Sie sind auf Seite 1von 29

INTERNATIONAL JOURNAL FOR NUMERICAL AND ANALYTICAL METHODS IN GEOMECHANICS

Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614


Published online 24 December 2009 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/nag.875

Explicit integration of bounding surface model for the analysis of


earthquake soil liquefaction

Konstantinos I. Andrianopoulos1, ∗, †, ‡ , Achilleas G. Papadimitriou2, §


and George D. Bouckovalas3, ¶
1 Geotechnical Department, School of Civil Engineering, National Technical University of Athens, Athens, Greece
2 Department of Civil Engineering, University of Thessaly, Greece
3 Geotechnical Department, School of Civil Engineering, National Technical University of Athens, Athens, Greece

SUMMARY
This paper presents a new plasticity model developed for the simulation of monotonic and cyclic loading
of non-cohesive soils and its implementation to the commercial finite-difference code FLAC, using its
User-Defined-Model (UDM) capability. The new model incorporates the framework of Critical State Soil
Mechanics, while it relies upon bounding surface plasticity with a vanished elastic region to simulate the
non-linear soil response. Stress integration of constitutive relations is performed using a recently proposed
explicit scheme with automatic error control and substepping, which so far has been employed in the
literature only for constitutive models aiming at monotonic loading. The overall accuracy of this scheme is
evaluated at element level by simulating cyclic loading along complex stress paths and by using iso-error
maps for paths involving change of the Lode angle. The performance of the new constitutive model and its
stress integration scheme in complex boundary value problems involving earthquake-induced liquefaction
is evaluated, in terms of accuracy and computational cost, via a number of parametric analyses inspired
by the successful simulation of the VELACS centrifuge Model Test No. 2 studying the lateral spreading
response of a liquefied sand layer. Copyright q 2009 John Wiley & Sons, Ltd.

Received 28 January 2008; Revised 30 July 2009; Accepted 9 November 2009

KEY WORDS: automatic substepping; bounding surface; critical state models; explicit integration; lique-
faction; plasticity

1. INTRODUCTION

The simulation of soil response during strong earthquakes is still a subject of intense research in the
field of Earthquake Geotechnical Engineering. A significant share of this effort is performed via

∗ Correspondence to: Konstantinos I. Andrianopoulos, Mourgkanas 6 str., Maroussi, Athens 15126, Greece.

E-mail: kandrian@tee.gr

Post-Doctoral Researcher.
§ Lecturer.
¶ Professor.

Contract/grant sponsor: General Secretariat for Research and Technology; contract/grant number: EAN-23
Contract/grant sponsor: N.T.U.A.; contract/grant number: 65/1506

Copyright q 2009 John Wiley & Sons, Ltd.


EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1587

numerical analyses, which can handle complex boundary value problems where simple analytical
solutions or empirical methodologies are too crude. For that purpose, the development of numerical
codes that can simulate the coupled interaction of mechanical and fluid response throughout a
severe ground shaking and perform non-linear effective stress analysis of soil behavior is of primary
importance. The corner stone of these codes is the incorporated constitutive model that controls
the qualitative and quantitative accuracy of calculations.
However, the large majority of available constitutive models which are currently implemented
in commercial codes oriented towards Geotechnical Engineering may be competent for static
analyses but cannot capture reliably all aspects of cyclic soil response. The accumulation of perma-
nent deformations and excess pore pressures, the degradation of soil moduli and the concurrent
increase of hysteretic damping as a function of the imposed cyclic shear strains, as well as the
evolution of soil fabric are some of these special aspects which require advanced constitutive
modeling. Hence, most often the solution to this problem is given via specialized codes devel-
oped in various research institutes and universities (inhouse codes), which aim at specific research
needs and consequently may have a rather narrow range of applicability while they are rarely
openly available to the technical community (e.g. GEFDYN [1], DYNAFLOW [2], LIQCA [3],
SUMDES [4], SWANDYNE [5]). Definitely, there are quite a few notable exceptions to this rule,
the number of which is expected to increase with time. For instance, Cyclic1D [6] focuses on
the liquefaction analysis of merely horizontally layered sands using the model of Yang et al. [7].
In addition, the OpenSees platform (http://opensees.berkeley.edu/) provides enormous capabilities
for performing even 3D fully coupled non-linear analyses (e.g. using the model of Manzari and
Dafalias [8]). Nevertheless, this finite-element platform has been used mostly by researchers rather
than practitioners until now, but this is expected to change in the years to come given its simulative
potential.
Kinematic hardening, pressure and density dependence of moduli and third stress invariant
influence of soil response are typical examples of model characteristics which are necessary for
accurate simulation of soil liquefaction, but may complicate considerably the numerical integration
of the constitutive equations [9–11]. For example, stability of the numerical solution may be difficult
to attain in the near zero effective stress regime related to earthquake-induced liquefaction, due to
the pressure dependence of the moduli leading to large induced strains. Implementation of these
complex features in commercial codes becomes even more difficult when two- or multi-surface
plasticity is used [8, 12–14]. The difficulty arises from the fact that, given the main algorithm
of the commercial code, a compatible stress integration scheme must be used, which will ensure
the uniqueness of the numerical solution without resorting to significant simplifications of the
constitutive model. Hence, a balance must be attained between the unavoidable complexity of the
constitutive model, the robustness of the integration scheme and the basic features of the main
algorithm.
In most cases, a robust integration of a constitutive model dictates the use of an implicit algorithm.
For instance, methods such as the closest point projection method [15] have become popular as
they can provide unconditionally stable integration of plasticity models [16–21]. Their popularity
stems from the fact that they can be used irrespective of strain increment size and as such they
may considerably improve the overall efficiency of the code. However, most of these algorithms
are focused on simple constitutive laws, which are mainly oriented towards monotonic loading.
For complex, highly non-linear constitutive laws aiming at cyclic loading the iterative process may
not converge and hence the efficiency of these codes may deteriorate. Furthermore, the consistent
tangent operator which is responsible for the quadratic convergence of Newton iterations, as well

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1588 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

as the second derivatives of the yield and plastic potential surfaces may be very difficult to derive
analytically. Hence, simplifications of the constitutive law may become necessary, at the cost of
deteriorating the predictive capabilities of the model.
Alternatively, semi-explicit and explicit methods can be used. Semi-explicit methods such as the
cutting-plane algorithm [15, 22] simplify the integration process and maintain partly the efficiency
of implicit algorithms. Nevertheless, the consistency condition is not satisfied and drift errors may
accumulate during the integration process, limiting the accuracy and efficiency of these algorithms.
Explicit integration schemes can be applied to complex constitutive models, as they are easy
to implement and do not require simplifications of the constitutive law. However, they are only
conditionally stable, since small strain increments are required for stability. This deficiency is
greatly enhanced if a sub-stepping technique with automatic error control is adopted [23, 24].
This paper deals with a new plasticity model developed for the simulation of monotonic and
cyclic loading of non-cohesive soils, as well as its implementation to the commercial finite-
difference code FLAC [25], using the User-Defined-Model (UDM) capability. In brief, the following
novelties are hereby introduced:
(a) The new plasticity model builds upon earlier work of the authors [26, 27], but introduces
previously unpublished constitutive alterations aiming primarily at facilitating its numerical
implementation, but also for enhancing its predictive capabilities.
(b) Unlike the original model [26, 27], whose accuracy was ascertained only at the element
test level, the accuracy of the new model is explored at the element test level, but also
for a coupled dynamic boundary value problem involving earthquake-induced liquefaction,
which has frequently served as a benchmark problem for validating pertinent constitu-
tive models (e.g. [22]).
(c) Although the explicit stress integration scheme [24] employed herein is not new, its use
with a constitutive model aiming at cyclic loading has never been previously demonstrated,
since similar efforts in the literature focus mostly on monotonic loading (e.g. [28]). More
importantly, this paper explores the efficiency of this integration scheme for extreme cyclic
loading conditions related to earthquake-induced liquefaction.
(d) Despite that a number of previous attempts to implement liquefaction-related constitu-
tive models in FLAC [25] can be found in the literature (e.g. Byrne et al. [29], Wang
et al. [30], Dakoulas and Gazetas [31]) fairly limited details are provided on the employed
implementation procedures. Hence, to fill this gap, this paper provides a detailed account
for the implementation strategy followed herein, which may guide independent attempts to
implement similar constitutive models in explicit numerical codes.
In terms of notation, tensors are written in bold face characters so that they can be easily
distinguished from scalars, while all presented stress quantities are effective.

