂σm ∂G ∂J
ng = = + . (3.90)
∂σ ∂σm ∂σ ∂J ∂σ
Recall that vectors n and ng represent normals to the yield and plastic potential functions
in the stress space, respectively. From Eq. (3.67) and (3.73) it is easy to show that
∂F ∂G PP ∂F ∂G
= MJP , = MJP , = = 1.
∂σm ∂σm ∂J ∂J
The values of ∂σm /∂σ and ∂J/∂σ are model independent and are given by
∂σm
= m, (3.91)
∂σ
T 1
∂J ∂( 21 σ Pσ) 2 1
= = Pσ. (3.92)
∂σ ∂σ 2
With the above expressions we finally get
∂F 1
= Pσ + MJP m, (3.93)
∂σ 2J
∂G 1 PP
= Pσ + MJP m. (3.94)
∂σ 2J
In addition, the second derivative of ng needed to determine the algorithmic tangent
matrix D cons is provided by
µ ¶1/2 T
∂ng 3 σ PσP − Pσσ T P
= . (3.95)
∂σ 2 (σ T Pσ)3/2
CHAPTER 3. CONSTITUTIVE MODELS 42

σ m = 0 MPa φ=5
0
n = 0.175
0
c = 10 Mpa σ1 φ = 20 n = 0.175
0
φ = 45 n = 0.175

σ3 σ2

Figure 3.12: Standard and modified Mohr Coulomb yield surfaces in the deviatoric plane.

3.4.6 Modified Mohr-Coulomb model

As shown in the previous section the Drucker-Prager model plots as a circle in the devia-
toric plane, thus it is invariant with respect to Lode’s angle θ. Although assumption about
smoothed yield surface in the deviatoric plane has been confirmed experimentally, a shape
of the yield surface shows variation with respect to the third invariant. The well known
Mohr-Coulomb model is probably the most widely accepted model that shows variation
with θ in the deviatoric plane. Recall that in the principal stress space the Mohr-Coulomb
model plots as a hexagonal cone. However, as evident from Fig. 3.10, it suffers from the
same drawback as the Tresca model as it experiences corners. To bring the modeling of
plastic behavior of soils closer to conventional soil mechanics and yet avoid corners of the
yield and potential surfaces we introduce a “Modified” Mohr-Coulomb model having a
smoothed surface with a shape somewhere between that of the hexagons and circles.

Yield surface
The yield function is constructed in such a way that it allows the present yield surface
and the one that corresponds to the Mohr-Coulomb law to be matched at all corners of
the yield surface in the deviatoric plane. Several examples are plotted in Fig. 3.12.
CHAPTER 3. CONSTITUTIVE MODELS 43

g( θ )
F 1
MPP
JP
1
G

Jc

−σ m
c cotg ϕ c cotg ϕ g( θ )

a pp

Figure 3.13: Relationship between the yield function and plastic potential in the modified
Mohr-Coulomb model.

The yield function that corresponds to smooth surfaces in Fig. 3.12 is defined as follows
b ϕ(κ2 )) = 0,
F = J + (σm − c(κ1 ) cot ϕ(κ2 ))g(θ, (3.96)
where a shape of the yield surface in the deviatoric plane is described by
g(θ,
◦
b −n ,
b ϕ) = (1 − δ(ϕ))n M θ=−30 [1 + δ(ϕ) sin(3θ)] (3.97)
JP

with ³ A − 1 ´n 3 + sin ϕ
δ= , A= . (3.98)
A+1 3 − sin ϕ
The value of θb is related to Lode’s angle, Eq. (3.11), in the form
" √ #
1 3 3 I 3S
θb = −θ = sin−1 . (3.99)
3 2 J3

Recall that quantities I3s and J in Eq. (3.99) are found from Eqs. (3.12) and (3.10),
respectively. The plastic potential of the modified Mohr-Coulomb model is assumed the
same as for the Drucker-Prager model (see Fig. 3.13).
" #
g( b ϕ)
θ,
c c PP
G = J + σm − σ m + (σm − c cot ϕ) P P MJP = 0, (3.100)
MJP
PP
with MJP being a constant. The searched hardening modulus can be obtained by replac-
ing constant MJP in Eq. (3.88) by function g(θ, b ϕ). This gives
" #
c dg( b ϕ)
θ,
b ϕ) − η2 h2
H = η1 h1 cot ϕ g(θ, b ϕ) + (σm − c cot ϕ)
g(θ, . (3.101)
sin2 ϕ dϕ
CHAPTER 3. CONSTITUTIVE MODELS 44

b
The last term that needs to be evaluated is dg(θ)/dϕ. After some lengthy algebra it can
be expressed as follows
√ £ 1 ¤n
b
dg(θ) 3 × 21+n 3 cos ϕ 1+C
=− h in × (3.102)
dϕ 1 + sin 3θ̄(C−1)
1+C

[[sin ϕ(C − 1)] − 3 [sin ϕ(C + 1)] + sin ϕ [3 − 3C + sin ϕ(C + 1)]]
× h h ii ,
b − 1)
(−3 + sin ϕ)2 (3 + sin ϕ) (C + 1) + sin 3θ(C
q
3+sin ϕ
where C = n
3−sin ϕ
.

Tangent stiffness matrix

The instantaneous tangent stiffness matrix can again be obtained by Eqs. (3.30) or (3.45).
The required normals to the yield and potential surfaces now become
∂F ∂F ∂σm ∂F ∂J ∂F ∂ θb
n = = + + , (3.103)
∂σ ∂σm ∂σ ∂J ∂σ ∂ θb ∂σ
∂G ∂G ∂σm ∂G ∂J
ng = = + . (3.104)
∂σ ∂σm ∂σ ∂J ∂σ
Referring to Section 3.4.5 we immediately see that
∂F ∂σm ∂G ∂σm 1 ∂F ∂J b ∂G ∂J PP
= = Pσ, = g(θ)m, = MJP m.
∂σm ∂σ ∂σm ∂σ 2J ∂J ∂σ ∂J ∂σ
To complete the formulation it remains to determine values of ∂F /∂ θb and ∂ θ/∂σ b in
b
Eq. (3.103). In doing so, we first differentiate the yield Eq. (3.96) with respect to θ to get
∂F h i−(n+1)
= (σm − c cot ϕ) (1 − δ)n MJP θ=−300 b
(−n) 1 + δ sin(3θ) b
(3δ cos(3θ)).
∂θb
b
The term ∂ θ/∂σ is again model independent and it receives the form
∂ θb 1 b
∂ sin(3θ)
= .
∂σ b
3 cos(3θ) ∂σ
Next, using Eq. (3.99) we get after some algebra
√
b
∂ sin(3θ) 3 3 J 3 (∂I3s /∂σ) − (3/2)JI3s Pσ
= .
∂σ 2 J6
Assuming again the plane strain conditions the term ∂I3s /∂σ attains, with the help of
Eq. (3.12), the following form
 1 2 

 3
[σ12 + 2sy sz − sx (sy + sz )]  
∂I3s  1 2 
3
[σ 12 + 2s x s z − s y (s x + s z )]
=bs= .
∂σ 
 −2σ12 sz 

 1 2 
3
[−2σ12 + 2sx sy − sz (sx + sy )]
CHAPTER 3. CONSTITUTIVE MODELS 45

Finally, employing the above expressions allows the two vectors n, ng in Eqs. (3.103) -
(3.104) to be written as
θ=−30 0
1 b (1 − δ)n MJP
n = Pσ + g(θ)m + (σm − c cot ϕ) h i(n+1) ×
2J b
1 + δ sin(3θ)
√ · ¸
3 3 3
×(−n)δ 6 J 3 bs − JI3s Pσ , (3.105)
2J 2
1 PP
ng = Pσ + MJP m. (3.106)
2J

3.5 Interface constitutive model

A proper modeling of soil–structure interaction requires a suitable treatment of relative
movement of the structure with respect to the soil that usually occurs. In the framework of
continuum mechanics, the most appealing way of treating interfaces is the use of interface
elements discussed in Section 2.3.4. The formulation of interface elements presented
therein will be completed in this section by introducing an interface material model that
can be used to simulate contact between two materials, e.g., concrete pile and soil.
In GEOmkp the interface material is based on Mohr-Coulomb failure criterion with
tension cut off, Fig. 3.14. An elastic – rigid plastic response of the interface material in
shear is assumed. Such a behavior is schematically illustrated in Fig. 3.15(a) showing
a variation of the shear stress as a function of the relative tangential displacement. In
tension or compression a purely elastic response of the interface material is considered.
When the tensile stress σ exceeds a certain allowable strength limit Rt , the initial yield
surfaces collapses to a residual surface which corresponds to dry friction, see Figs. 3.14
and 3.15(b).

3.5.1 Yield surface and stress update procedure

The mathematical representation of the initial yield surface displayed in Fig. 3.14 is given
by
F = |τ | + σ tan ϕ − c, (3.107)
where ϕ and c are the angle of internal friction and cohesion of the interface material,
respectively. The direction of the plastic flow depends on the shape of plastic potential
surface. Here, a nonassociated plastic flow rule is assumed with the plastic potential
function written as
G = |τ | + σ tan ψ, (3.108)
where ψ is the angle of dilation. The angle of dilation controls the magnitude of the
irreversible (plastic) volume expansion. As stated in the previous paragraph, the plastic
response is limited to shear only which corresponds to the value of dilation angle ψ equal
to zero (volume preserving return mapping), see Fig. 3.14. Thus, setting ψ = 0 gives the
CHAPTER 3. CONSTITUTIVE MODELS 46

Trial stress (τitr σtri )

ψ=0
Updated stress (τi σi ) volume preserving
return mapping

τ
tanϕ
initial failure surface 1
(τi−1 σi−1 )

c
final failure surface
Rt −σ
c cotg ϕ

Figure 3.14: Failure surface for interface model

τ σ
Ks Kn
1 1
τ max Rt

us un
τ max

Figure 3.15: Constitutive model for interface

CHAPTER 3. CONSTITUTIVE MODELS 47

normals to the yield and potential surfaces, Eq. (3.31), in the form
½ ¾ ½ ¾
∂F τ /|τ | ∂G τ /|τ |
n= = , ng = = , (3.109)
∂σ tan ϕ ∂σ 0

where the stress vector σ follows from Eq. (2.54)

σ = {τ, σ} T .

The normal to the plastic potential function ng then provides the direction of the plastic
flow governing the return mapping algorithm. This algorithm is schematically depicted
in Fig. 3.14. In particular, when solving a plasticity problem the analysis is carried out in
several load increments. To that end, suppose that stresses at state i − 1 are known and
we wish to proceed to a new stress state i by applying a new load increment. This step
results into an increment of the vector of relative displacements ∆ [[u]] = {∆ [[u]] , ∆ [[v]]} T .
The elastic “trial” stresses then follows from

τtri = τ i−1 + Ks ∆ [[u]]i , (3.110)

i
σtr = σ i−1 + Kn ∆ [[v]]i . (3.111)

The particular form of ng just confirms the elastic response in normal direction so that

σ i = σtr
i
,

as evident from Fig. 3.14. The shear stress then follows from the yield condition (3.107).
Note that during plastic flow the stresses must remain on the yield surface. Therefore

F i = |τ i | + σ i tan ϕ − c = 0. (3.112)

Next, multiplying both sides of Eq. (3.112) by the first component of ng , noting that
τ i /|τ i | = τtri /|τtri | and then solving for τ i gives

¡ i ¢ τi
τ i = −σtr tan ϕ + c tri .
|τtr |

Thus, in the absence of pore pressure the stresses at the end of the ith load increment are
given by
½ ¾i ½ ¾ ½ ¾
τ i 0 ¡ i ¢ τtri /|τtri |
= Kn [[v]] + −σtr tan ϕ + c . (3.113)
σ 1 0

3.5.2 Tangent stiffness matrix

Following [24] the algorithmic tangent stiffness matrix Eq. (3.45) can be found from the
expression
∂σ i
Dcons = . (3.114)
∂ [[u]]i
CHAPTER 3. CONSTITUTIVE MODELS 48

Referring to Eq. (3.113) it becomes evident that

µ ½ ¾¶
i 0 T
∂ Kn [[v]] · ¸
1 0 0
½ ¾ = Kn , (3.115)
[[u]]i 0 1
∂
[[v]]i
µ ½ i ¾¶
i τtr /|τtri | T
∂ (−σtr tan ϕ + c) · ¸
0 a11 a12
½ ¾ = , (3.116)
[[u]]i 0 0
∂
[[v]]i

where
µ ¶
i ∂ τtri ∂τtri
a11 = (−σtr tan ϕ + c) i
∂τtr |τtri | ∂ [[u]]i
i
∂(−σtr tan ϕ + c) ∂σ i τtri
a12 = .
∂σ i ∂ [[v]]i |τtri |

After expanding individual derivatives in the above expressions we get

µ i ¶ µ ¶
∂ τtr 1 τtri (∂|τtri |/τtri ) 1 τtri τtri
= − = i 1− i i = 0,
∂τtri |τtri | |τtri | |τtri ||τtri | |τtr | |τtr ||τtr |
i
∂(−σtr tan ϕ + c)
= − tan ϕ,
∂σ i
∂τtri
= Ks ,
∂ [[u]]i
∂σ i
= Kn .
∂ [[v]]i

Finally, introducing the above expressions back into Eq. (3.116) and then adding to
Eq. (3.115) provides the desired tangent stiffness matrix in the form
· ¸
cons 0 −Kn tan ϕ (τtri /|τtri |)
D = . (3.117)
0 Kn

It is interesting to note that the same result will be recovered if starting from Eq. (3.30)
and setting H = 0. In other words, there is no difference between the standard and
algorithmic tangent stiffness matrices for the selected flow rule (ψ = 0).
Chapter 4

Solution strategies

The purpose of this section is to review the basic solution strategies adopted in GMKP
for the solution of the nonlinear systems of equations. Note that both the linear and
nonlinear systems of equations are solved employing standard direct solvers based on the
Cholesky decomposition. Thus, usual skyline and double-skyline storage are used for the
entries of the structural stiffness matrix in both symmetric and non-symmetric problems,
respectively.

k
F, R
k KT1
KT0

k k k
∆ f1 = F − R
∆ f0 = ∆ λ . F
k

k
F

k
R

∆u 1 ∆u 2

Figure 4.1: Incremental procedure: reason for out-of balanced forces

As typical for the analysis of nonlinear problems a standard step by step incremental
procedure (the total load can be split into a set of increments) is adopted in GMKP. Such
approach, consistent with the finite element formulation given in Section 2, results in the

49
CHAPTER 4. SOLUTION STRATEGIES 50

σ1
plastic 


  


B
  



  



elastic
corrector
current yield
        
surface     C  



predictor
      

      


      

       


       
       A A − initial stress state
       B − stress state after elastic predictor
C − stress state at the end of the plastic increment

σ2 σ3

Figure 4.2: Elastic predictor – plastic corrector approach

following set of governing equations

KkT i ∆uki = ∆f ki , (4.1)

where KT i is the instantaneous tangent stiffness matrix, ∆ui is the current increment of
the displacement field and ∆f ki is the current increment of the applied load. The scripts k
and i stand for the current loading increment and the current iteration step, respectively.
Schematic representation of this procedure is evident from Fig. 4.1, in which ∆λ is the
coefficient of proportionality and the term F − R represents a vector of out-of-balanced
forces. It arises as a direct consequence of the nonlinear behavior. Here, the attention is
limited to standard plasticity only. The reason for the presence of out-of-balanced forces is
explained in the next section. The Newton-Raphson and the Arc length methods needed
in the solution of Eq. (4.1) are discussed next.