2. OUTLINE OF CONSTITUTIVE MODEL

2.1. Background
The employed model incorporates the framework of Critical State Soil Mechanics, while it relies
upon bounding surface plasticity with a vanished elastic region (a concept first proposed by
Dafalias [32], Dafalias and Popov [33]) to simulate the non-linear soil response. It is based on

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1589

Figure 1. Definition of the state parameter  (Been and Jefferies [35]) as the ‘distance’ of the current
state ( p, e) from the CSL in the [e-lnp] space.

the original two-surface model of Papadimitriou et al. [26] and Papadimitriou and Bouckovalas
[27], who demonstrated in their papers the capability of the model to simulate the cyclic behavior
of non-cohesive soils (sands and silts), under small, medium and large cyclic strain amplitudes
(see thresholds of cyclic strain amplitudes in Vucetic [34]) and various initial stress and density
conditions, using a single, soil-specific set of model constants, at the element level.
The original model inherited the following constitutive ingredients from Manzari and Dafalias
[8]: (a) the four (open-cone) surface formulation (yield, critical state, bounding, dilatancy) and
the inter-dependency of the last three surfaces on the basis of the state parameter  [35], defined
in Figure 1(b) the kinematic hardening and mapping rules and (c) the concept of allowing
the stress state to move outside the bounding surface in order to simulate strain softening. It
then emphasized on the cyclic response under a wide range of cyclic strain amplitudes, by
introducing two key constitutive elements: (a) a Ramberg-Osgood type non-linear hysteretic
formulation of the ‘elastic’ moduli, which governs shear modulus degradation and hysteretic
damping increase for small-to-medium cyclic shear strains and b) a scalar multiplier of the plastic
modulus, which reflects macroscopically the effect of sand fabric evolution during shearing and
governs the response at medium-to-large cyclic shear strains leading to liquefaction and cyclic
mobility.
Since the publication of the original model, some constitutive models have appeared in the
literature with ingredients that are inspired or directly based on the foregoing two key elements. For
example, Loukidis and Salgado [36] incorporate the Ramberg-Osgood type formulation unaltered
in their two-surface plasticity model, while Lopez-Querol and Blazquez [37] use it in a simplified
manner in their endochronic liquefaction model. Hence, the new model that has been implemented
to FLAC retains the foregoing two key constitutive elements, but also introduces certain alterations
to the whole formulation, the most important being that the conical yield surface of the original
model has been reduced to a point (zero elastic range). On the whole, the incorporated alterations
aim at ameliorating the simulative capabilities of the model, but mainly at facilitating its implemen-
tation to FLAC. The remaining of Section 2 outlines the new model. A more detailed presentation
of the model and its predictive capabilities in relation to various element tests and boundary
value problems involving earthquake-induced liquefaction is presented in Andrianopoulos [38] and
Andrianopoulos et al. [39].

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1590 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

2.2. Model surfaces


In its current form, the constitutive model incorporates three open cone-type surfaces with apex
at the origin of stress space: (i) the critical state surface at which deviatoric deformation develops
for fixed stresses and no volume change (constant void ratio e), (ii) the bounding surface which
locates the peak deviatoric stress ratio states and (iii) the dilatancy surface which dictates the sign
of the plastic volumetric strain increment during loading. The shape of these surfaces in the triaxial
[q, p] space is shown in Figure 2, where q = (a −r ) is the deviatoric and p = (a +2r )/3 is the
mean effective stress, with a and r denoting the axial and the radial effective stresses. It becomes
obvious that these model surfaces for triaxial compression are uniquely defined by deviatoric stress
ratios Mcc , Mcb and Mcd (collectively Mcc,b,d ), which are interrelated via the state parameter  [35],
according to [8]:
Mcb = Mcc +kcb − (1)

Mcd = Mcc +kcd  (2)


 = e −ecs = e −cs + ln( p) (3)
where   are the Macauley brackets yielding x = 0.5(|x|+ x) for any variable x, while kcb ,
kcd ,
Mc , cs and  are model constants, with the last three being related to the constant location of the
c

critical state surface in the [e-p-q] space, and the first two related to the rate of evolution (shrinking
or expanding) of the bounding and the dilatancy surfaces, with respect to the constant critical state
surface as a function of  (note that < 0 leads to dilative response, > 0 to contractive, while
 = 0 is a condition prevailing at the critical state). The multiaxial generalization of these surfaces
is shown graphically in Figure 3, which presents their shape in the -plane (perpendicular to the
hydrostatic p axis) of the deviatoric stress-ratio r space, where r = s/ p, with s = r− pI being the
deviatoric stress tensor (r and I are the effective stress and the identity second-order tensors). As
shown in this figure, the multiaxial generalization is based on a Lode angle -dependency of all
surfaces, introduced analytically by:
Mc,b,d = g(, c)M c,b,d
c (4)