4.1 Stress return mapping: an algorithm for stan-

dard plasticity
One of the most important steps in the nonlinear analysis is a correct evaluation of
the current stress state. In the framework of incremental solution process the standard
predictor-corrector procedure is usually used. In such a case, the evaluation of the current
stress increment is split into two steps. The first step assumes elastic behavior of the
material. The resulting stress, however, may fall outside of the current yield surface. Such
stress state in unacceptable and must be brought back to the yield surface as illustrated in
Fig. 4.2. This step generates a new increment of the plastic strain and is usually termed
the plastic corrector.
CHAPTER 4. SOLUTION STRATEGIES 51

elastic
σ predictor De step iteration
1 (plastic corrector)

i th step ∆σ i

σtr
i−1 σ i
ε pl i−1
εe

ε i−1 ∆ε i ε
εi

Figure 4.3: Stress return mapping algorithm

4.1.1 Cutting plane algorithm

To determine a new increment of plastic strains and the new current stress state, the return
mapping algorithm must be applied. A popular method of attack is an explicit algorithm
resulting from a Taylor expansion of the yield function around the current values of stress
and hardening parameters. This approach is called the tangent-cutting-plane algorithm
and the interested reader may consult [17, 18] for more details. The essential points of
this algorithm are reviewed below.
To proceed consider the k th load increment and the ith iteration step within this
increment. The goal is to satisfy the yield condition at the end of the i − th iteration.
A one-dimensional illustration of the return mapping algorithm appears Fig. 4.3. After
expanding the yield surface at the end of the ith iteration step via Taylor series expansion
we arrive at the estimate of F (σ ki , κki ) in the form
∂F ¡ ¢
F (σ ki,j+1 , κki,j+1 ) = F (σ ki,j , κki,j ) + |σ k ∆σ ki,j+1 − ∆σ ki,j
∂σ i,j
∂F ¡ ¢
+ |κk ∆κki,j+1 − ∆κki,j = 0. (4.2)
∂κ i,j
To continue, recall Eqs. (3.20), (3.22), (3.24) and (3.31) to get
 

F (σ j , κj ) + nj T Del ∆εj+1 − ∆εj − (∆λj+1 − ∆λj ) ng,j 

| {z }
=0
−Hj (∆λj+1 − ∆λj ) = 0. (4.3)
CHAPTER 4. SOLUTION STRATEGIES 52

Note that scripts k, i were omitted for brevity. Eq. (4.3) now readily provides an estimate
of the current increment of the plastic multiplier ∆λi in the form

F (σ j , κj )
∆λj+1 = ∆λj + . (4.4)
nj Del ng,j +
T Hj

It remains to mention the subscript j that implies an iteration within a single return
mapping step. This is attributed to the fact that the hardening/softening parameters
κ and therefore also the hardening/softening modulus H may in general depend on the
current state of stress which is not known a priory. Note, however, that if the modulus
H remains constant during the return mapping then the algorithm converges in a single
step. The entire algorithm is summarized in Table 4.1.1.

Table 4.1: An algorithm for standard plasticity (k th load increment, ith equilibrium iter-
ation)

1 For each integration point compute:

2 σ trial = σ i−1 + Del ∆εi (trial stress)

3 if F (σ trial , κ0 ) ≥ 0

4 then plastic state: set j = 0, ∆λ0 = 0, σ 0 = σ trial

F (σ j , κj )
do { ∆λj+1 = ∆λj +
nj Del ng,j +
T Hj

σ j+1 = σ trial − ∆λj+1 Del ng,j

κj+1 = κ0 + η∆λj+1

j =j+1

} while F (σ j , κj ) ≥ ε

∆λki = ∆λj , σ ki = σ j

5 else elastic state: σ ki = σ trial

CHAPTER 4. SOLUTION STRATEGIES 53

K T1
F, R K T0

∆f − imbalance
∆λ. F forces = F − R
surcharge

F − load
∆ λ − size of the
load step iteration of R − internal
equilibrium forces

∆u

u0 u1 u2 u − displacement

Figure 4.4: Standard full Newton-Raphson method

4.2 Newton–Raphson method

The Newton-Raphson method is the most frequently used method for the solution of non-
linear finite element equations. The standard full Newton-Raphson method, the modified
Newton-Raphson method and the initial stress method are implemented in GMKP.

4.2.1 Standard Newton–Raphson method

As intimated in the introductory part of this chapter the basic equations to be solved in
nonlinear analysis are
F k − Rk = 0, (4.5)
where F k is the vector of externally applied nodal forces at the end of k − th loading step
and Rk is the vector of nodal forces found from the element stresses such that
Z
k Ne
R = Ae=1 Bσ ke dV. (4.6)
Ve

Fig. 4.1 suggests that Eq. (4.5) might not in general be fulfilled at every step of the
solution process, since the nodal forces Rk depend nonlinearly on the nodal displacements.
Therefore, an iteration is required within a given load increment. The iterative procedure
as shown in Fig. 4.4 arises from the consistent linearization of the nonlinear response of
the finite element equations at iteration i − 1. The consistently linearized tangent stiffness
matrix is therefore formed at the beginning of every iteration step. The resulting iterative
CHAPTER 4. SOLUTION STRATEGIES 54

scheme reads

∆f i−1 = F k − Rki−1 , (4.7)

KkTi−1 ∆ui = ∆f i−1 , (4.8)
uki = uki−1 + ∆ui , (4.9)
uk0 = u k−1
, Rk0 = R k−1
=⇒ ∆f 0 = F − F k k−1
+ ∆f εk−1 . (4.10)

Vector ∆f εk−1 represents the out-of-balance forces found at the end of the previous loading
that are linked to the selected solution accuracy ε.

4.2.2 Modified Newton–Raphson method

Recall that the full Newton-Raphson requires an assembly and factorization of the tan-
gential stiffness matrix KT for each iteration, which might prove to be rather demanding
in terms of computer time. The basic idea of the modified Newton-Raphson method is to
reduce this effort such that a new tangent stiffness matrix is formed at the beginning of
the current load step and then is kept constant for all subsequent iteration steps. Eq. (4.8)
then receives the form
KkT0 ∆ui = ∆f i−1 . (4.11)
The algorithm is summarized in Fig. 4.5. As evident from this figure the pay-off in
reducing the computational cost per iteration, however, may be spoiled by an excessive
increase in the number of iterations.

KT0
F, R
KT0

∆f − imbalance
∆λ. F forces = F − R
surcharge

F − load
∆ λ − size of the
load step iteration of R − internal
equilibrium forces

∆u

u0 u1 u2 u − displacement

Figure 4.5: Modified Newton-Raphson method

CHAPTER 4. SOLUTION STRATEGIES 55

4.2.3 Initial stress method

To further reduce the computational cost one may use the initial elastic stiffness matrix
K el throughout the entire iterative process. Thus it is necessary to assembly and eliminate
the stiffness matrix only one. Eq. (4.8) then becomes

Kel ∆ui = ∆f i−1 . (4.12)

The algorithm is summarized in Fig. 4.6. Note a very slow convergence of the iterative
process compared to the full Newton-Raphson.

F, R K
el

∆f − imbalance
∆λ. F forces= F − R
surcharge

∆ λ − size of the
load step F − load
iteration of
equilibrium
R − internal
∆u forces

u0 u1 u2 u − displacement

Figure 4.6: Initial stress method

4.2.4 Optimal step-length (line search)

The principal idea behind the line search method is to find a scaling factor η of the
current increment of the displacement field ∆uki such that the stationarity of the total
energy functional is preserved at the end of each iteration step. Suppose that a new
displacement uki at the end of the iteration step i is expressed as

uki = uki−1 + η∆ui . (4.13)

Recall Eq. (4.9) in which η is assumed to be equal to one. Next, introducing Eq. (4.13)
into stationary condition of the total potential energy at the end of the i − th step gives

∂Π(uki (η)) ∂uki

δΠ(uki (η)) = δΠ(uki−1 + η∆ui ) = δη = 0. (4.14)
∂uki (η) ∂η
CHAPTER 4. SOLUTION STRATEGIES 56

∆R(ui−1) ∆u i = s(0)
η

η1
∆R(ui−1 + ∆u i ) ∆u i = s(1)

Figure 4.7: Line search method

Since
∂Π(uki (η))
= ∆f i (η), (4.15)
∂uki (η)
∂uki
= ∆ui , (4.16)
∂η
we get
δΠ(uki (η)) = ∆f i (η)∆ui δη = 0. (4.17)
Thus the stationary condition (4.14) corresponds to the condition of zero work of out-
of-balance forces on the displacement increment ∆ui . Eq. (4.17) does not need to be
solved exactly. An estimate of the scaling parameter η can be found from a simple linear
interpolation as displayed in Fig. 4.7. Referring to Fig. 4.7 the first estimate of η is
provided by
−s(0)
η1 = . (4.18)
s(1) − s(0)
A recursive application of Eq. (4.18) then leads to a more accurate value of η, see Fig. 4.7,

−s(0)
ηi+1 = ηi . (4.19)
s(i) − s(0)

Iteration in Eq. (4.19) is usually terminated when the ratio

|si+1 |
< 0.8,
|s0 |

is reached. Clearly, the line search method can either damp (η < 1) or accelerate (η > 1)
the speed of analysis. The latter option, however, is not recommended. Details can be
found in [8].
CHAPTER 4. SOLUTION STRATEGIES 57

4.2.5 Convergence criteria

An additional ingredient of the successful implementation of iterative methods are realistic
convergence criteria to terminate the iterative process. In particular, at the end of each
iteration step, it should be checked whether Eq. (4.5) is satisfied within preset convergence
criteria. The following criteria are implemented in GMKP:
s
∆ui T ∆ui
Pi ≤ εd , (4.20)
T
j=1 ∆uj ∆uj
s
∆f i T ∆f i
≤ εf , (4.21)
∆f 0 T ∆f 0
s
∆ui T ∆f i−1
Pi ≤ εe , (4.22)
| j=1 ∆uj T ∆f 0 |

where εd , εf , εe are preset displacement, out-of-balance forces and energy convergence

tolerance, respectively. The first criterion naturally requires the displacement at the end
of iteration to be found within a certain tolerance, while the second criterion is a measure
of the state of equilibrium at the end of iteration given in terms of out-of-balance forces.
The precision can also be measured by the work of out-of-balance forces on the current
displacement increment as suggested by the third criterion. This is a rather appealing
criterion as it is written in terms of both the displacements and forces. Similar convergence
criteria are proposed in [2].
Note that the selection of values for individual tolerances may be crucial for the success
of computation. Setting these tolerances too large values may lead to inaccurate results,
while selecting rather small values may result in time consuming iterative process in
search for unnecessary accuracy. Also note that there is no guarantee that the process
will converge either due to the excessive number of iterations or due to the divergence.
Therefore, an appropriate divergence check to terminate the iterative process and restart
option should be built into the solver, so that it is possible to repeat a failed increment
with a shorter step. This is also the solution strategy in GMKP.
In particular, the solution is start with some preset increment of the applied load. If
the solution diverges or fails to converge for a given number of step iterations, the load
increment is reduced and the solution is restarted from the last converged step. The
divergence in GMKP is checked against the out-of-balance forces. If the norm out-of-
balance forces Eq. (4.21) increases in two successive iterations, the iterative process is
thought to diverge, the iteration is terminated and the solution is restarted. Similarly, if
the number of iterations needed for the convergence is less then a certain present number,
the load increment can be increased to accelerate the solution process. This option,
however, should be used with caution.
CHAPTER 4. SOLUTION STRATEGIES 58

4.2.6 Additional remarks

Since the Newton-Raphson method and the closely related techniques discussed previously
are widely used, we conclude this section with several useful remarks to shed the light on
the applicability and use of these methods. Further details can be found in [2, 4].

• The advantage of these methods is their applicability in the solution of time-dependent

problems such as consolidation and in the problems where non-proportional loading
is to be applied.

• Figs. 4.4–4.6 illustrate an iterative process for a single degree of freedom that shows
rather good convergence characteristics. For more complicated problems with many
degrees of freedom the convergence of the iterative procedure is not generally as-
sured. The major properties of the full Newton-Raphson method, see e.g. [2] for
more details, suggest that if the current iterative displacement is close to the “ex-
act” one and if the current tangent stiffness matrix consistent with the stress update
procedure is nonsingular and with no abrupt changes, one may expect quadratic con-
vergence. This property, however, is lost with two other techniques. Thus the use
of full Newton-Raphson method is strongly recommended.

• Generally, the choice of the method depends on the degree of nonlinearity. When the
nonlinear response increases the method that allows frequent stiffness update such
as the full Newton-Raphson method should be used. Even if the problem shows
a monotonic convergence the simplified methods, e.g., initial stress method may
fail due to excessive number of iterations needed for convergence. This conclusion,
however, is not universal as there exist problems for which the modified Newton-
Raphson or initial stress method work well. If this is the case then the simplified
methods become particularly attractive as they lead to substantial reduction of
the computational cost (recall fewer or none stiffness updates involved in modified
Newton-Raphson or initial stress method, respectively)

Some of the remarks discussed above will be now illustrated on several examples. As
the first example we consider a simple problem of uniform strip loading applied to the
flat ground. Geometry, loading and boundary conditions are evident from Fig. 4.8(a).
Material parameters of the selected soil are listed in Table 3.44. The Drucker-Prager
plasticity model was selected to represent the soil behavior. The results appear in Figs. 4.9.
Fig. 4.9(a) shows deformation pattern and spread of plastic deformation given in terms
of equivalent plastic strain obtained with the full Newton-Raphson method. The blue
color represents a region with the highest plastic strain, whereas the dark red color marks
the elastic region. The results displayed in Fig. 4.9(b) were found with the initial stress
method. The deformation pattern and the distribution of the equivalent plastic strain
correspond to 87.5% of the total applied load. This is the result of the lack of convergence
due to excessive number of iterations required. The total applied load was not reached
even when substantially reducing the load increment. On the contrary, no difficulties were
encountered when employing the full Newton-Raphson method.
CHAPTER 4. SOLUTION STRATEGIES 59