where Mc,b,d defines (collectively) the aperture of the critical-state, bounding and dilatancy surfaces
along radii related to Lode angle  (see Figure 3) on the basis of their respective aperture for
triaxial compression Mcc,b,d (defined in Equations (1) and (2)) and c, which is a model constant
relating the slope of the critical-state surface in triaxial extension (Mec ) to that in compression
Mcc via c = Mec /Mcc . The Lode angle  dependence function g(, c) is given the form proposed by
Papadimitriou and Bouckovalas [27]:
 
2c 1+c 1−c
g(, c) = − + cos(3) (5)
1+c 1−c 2 2
− cos(3)
2 2
where the Lode angle  is defined in terms of the conjugate (image) deviatoric stress ratio tensor
rib (similarly to Li [10]) according to:

3 3 J3
cos(3) = (6)
2 J 3/2
2

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1591

c,b,d
Figure 2. Model surfaces in the triaxial [q − p] space and definition of deviatoric stress ratios Mc,e .

Figure 3. Model surfaces and mapping rule in the -plane of the deviatoric stress ratio space.

with J2 = 1/2 (rib : rib ) and J3 = 1/3 (rib : rib : rib ) being related to the second and third invariants
of rib (see Figure 3), since the symbol ‘:’ denotes the double inner product of two second-order
tensors. The procedure for estimating rib (mapping rule) will be fully explained in Section 2.3
that follows. What is important to note here is that the Lode angle varies from  = 0 for triaxial
compression to  = /3 for triaxial extension, according to Equation (6).

2.3. Elastoplastic formulation


As mentioned above, there is no yield surface in the new model and hence no purely elastic
region. Thus, the response of soil during loading is continuously elastoplastic, i.e. irrecoverable
deformations occur at every incremental step. The choice of a vanished elastic region provides a
smooth transition from non-linear hysteretic behavior at lower strains to fully elastoplastic behavior
at larger strains, and hence improves the numerical robustness and efficiency of the code [40].
In this way, it was possible to satisfactorily overcome a number of stability-related problems that
significantly increase the required computational effort, namely the estimation of a stress path
crossing point of the yield surface, the drift correction resulting from the weak enforcement of the
consistency condition and the subsequent necessary sub-stepping in the integration scheme.
The need for reducing the elastic region to a stress point was further pronounced by the mixed
discretization algorithm [41] adopted in FLAC that appropriately resets the isotropic (effective

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1592 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

stress and strain) components at the end of each time increment in order to solve the problem
of hourglassing (see details in Section 3). What is important to underline here is that resetting
the isotropic (mean) effective stress at the end of each time increment makes the requirement for
enforcing the consistency condition even more pronounced in constitutive models implemented in
FLAC than in any other explicit numerical code that does not perform such an action. In fact, this
is the main reason for adopting a vanished yield surface, the main feature that differentiates the
new model from the original one of Papadimitriou and Bouckovalas [27].
Hence, for any given strain rate ε̇ the effective stress rate ṙ is always computed as:

ṙ = 2G t ė+ K t ε̇ p I−(2G t n+ K t DI) (7)

where ε̇ p and ė are the volumetric and deviatoric strain rates, K t and G t are the bulk and shear
non-linear hysteretic ‘elastic’ moduli (see Section 2.4) and  is the loading index defined as:
L : ṙ 2G t n : ė− V K t ε̇ p
= = (8)
Kp K p +2G t − V K t D
where K p is the plastic modulus. The foregoing equations for ṙ and  are based on the following
generalized definitions of the loading direction L and the plastic strain rate direction R:
n:r V
L = n− I = n− I (9)
3 3
D
R = n+ I (10)
3
where D is the dilatancy function (defined in Section 2.5) and n is the unit deviatoric stress-ratio
tensor (n : n = 1) that is defined on the basis of the direction of the conjugate (image) deviatoric
stress ratio point on the bounding surface rib , namely:

rb
n=  i (11)
rib : rib

Based on the mapping rule shown graphically in Figure 3, the estimation of rib presupposes the
definition of a relocatable projection centre rref defined as the deviatoric stress-ratio state where the
last load reversal took place (see Section 2.6, for triggering of load reversal). Once the reference
state rref has been defined, the conjugate (image) point on the bounding surface rib is found as the
crossing point of the bounding surface and the line originating from r and projecting along the
(r−rref ) direction (as proposed by Wang et al. [42]). Having defined rib enables the depiction of
rid and ric , i.e. the conjugate (image) points on the dilatancy and the critical state surfaces, as the
points on these surfaces that correspond to the same Lode angle  (see Equation (6) and Figure 3).
Analytically, the conjugate (image) points on all three surfaces are given by:

b,c,d 2 b,c,d
ri = M n (12)
3 
As in all bounding surface formulations, a conjugate (image) point is used for the definition of
the distance of the current stress point from the respective surface. In the new model, these scalar

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1593

distances d b,d from the bounding and the dilatancy surfaces are defined in the deviatoric stress-ratio
space and along the direction of n (see Figure 3), as:

db,d = (rib,d −r) : n (13)

Note that distances d b,d may become positive or negative, depending on whether r is within or
b
outside the respective surface. Besides the foregoing distances, of interest are also dmax d
and dmax
b,d
(collectively dmax ), which correspond to the ever-current maximum possible values of d b,d along
the direction of n, given by [8, 43]:

2 b,d
b,d
dmax = (M + Mb,d
+ ) (14)
3 

2.4. Non-linear hysteretic ‘elastic’ moduli


The tangential ‘elastic’ shear modulus G t is assumed to decrease smoothly, from its maximum
value G max (at the last load reversal, when r = rref ) to its ever-current value G t (<G max ) according
to the following generalized Ramberg-Osgood relations (based on [27]):

   
G max Bpa p 1
Gt = = (15)
T 0.3+0.7e2 pa T
⎛ ⎞
 ⎜ ⎟
1 ⎜ 1/2(r−rref ) : (r−rref ) ⎟
T = 1+2 −1 ⎜  ref  ⎟ (16)
a1 ⎝ G max ⎠
2a 1 1
p ref

where B, a1 , 1 are model constants, pa = 98.1 kPa is the atmospheric pressure and T is a scalar
variable which introduces shear modulus degradation as a function of the distance of the current
state (tensor r) from the last load reversal (rref ) on the -plane of the deviatoric stress-ratio space.
Note that p ref and G ref
max correspond to the values of p and G max (using Equation (15)) at the last
load reversal state. Compared to the form of these equations in Papadimitriou and Bouckovalas
[27], Equation (16) has no longer a maximum value for T equal to (2/a1 −1), a fact leading to a
smoother transition in the cyclic response from medium-to-large cyclic shear strains.
Finally, adopting their interrelation from isotropic elasticity, the tangential elastic bulk modulus
K t is estimated from G t via a constant elastic Poisson’s ratio  (a model constant), according to:

2(1+)
Kt = Gt (17)
3(1−2)

The foregoing non-linear hysteretic form of the ‘elastic’ moduli (Equations (15)–(17)) governs
shear modulus degradation and hysteretic damping increase for small-to-medium cyclic shear
strains, and is the first of the two key constitutive ingredients first proposed by [27] providing
enhanced accuracy for the simulation of cyclic response.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1594 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

2.5. Plastic modulus and dilatancy


The basis of the constitutive response for large cyclic strains is the form given to the plastic
modulus K p (see Equations (7) and (8)) and the dilatancy function D (see Equations (7), (8) and
(10)). Specifically, the plastic modulus K p is given as [27]:

K p = ph b h f d b (18)
where h b and h f are dimensionless positive functions to be defined in the sequel, while the sign of
distance d b from the bounding surface (Equation (13)) controls the sign of K p . Specifically, when
d b is negative then K p becomes negative as well, thus simulating post-peak shear strain softening
response of dilative soils, an attribute of the formulation adopted by Manzari and Dafalias [8].
Function h b introduces the highly non-linear interpolation rule of the model according to:

|d b |3
hb = ho (19)
dmax
b −|d b |

with h o being a model constant, and distances d b and dmax


b being estimated by Equations (13) and
(14), respectively.
Function h f is the second key constitutive ingredient providing enhanced accuracy for the
simulation of cyclic response (see Section 2.1) that was first proposed by [27]. Specifically,
this function acts as an empirical macroscopic indicator of the effect of sand fabric evolution
during shearing on the magnitude of plastic strains, since it is a plastic modulus multiplier
(Equation (18)). In more detail, h f is related to an evolving fabric tensor F = f+( f p /3)I, where f
is the deviatoric component of F and f p = trace(F) = F : I, according to [27]:

1+F : I2 1+ f p 2


hf = = (20)
1+F : n 1+f : n
with components f p and f having independent rates of evolution [27]:

f˙p = N ε̇ p
p
(21)
p
ḟ = −N −ε̇ p [F m n+f] (22)
Quantities Fm and N in the above equations are given by:

Fm = 4 max| f p |2 (23)
 
pa
N = No −o  (24)
1o
where No is a model constant, Fm corresponds to the maximum norm allowed for fabric tensor f,
while o and 1o correspond to the values of the state parameter and the major principal stress at
the beginning of cyclic loading (e.g. at the end of consolidation), both quantities being used for
appropriately scaling the rate N of fabric evolution according to the prevailing initial conditions.
As deduced by Equations (20) and (21), the numerator of function h f evolves with the accu-
mulation of plastic volumetric strain ensuring that the history of cyclic shearing leads to stiff-
ening response (i.e. increase of K p ) for small-to-medium cyclic strain amplitudes. In parallel,

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1595

the form of Equation (22) makes the deviatoric component f evolve only upon dilation (due to
the Macauley brackets), thereby simulating conclusions from micromechanical studies depicting
important changes in the sand particle contact normal orientation distribution only upon the dilative
phase of monotonic shearing (e.g. Nemat-Nasser and Tobita [44]). Because of the negative sign
in front of N in Equation (22), the f component evolves in the opposite sense to tensor [Cn+f],
and in this way, the f : n term in the denominator of Equation (20) becomes non-zero, only upon
shear reversal following a dilative shearing phase, and this is due to the change in the direction
of n. In such a case, the response is characterized by larger plastic strains (due to a decrease
of K p ), which under undrained conditions provides the well-known intense contraction following
dilation at small p values, which may eventually lead to liquefaction or cyclic mobility. It should
be noted that, in parallel, the deviatoric fabric evolution f concept evolving as per Equation (22)
has also been proposed by Dafalias and Manzari [45] for the formulation of a scalar multiplier
of the dilatancy function D (see Equation (25)), instead of the plastic modulus K p , for similar
purposes. Finally, note that the form of Equations (20)–(24) in Papadimitriou and Bouckovalas
[27], Equation (23) allows for a 4 times larger maximum norm of f, providing potential for more
intense sand fabric evolution during shearing, a fact leading to more accurate simulation of cyclic
mobility, given the new interpolation rule (Equation (19)).
Adopting a non-associative flow rule, the dilatancy function D is not related to the volumetric
parameter V of the loading direction (see Equations (9) and (10)), but is defined explicitly as [43]:
  
d d 
D = Ao dd 2− d
(25)
dmax

Note that it is the sign of d d that defines the sign of D, thus ensuring negative plastic volumetric
p
strain rates ε̇ p <0 for negative values of d d , i.e. when the stress point is outside the dilatancy
surface.

2.6. Triggering of load reversal


Both the original and the new model require the definition of a relocatable load reversal point
that affects the values of the elastic and plastic strain rates. Specifically, the original model
[26, 27] had different triggering mechanisms of load reversal for the elastic and the plastic strain
rate. Nevertheless, given the vanished elastic region of the new model, it was considered more
appropriate to adopt a unique triggering mechanism. More specifically, in the proposed model, by
definition, load reversal takes place when the scalar loading index  takes a negative value (e.g.
Wang et al. [42] and Pastor et al. [9]). If this occurs, then the load reversal point (reference state) is
automatically relocated, giving updated values to rref , p ref and G ref
max . The update in r
ref affects the
b
location of the conjugate (image) point on the bounding surface ri (Figure 3), which in turn affects
the direction of n (Equation (11)), the cornerstone of the equations estimating the plastic strain
rate. In addition, load reversal affects the elastic strain rate, since this adopts a (Ramberg-Osgood
type) non-linear hysteretic formulation that reduces the values of the elastic moduli G t and K t
as a function of the |r−rref | distance (see Equations (15)–(17) that also employ p ref and G ref max ).
Note that  remains negative only momentarily, since there is no yield surface and the response
at every incremental step is elastoplastic. Thus, immediately after relocating rref , the direction of
the unit vector n changes and the scalar loading index  becomes non-negative again.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1596 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