Table 4.2: Material properties of selected soil

E [MPa] ν c [MPa] ϕ [ 0 ] ψ [0 ] γ [kN/m3 ]

strip loading 25 0.35 10 25 0 20
excavation 25 0.3 10 20 0 20

Table 4.3: Per cent of the total applied load reached

Lack of convergence Newton-Raphson method initial stress method

fine mesh - 3rd level of excavation 87.5% 81.5%

coarse mesh - 4th level of excavation 56.25% 25%

Similar conclusions can be drawn from the second example devoted to the excavation
problem. See Fig. 4.8(b) for the respective geometry and assumed excavation steps.
Figs. 4.10 display the results derived with relatively fine mesh (average size of element
equal app. to 1m was selected). As in the previous example the influence of the number
of stiffness updates on the finite element response was examined. The results provided
by the full Newton-Raphson method appear in Fig. 4.10(a). Fig. 4.10(c) then shows
the respective results found employing the initial stress method. Despite the method
selection, however, it is evident from Figs. 4.10 that the loss of stability occurred in the
3rd stage of excavation. The deformation pattern and the spread of equivalent plastic
strain correspond to the per cent of the total applied load (amount of excavated soil)
reached in the third stage of excavation, see Table 4.3. The resulting slip surface in
Fig. 4.10(a) manifested by the localized plastic deformation is clearly visible. As in the
previous example, the initial stress method was not able to reach the same level of load at
convergence as the full Newton-Raphson method primarily due to the excessive number of
iterations needed. Neither method, however, provides a reliable estimate of the collapse
load. Such a problem can be well treated only with the help of the Arc-length method
discussed in the next section.
The last example is concerned with the effect of the finite element mesh refinement on
the prediction of material response. In doing so the relative element size was increased
up to two meters. The corresponding material response is shown in Figs. 4.11. Note that
larger elements force the plastic region to spread over a larger area of the finite element
mesh. As a consequence the point of instability was moved to the 4th stage of excavation.
Thus, in this particular case, an increase of the element size largely overestimates the
value of the collapse load. Proper selection of the finite element mesh is therefore a very
important task and plays a key role in successful and reliable modeling of the material or
structural response.
CHAPTER 4. SOLUTION STRATEGIES 60

4m

300 kN/m2
15m

γ = 20 kN/m3

30m

(a)

stage of excavation

1 2m
2
3
15m

4
γ = 20 kN/m3

30m

(b)

Figure 4.8: Problem setup: a) uniform strip loading, b) excavation

CHAPTER 4. SOLUTION STRATEGIES 61

(a)

(b)

Figure 4.9: Strip loading: a) full Newton-Raphson method, b) initial stress method
CHAPTER 4. SOLUTION STRATEGIES 62

(a)

(b)

Figure 4.10: Excavation with fine mesh: a) full Newton-Raphson method, b) initial stress
method
CHAPTER 4. SOLUTION STRATEGIES 63

(a)

(b)

Figure 4.11: Excavation with coarse mesh: a) full Newton-Raphson method, b) initial
stress method
CHAPTER 4. SOLUTION STRATEGIES 64

4.3 Arc-length method

In situations where we seek for the unknown collapse load, e.g., stability analysis of earth
slopes, the Newton-Raphson method introduced in the previous sections may experi-
ence rather poor behavior. This can be attributed to the fact that the solutions by the
Newton-Raphson method and closely related techniques are driven by load increments.
The difficulty that arises around the collapse point can be overcome when driving the
solution by displacement increments. This is the essential ingredient of the arc-length
method discussed hereafter. In particular, the method fixes both the loading and dis-
placement at the end of the current load increment by introducing a scalar multiplier
that controls the magnitude of the applied load. The load multiplier now becomes an ad-
ditional unknown and calls for introduction of additional equation for its determination.
There exist several constraint equations in the literature employed for evaluation of λ. In
what follows the constraint equations that arise from so called spherical (Crisfield) and
linearized arc-length methods will be reviewed as they are implemented in GMKP.

4.3.1 Spherical ALM (Crisfield)

∆L − arc length
K T (u start ) (s−1)
F g 0 = ∆f0 = F + λ 0F − R0
1
(s−1)
− δ u 1F g 1 = ∆f1 = F + λ 1F − R1

∆λ 0F ∆λ 1F
∆λ F
∆u 0 δ u1
λ 0F

∆u 1 = ∆u 0 + δ u 1
R1
R(u0 )=R 0
λ start F
∆u
u 0 = ustart + ∆u 0
u 1 = ustart + ∆u



1


u
F+F

(s−1) Rstart
F
~
(s−1)

~
F=

u start u

Figure 4.12: Arc length method with stiffness update after each iteration
CHAPTER 4. SOLUTION STRATEGIES 65

The basic idea behind the arc-length method is best understood from Fig. 4.12. A funda-
mental assumption of the method is that the load vector varies proportionally during the
analysis. To that end, suppose that the solution process is split into several construction
stages and denote the external load applied up to the beginning of current construction
stage s as s−1 F . The total load F = s F = s−1 F + F represent a certain “reference” or
expected load vector at the end of the current load stage s and F is the corresponding
reference load increment applied at the beginning of the s-stage. The goal now becomes
to determine a fraction λ of the applied load F such that the vector of out-of-balanced
forces ∆f = g converges to zero in some appropriate norm measure, Section 4.2.5, at
the end of the current construction stage, Fig. 4.12. Individual vectors and parameters
introduced in Fig. 4.12 are:

• F – total load applied in a given solution stage.

• s F = s−1 F + F – total load expected at the end of a given solution stage.

s−1
• F – load applied up to stage (s − 1).

• λstart – fraction of F at the beginning of a new load increment.

• λki = λkstart + ∆λki – fraction of F at the end of the ith iteration within the k th load
increment.

• Rstart = λkstart F + s−1 F – internal forces at the beginning of a new load increment.

• g i – out-of-balanced forces at the end of the ith iteration.

bi = g i−1 + δλi F – load increment in the ith iteration.

• g

The starting point of the method is an incremental expression of a differential of the

arc-length that provides the additional constraint equation is given by

η 2 ∆ui T ∆ui + β 2 ∆λ2i F T F = ∆L2 , (4.23)

where

• ∆L – represents a radius of a spherical hyper-surface in (u, λ) (standard arc-length

method as introduced by Crisfield in [8]; when setting β = 0 =⇒ ∆L becomes
a radius of a cylinder (Cylindrical ALM)). This parameter is an a priory set step
length and serves to evaluate a corresponding fraction of the current load increment
∆λ. Note that the selection of this parameter is essential for the success of solution.

• β – is a scalar parameter describing the ratio of selected scales for λ and u. The
default setting in GMKP is β = 0. When β - optimize option is chosen then β is
set to the Bergan current stiffness parameter introduced later in this section.