It is important to note that, upon load reversal, both the elastic and the plastic strain rates
take values much smaller than those in the preceding time step, thus accurately simulating the
stiffer soil response upon load reversal, an important constitutive attribute for accurate simulation
of cyclic loading, especially for small and medium cyclic shear strains. Yet, due to the adopted
interpolation rule (Equation (19)), the plastic strain rate upon load reversal remains non-zero, as
opposed to other discrete memory models that introduce a K p → ∞ condition upon load reversal
(e.g. Wang et al. [42], Li [10], Dafalias and Manzari [45]). It is believed that the selected form of
the interpolation rule reduces the unrealistic (instability) effects of numerical perturbations causing
false triggering of the load reversal criterion in boundary value problems, as compared to discrete
memory models that introduce a K p → ∞ condition upon load reversal in favor of enhanced
accuracy in the simulations of (fully controlled) cyclic loading element tests. On the whole, this
selection is considered a compromise between accuracy and stability for numerical analyses of
seismic ground response. In fact, the mapping and interpolation rules, as well as the related load
reversal criterion which introduces a discontinuously moving rref point, were adopted with seismic
ground response analyses in mind, for which the cyclic loading path of a soil element remains
predominantly proportional (when multi-directional shaking is ignored). For non-proportional paths
(e.g. ratcheting, rotational shear) the constitutive model has not been fully tested, but Wang et al.
[42] showed that the mapping rule based on a discontinuously moving rref point hereby adopted
may provide successful simulations, at least for certain rotational shear paths. Nevertheless, it is
believed that successful simulations of various non-proportional paths may require some additional
constitutive alterations (e.g. the incorporation of inherent anisotropy, as per Dafalias et al. [43],
and/or an additional loading mechanism, as per Li and Dafalias [46]).

3. MODEL IMPLEMENTATION IN FINITE DIFFERENCE CODE

3.1. Overview
The new constitutive model is implemented in the commercially available finite-difference computer
code FLAC [25], using the UDM capability. Thus, the user of FLAC must supply an external
UDM subroutine that will integrate the constitutive equations at each incremental solution step. It
is then the role of this UDM to accurately estimate the effective stress increment ṙ (Equation (7))
and supply an updated set of values for the state variables (r and void ratio e) and the hardening
parameters ( f p and f), given their old set of values and the applied strain increment ε̇.
The equations of motion in FLAC are integrated using the explicit central difference integration
rule. Therefore, small time increments are used to ensure stability. Thus, the use of an implicit
stress integration scheme for the implementation of the highly non-linear new model would prove
unnecessary, while it would require increased computational effort for performing iterative calcula-
tions with complex derivatives of the various constitutive ingredients. No global stiffness matrix is
formulated with this computer code. Darcy’s law is invoked for fluid flow in a porous solid, while
the incremental formulation of coupled deformation-diffusion processes provides the numerical
representations for the linear quasi-static Biot theory.
In the current work, the sub-stepping technique with automatic error control [23, 24] was adopted.
This algorithm belongs to the family of explicit stress integration schemes and divides automatically
the applied strain increment into sub-increments (substeps). An appropriate size for each substep
is found through the use of a modified Euler formula, which is specially constructed to provide

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1597

an estimate of the local error. This scheme is particularly effective in handling multiple hardening
parameters along with complicated hardening laws [28], as required by the new constitutive model.
With the foregoing computer code, the continuum is divided into a finite difference mesh
composed of quadrilateral elements (or ‘zones’ in FLAC terminology). Mixed discretization [41] is
used to solve the problem of hourglassing, which may occur with constant-strain finite difference
quadrilaterals. Namely, each element is automatically subdivided into two overlaid sets of constant-
strain triangular subzones, and stress integration is performed separately for each of the four
subzones of the element. However, the isotropic stress (p) and strain components (ε p ) are taken
to be uniform over the whole quadrilateral element and equal to their average value over the four
triangular sub-zones, while the deviatoric components (s and e) are treated separately for each
triangular sub-zone. This averaging procedure of p and ε p is inherent in FLAC and is performed
after the end of each applied strain increment ε̇. Similarly, the UDM subroutine of the proposed
model in FLAC introduces averaging of the hardening parameters f p and f, to ensure that each
quadrilateral element possesses uniform hardening parameters at the end of each strain increment ε̇.
The foregoing averaging procedure is code-related and will not be elaborated further. What is
not code-related is the explicit stress integration scheme employed for the new constitutive model,
i.e. the second order modified Euler scheme with automatic error control and substepping, which is
presented in Section 3.2. Furthermore, Section 3.3 presents the procedure followed for estimating
the image point on the bounding surface of the new model.