• η – is a scalar parameter which comes from the line search method, recall Sec-
tion 4.2.4.
CHAPTER 4. SOLUTION STRATEGIES 66

With the help of Eq. (4.23) the graphical representation of iteration process displayed
in Fig. 4.12 can be put forward in mathematical terms as follows. Suppose you wish to
determine δui , i = 1, 2, . . . n, in the k th load increment. Referring to Fig. 4.12 we have
¡ ¢
δui = K−1 T (ui−1 ) g i−1 + δλi F , (4.24)
z }| {
s−1
F + λi−1 F mech −Ri−1
| {z }
mechanical part of the total applied load
= K−1 (ui−1 )g +δλi K−1 (u )F , (4.25)
| T {z i−1} | T {zi−1 }
δwi δv i
δui = δwi + δλi δv i =⇒ ∆ui = ∆ui−1 + δui . (4.26)

Next, set
∆λi = ∆λi−1 + δλi , (4.27)
and introduce Eq. (3.27) into Eq. (4.23) to get

L2 = η 2 {∆ui−1 + δw i + δλi δv i } T {∆ui−1 + δwi + δλi δv i }

+ β 2 {∆λi−1 F + δi F } T {∆λi−1 F + δi F } . (4.28)

Rearranging the above equation to collect the terms δλi with the same power gives

a2 δλ2i + a1 δλi + a0 = 0, (4.29)

where

a2 = η 2 δv i T δv i + β 2 F T F (4.30)
¡ ¢
a1 = η 2 {∆ui−1 + δwi } T δv i + δv i T {∆ui−1 + δw i } δλi + 2∆λi−1 β 2 F T F (4.31)
| {z }
T
2δv i {∆ui−1 + δwi }
a0 = η 2 {∆ui−1 + δwi } T {∆ui−1 + δwi } + ∆λ2i−1 β 2 F T F − L2 . (4.32)

Two cases are possible when solving Eq. (4.29) for unknown δλi .

• Both roots δλ1 , δλ2 are real, or

• both roots δλ1 , δλ2 are imaginary, see [4].

When the latter possibility occurs, the computation must be restarted with a shorter step
∆L. As for the real roots it is necessary to choose which one to use. In GMKP the one
which maximizes the angle ϑ found from

∆uT
i−1 ∆ui
cos ϑ = (4.33)
∆L
is used. A complete algorithm requires to determine a starting value of ∆L0 . In GMKP,
the magnitude of ∆L0 can be either prescribed manually, or depending on the prior history
CHAPTER 4. SOLUTION STRATEGIES 67

of iterations the value of ∆L found at the end of stage s − 1 can be used as a initial value
of arc-length step in the s-stage or it can be estimated from a prescribed load increment
∆λ0 . The latter procedure is presented next. To begin set

δλ0 = 0 δw0 = 0,

and write
δu0 = ∆u0 = ∆λ0 KT−1 (ustart )F . (4.34)
From Eq. (4.29) it follows that

a0 = a 1 = 0
a2 = η 2 ∆u0 T ∆u0 + ∆λ20 β 2 F T F − ∆L20 = 0.
z }| {
∆λ δv
z 0 }|1 {
K−1 T (u start )F

Finally, solving for ∆L0 from Eq. (4.29) we get

p
∆L0 = ∆λ0 η 2 δv 1 T δv 1 + ∆β 2 F T F . (4.35)

The subsequent magnitude of ∆L depends on the history of iteration and may either
increased or be reduced if convergence difficulties are encountered. This strategy is also
adopted in GMKP.
Some final remarks should be made regarding the choice of the structural stiffness
matrix. In analogy with the Newton-Raphson method the matrix K can be updated
either at every iteration (full Newton-Raphson method) or at the beginning of a new
load increment (modified Newton-Raphson method) or can be fixed throughout all load
increments in a given stage of construction (initial stress method). Compare Figs. 4.12-
4.14.

4.3.2 Linearized ALM

A simplified version of the arc-length method was proposed by Ramm, [21]. It assumes
that the iterative change of (δui , δλi ) is perpendicular to the secant vector (∆ui−1 , ∆λi−1 ).
The linearized step-length constraint δλ then follows from the formula
T
η∆ui−1 δwi
δλi = − T . (4.36)
η∆ui−1 δv i + ∆λi−1 β 2 F T F

Note that the difficulty with the selection of the root from Eq. (4.29) is avoided. The
method, however, though more simple than the Crisfield method is not as robust as the
full ALM.
CHAPTER 4. SOLUTION STRATEGIES 68

∆L − arc length
F

∆ f 1 = λ 1F − R1

∆λ 0F ∆λ 1F
∆λ F

λ 0F λ 1F =λ start F + ∆λ 1F
∆u 0 δ u1
R1
λ start F ∆u 1 = ∆u 0 + δ u 1

∆u

Figure 4.13: Arc-length method with stiffness update after each load increment

∆L − arc length
F

∆ f 1 = λ 1F − R1
∆λ 0F
∆λ 1F
∆λ F

λ 0F
λ 1F =λ start F + ∆λ 1F
∆u 0 δ u1
R1
λ start F ∆u 1 = ∆u 0 + δ u 1

∆u

Figure 4.14: Arc-length method with no stiffness update

CHAPTER 4. SOLUTION STRATEGIES 69

4.3.3 Automatic step-length control

An automatic procedure that adaptively controls the value of the step-length depending
on the history of iterations can also be very valuable and in favor of the success of compu-
tation. The literature offers several formulae to self-adaptively adjust the value of ∆L for
every new load increment. In GMKP the formula advocated by Ramm is implemented.
It receives the form s
Ibnk
∆Lk = ∆Lk−1 , (4.37)
Ink−1
where Ink−1 is the number of iterations needed for convergence in the k − 1 load step, while
Ibnk represents the number of iterations required in the k th load increment. Although this
formula proved to be quite useful in many problems devoted to the analysis of the collapse
response of structures it is not universal and may fail in some situations. Inexperienced
user should use this option with caution.
Acceleration or damping of the iteration process can be further controlled through the
Bergan current stiffness parameter [3, 4] given by

S k−1
C= , (4.38)
S k−2
where, e.g., S k−1 provides a scalar measure of the structure stiffness at the end of the
k − 1 load step and is provided by

∆λk−1 F T ∆uk−1
S k−1 = . (4.39)
(∆uk−1 )T ∆uk−1

The value of C can be then used to adaptively adjust the parameter β for a new load
increment such that
β k = Cβ k−1 . (4.40)
Recall that parameter β is a scaling factor representing a ration of selected scales for
λ and u. The larger the β the larger the importance of the loading space. If β is
set sufficiently large the solution process is essential driven by the load control. Such
a choice is justifiable in the initial stage of solution. When the degree of nonlinearity
increases or the collapse load is approached, this parameter should be effectively reduced
as suggested by Eq. (4.40). Setting β = 0 (cylindrical ALM) forces the solution to be
driven predominately by the displacement control. For inexperienced user, this option is
recommended.

4.3.4 Additional remarks

As for the Newton-Raphson method there are certain drawbacks as well as advantages
when using the arc-length method. Those pertinent to geotechnical engineering are dis-
cussed next.
CHAPTER 4. SOLUTION STRATEGIES 70

Drawbacks: recall that the application of arc-length method requires the load to change
proportionally. This means that a load increment or a portion of the total applied load
can be written as

∆F = ∆λF
F i = λi F .

Although in static problems such requirement does not usually cause any trouble, it
becomes a severe limitation in the time dependent problems such as consolidation. In
particular, this method becomes inappropriate when assigning different time histories to
various loads within a single construction stage. Such a task can be satisfactorily solved
only with the Newton-Raphson method.

Advantages: the arc-length method is particularly useful in problems that involve the
search for collapse loading such stability problems. Note in stability problems the load
leads to loss of stability is not known a priory. Figs. 4.15 show the application of the arc-
length method to the excavation problem introduced in Section 4.2.6. As in the previous
example, Fig. 4.15(a) displays results found with a relatively fine mesh, while Fig. 4.15(b)
shows the same results obtained with rather coarse mesh. In both examples, the cylindrical
ALM, (β = 0), combine with full NRM, see Fig. 4.12, was used. Comparison with the
full Newton-Raphson method in terms of the per cent of the total applied load reached
at failure appears in Table 4.4. Evidently, both the NRM and the ALM detected the
same collapse load when employing the fine mesh. However, when using the coarse mesh
the NRM over-predicts by large the load at failure. Nevertheless, this result is rather
surprising as one would expect the NRM to fail due to convergence problems before
reaching the collapse load predicted by the ALM. Here we present these results mainly
to highlight problems encountered in the numerical analysis of structures loaded close to
collapse or limit loads.

Table 4.4: Per cent of the total applied load reached

Lack of convergence NRM - FM ALM - FM NRM - CM ALM - CM

3rd level of excavation 87.5% 87.7% 98.1%

4th level of excavation 56.25%

FM - fine mesh
CM - coarse mesh
NRM - full Newton-Raphson method
ALM - cylindrical ALM with full NRM for the stiffness update
CHAPTER 4. SOLUTION STRATEGIES 71

(a)

(b)

Figure 4.15: Excavation with cylindrical ALM combined with full NRM: a) fine mesh, b)
coarse mesh
Chapter 5

Miscellaneous topics

Several miscellaneous topics that may contribute to the clarity of the program performance
are discussed in this section. In particular, we address the problem of stability of earth
slopes, generation of initial geostatic stresses, and implementation of unloading-reloading
procedure. This is an open section and might be further modified.

5.1 Stability analysis of earth slopes

Several examples devoted to the stability analysis of earth slopes are presented here to
promote the applicability of relatively simple plasticity models discussed in Section 3.4.
The stability analysis implemented in GMKP draws on the assumption that the forces
generated by the self weight represent the only source of loading and are applied in a single
increment to an initially stress-free slope. It has been argued in [14] that, in comparison
with limit equilibrium solutions which in general do not account for loading sequence, the
predicted factor of safety derived from simple plasticity models (Drucker-Prager, modified
Mohr-Coulomb) is insensitive to the form of gravity application, thus justifying the use of
gravity turn-on procedure. To arrive at the desired value of the Factor of Safety (F OS) we
follow the method called the shear strength reduction technique. With this procedure the
F OS of a slope is defined as the factor by which the original shear strength parameters c, ϕ
must be reduced to achieve the slope instability. This definition of the F OS is therefore
identical to that of simple limit equilibrium methods in that, that it is defined as the
ratio of restoring and driving moments (see [14] for additional discussion). It follows from
above that the value of F OS is computed, e.g., as

tan ϕreal
F OS = . (5.1)
tan ϕf ailure
To check the influence of individual parameters on the slope stability one may choose to
either one of the two parameters or both parameters. If the latter option is selected then
both c and φ parameters are scaled with the same scaling factor. The analysis proceeds in
such a way that if the solution converges for given magnitudes of soil parameters c, ϕ then

72
CHAPTER 5. MISCELLANEOUS TOPICS 73

20m

15m

10m
G

5m
50m

Figure 5.1: Loading and boundary conditions for stability analysis.

the parameters are reduced and the solution is restarted with new values of c, ϕ given by

ci = scale ci−1 ,
ϕi = scale ϕi−1 , scale < 1 (e.g. 0.9).

This procedure is successively repeated until instability, manifested by no convergence,

occurs.
Several example problems are discussed to illustrate the method performance. Loading
and boundary conditions common to all problems are plotted in Fig. 5.1. In all cases, the
analysis assumes elastic-perfectly plastic behavior of a soil so that no hardening/softening
in involved. The material parameters are listed in Table 5.1. The non-associated Drucker-
Prager model with the parameter MJP evaluated at θ = 300 and dilation angle ψ = 00 is
used to derive the searched F OS.

Table 5.1: Material properties of selected soil

E [MPa] ν c [MPa] ϕ [0 ] ψ [0 ] γ [kN/m3 ]

25 0.3 10 20 0 20

As a first example we assumed a simple slope geometry displayed in Fig. 5.1. In this
particular case, no convergence stopping criterion was activated when the F OS exceeded
the value of 1.44 which agrees well with the result provided by the Sarma method. The
corresponding deformed mesh appears in Fig. 5.2. Such a graphical representation which
allows better understanding of the failure mechanisms is a further advantage of using the
Finite Element Method over limit equilibrium approaches.
The last set of examples is used to examine sensitivity of the element size on the final
response. The following conclusions can be drawn. First, a reliable solution that is very
CHAPTER 5. MISCELLANEOUS TOPICS 74

Figure 5.2: Deformed mesh corresponding to F OS = 1.44

close to the failure can be achieved with conventional finite elements based formulated on
the basis of classical continuum theory. In particular, 6-noded triangular elements with 7
integration points were used in this study. Next, the resulting F OS is virtually insensitive
to the coarseness of the finite element mesh. As evident from Fig 5.3 even with relatively
coarse mesh a satisfactory value of the F OS can be attained. The results also suggest
evolution of localized region even before the peak of stress-strain curve is reached (recall
elastic-perfectly plastic assumption of the soil behavior). This region is, however, highly
dependent on the finite element mesh. Figs. 5.3 - 5.6 show plots of the equivalent plastic
stress εpl
eq . The dark red color corresponds to elastic region while the blue color marks
the region with the highest value of εpleq . The mesh sensitivity to localization of inelastic
deformation is one of the drawbacks of the continuum theory based approaches.

5.2 K0 procedure to generate initial geostatic stress

state
It is often desirable to generate an initial stress state that differs from the one provided
by standard elasticity analysis. It is a well known fact that in the rock analysis the lateral
earth pressure often exceeds the vertical stress as a consequence of various deformation
processes that took place in the past. Such a stress state, however, cannot be attained
for a general class of materials when adopting classical constitutive equations of elasticity.
Recall that in the case of linear elasticity the following relation holds
ν
σx = σy , (5.2)
1−ν
where σx and σy represent the lateral and vertical normal stress, respectively and ν is the
Poisson ratio. Eq. (5.2) can be generalized to get
σx = K 0 σy , (5.3)
where K0 is known as coefficient of lateral earth pressure at rest. Thus setting
ν
K0 = , (5.4)
1−ν
CHAPTER 5. MISCELLANEOUS TOPICS 75

Figure 5.3: Selected mesh and corresponding deformation pattern, F OS = 1.60.

Figure 5.4: Selected mesh and corresponding deformation pattern, F OS = 1.58.

Figure 5.5: Selected mesh and corresponding deformation pattern, F OS = 1.55.

Figure 5.6: Selected mesh and corresponding deformation pattern, F OS = 1.58.

CHAPTER 5. MISCELLANEOUS TOPICS 76

corresponds to standard elasticity approach. The third principal stress follows from the
geometrical symmetry
σz = σ x = K 0 σy . (5.5)
The stress τxy is obviously zero.
Adopting Eq. (5.4) the K0 variable may in general attain values ranging from 0 to 1.
To overcome this limitation the program GMKP offers an option called K0 - procedure that
allows an arbitrary selection of K0 when generating an initial geostatic stress state prior
to any construction stage (stress state that exists in the earth body prior to any mankind
activities). This option, however, should be used with caution. In general, combining
K0 - procedure with externally applied loading, anchors, water and other external effects
in not allowed in GMKP. In such a case one should use the standard approach with the
possibility of including inelastic deformation of soils or rocks.

5.3 Unloading – reloading

A treatment of unloading in geotechnical problems deserves a special attention. Regardless
of the current deformation state the unloading may either follow the path consistent with
the primary elastic loading (type 1 of unloading in Fig. 5.7) or the increase of stiffness
upon unloading is assumed (type 2 of unloading in Fig. 5.7). The latter case suggests an
inelastic response upon unloading. Nevertheless, it is generally accepted that the linear
elastic model with unloading-reloading modulus Eur different from the one used for the
primary elastic loading Eelprimary looading provides reasonable approximation to the true soil
behavior upon unloading. Two specific problems will be now addressed depending on the
current deformation state: unloading from an elastic state (A) or unloading from a plastic
state (B). See Fig. 5.7 for graphical representation.
primary loading
Eel σ1
primary loading
ET
σ 
B
elastic  
stiffness
B 
unloading
from plastic state current yield A


surface

 A
 E ur
type 1 unloading 1
from elastic state
type 2

ε σ2 σ3

Figure 5.7: Unloading problem

CHAPTER 5. MISCELLANEOUS TOPICS 77

Unloading from an elastic state: unloading from the elastic state along the same path
as used for the primary loading (type 1) does not require any special attention. On the
contrary, the need for simulating an increase of stiffness upon unloading (type 2) calls
for a viable criterion to check for unloading. Note that incremental loading procedure
as used in GMKP may prove to be particularly helpful in problems in which the whole
soil volume either compacts or expands at the same time. To reliably determine the
point of unloading in such cases one should apply reasonably small load increments. The
point of unloading at a given integration point might be judged, e.g., by a decrease of
equivalent elastic stress. When reloading takes place the onset for switching from E ur to
Eelprimary loading can be set on by the same criterion as for unloading. Clearly, the present
criterion for reversing from unloading-reloading to primary loading and vice versa needs
to store the maximum equivalent elastic stress ever reached at each integration point.
Also note that changing moduli locally even in elastic regime produces an imbalanced
forces and leads to global iteration of equilibrium. When the initial stress method or the
modified Newton-Raphson method is used the program should be able to recognize the
change in stiffnesses at a respective number of integration points and allow for reformation
of the structural stiffness matrix.

Unloading from a plastic state: unloading from the plastic state appears somewhat simple
in comparison with the above problem. In particular, no need for keeping any information
about the current stress state exists. Clearly, the yield condition naturally serves to
distinguish between loading and unloading. In particular, when unloading from the plastic
state takes place the following conditions must be satisfied

F (σ ki , κki ) < 0,
∂F (σ ki , κki )
∆σ i < 0. (5.6)
∂σ
Eq. (5.6) is nothing else but the unloading condition introduced in Eq. (3.18). The discrete
form of reverse loading (primary loading in plastic state) assumes the form

F (σ ki , κki ) = 0,
∂F (σ ki , κki )
∆σ i > 0. (5.7)
∂σ
Eqs. (5.6) and (5.7) represent natural criteria for replacing the current tangent “modulus”
ET by the modulus for unloading-reloading Eur when unloading occurs and back when
a reloading in to a primary plastic loading is encountered. The exact form of neutral
loading condition as given by Eq. (3.18) cannot be essentially attained in the numerical
analysis due to the roundoff error.
Bibliography

[1] M. Audy, J. Krcek, J. Zeman and M. Sejnoha, ’Constant strain triangular element
with embedded discontinuity based on partition of unity’, submitted for publication
in the Building Research Journal, (2002).

[2] K.J. Bathe, ’Finite Element Procedures’, Prentice hall, (1996).

[3] P.G. Bergan, ’Solution algorithms for non-linear structural problems’, Computers
and Structures, Vol. 12, pp. 497–509, (1987).

[4] Z. Bittnar, J. Šejnoha, ’Numerical methods in structural engineering’, ASCE Press,

(1996).

[5] R.I. Borja, S.R. Lee, R.B. Seed, ’Numerical simulation of excavation in elastoplastic
soils’ International journal for numerical methods in geomechanics, Vol. 13, pp. 231–
249 (1989).

[6] J. Cervenka, ’Discrete crack modeling in concrete structures’, Ph.D. thesis, University
of Colorado, Boulder (1994).

[7] ČSN P ENV 1997-1 (731000), Navrhovánı́ geotechnických konstrukcı́ Část 1: Obecná
paravidla, ČNI, Praha (1996).

[8] M.A. Crisfield, ’Non-linear finite element analysis of solids and structures, Vol. 1,
John Willey & Sons.

[9] J.M. Dluzewski, ’Nonlinear consolidation in finite element modeling’, the IX INter-
national conference on computer methods and advances in geomechanics, Wuhan,
China, pp. 1089–1094 (1997).

[10] day

[11] R.A. Day, D.M. Potts, Zero thickness interface elements - numerical stability and ap-
plications’, International journal numerical and analytical methods in geomechanics,
Vol. 18, pp. 689–708 (1994).

[12] Fine Ltd., User Manual, http://www.fine.cz.

[13] Fine Ltd., GEO4, http://www.fine.cz.

78
BIBLIOGRAPHY 79

[14] D.V.Griffths, ’Stability analysis of highly variable soils by elsato-plastic finite el-
ement’, International Journal for Numerical Methods in Engineering, Vol. 50, pp.
2667–2682 (2001).

[15] J.M. Hohberg, A note on spurious oscillations in FEM joint elements’, Eartq. Engi-
neering Struc. Dynamic., Vol. 19, (1990).

[16] M. Manzari, ’Post localization analysis of soil slopes’, 15th ASCE Engineering Me-
chanics Conference, Columbia University, New York, June 2-5, (2002).

[17] M. Ortiz, J.C. Simo, ’An analysis of a new class of integration algorithms for elasto-
plastic constitutive relations’, Int. J. Num. Meth. Eng., Vol. 23, pp. 353–366, (1986).

[18] J. Pamin, ’Gradient-dependent plasticity in numerical simulation of localizatio phe-

nomena’, Ph.D. Thesis, Technical University Delft, (1994).

[19] D.M. Potts, L. Zdravkovic, ’Finite element analysis in geotechnical engineering, The-
ory’, Thomas Telford Publishing, London, (1999).

[20] D.M. Potts, L. Zdravkovic, ’Finite element analysis in geotechnical engineering, Ap-
plications’, Thomas Telford Publishing, London, (1999).

[21] E. Ramm, ’The Riks/Wemper approach-an extension of the displacement control

method in non-linear analysis’, In: Non-linear computational mechanics (ed. E. Hin-
ton et al.), Pineridge Press, Swansea (1982).

[22] S.K. Sarma, ’Stability analysis of embankments and slopes’, JGED ASCE, 105 GT
12, pp. 1511–1524 (1979).

[23] M. Sejnoha, K. Matous, ’Nonlinear analysis of initially prestressed laminates’, in

CTU Report, Czech Technical University in Prague, Vol. 3(4), pp 55–68 (1999).

[24] J.C. Simo, R.L. Taylor, ’Consistent tangent stiffness operators for rate-indpendent
elastoplasticity’, Comp. Method in Applied Mechanics and Engineering, 48, 101–118
(1985).

[25] I.M. Smith, D.K.H. Ho, ’Influence of construction technique on the performance of
a braced excavation in marine clay’, International journal for numerical methods in
geomechanics, Vol. 16, pp. 845–867 (1992).

[26] Z.Stepanek, ’Zakladani staveb 10’, Faculty of civil enegenering in Prague, 75-79
(1997).