3.2. Stress integration


Integration of the constitutive relations at each zone for a given strain increment ε̇ is performed
via a second order modified forward Euler method. Integration is accomplished in one or more
sub-increments (or substeps), in order to maintain the local truncation error at each step below
a desired tolerance level STOL . The integration scheme is fully presented in Table I, while a
conceptual graphical description of the scheme is given in Figure 4. Specifically, to facilitate the
integration process and especially the error control process that follows, a pseudo-time T is defined,
with 0T 1 for each strain increment ε̇ (with T = 0 and T = 1 marking the beginning and the
end of the integration process for each strain increment). If the error control process requires it,
each strain increment ε̇ is further divided to sub-increments ε̇a , with the aid of a pseudo-time
sub-increment Tn (0<Tn 1) according to:
 
ε̇ap ε̇ p
ε̇ = ė + I = ε̇Tn = ė+ I Tn
a a
(26)
3 3

Thus, stress integration for each strain sub-increment is performed between pseudo-times Tn−1
and Tn = Tn−1 +Tn , with subscripts n −1 and n denoting the start and the end of the sub-increment,
respectively. Note that if the error control process shows sufficient accuracy, then substepping is not
performed, by setting T = 0 and Tn = 1 in the algorithm of Table I. Initiating a sub-increment from
a current stress state rn−1 = rn−1 + pn−1 I, a first approximation of the effective stress increment
ṙ1 is calculated through the first order accurate Euler solution using constitutive relations (1)
through (25), with all state (stress and void ratio) dependent quantities and hardening parameters
estimated at the current state rn−1 . In this manner, ṙ1 , its deviatoric stress ratio component ṙ1
and the respective increments of the hardening parameters ( f˙p,1 and ḟ1 in this model) are given
by setting subscript j = 1 in equations (I), (II), (III) and (IV) in step 3 of Table I.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1598 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

Table I. Section of UDM routine that performs the second order modified forward Euler with automatic
error control and substepping stress integration of the new constitutive model.
Step Description
1 Enter with initial stresses ro , initial hardening parameters f p,o , fo , given total strain increment ε̇ -
Initialize values of T = 0 and Tn = 1
2 If T <1, then continue integration (go to step 3)
3 Compute first order ( j = 1) and second order ( j = 2) estimates of stress and hardening parameter
increments, given the respective image point ri, b of the current r on the bounding surface (see Table II),
j j
according to:
ṙ j = 2G t, j ėa + K t, j (ε̇ap )I−< j >(2G t, j n j + K t, j D j I) (I)
ṙ j = r j −rn−1 (II)
f˙p, j = N (ε̇ p )
p,a
(III)
p,a
ḟ j = −N <− ε̇ p >[F m n j +f j ] (IV)
with
ε̇a = Tn ε̇
b , n ,  , f
and values of G t, j , K t, j , ri, j j j p, j , f j and D j for first order estimate ( j = 1) evaluated at the
stress-state rn−1 , while for second order estimate ( j = 2) evaluated at the temporary updated stress-state
(rn−1 + ṙ1 ).
4 Compute new stresses and hardening parameters and hold them in temporary storage according to:
rn = rn−1 +0.5(ṙ1 + ṙ2 ) (V)
rn = (rn − pn I)/ pn (VI)
f p,n = f p,n−1 +0.5( f˙p,1 + f˙p,2 ) (VII)
fn = fn−1 +0.5(ḟ1 + ḟ2 ) (VIII)
5 Determine the relative error Rn for the current substep from: Rn = max(R , Rr , R f p , R f ) where
   
Rr = 0.5 ṙ2r−ṙ1  = 0.5 (ṙ2 −ṙ1 ):(ṙ2 −ṙ1 )
rn :rn (IX)
n
   
Rr = 0.5 ṙr2 −ṙ1 
 = 0.5 (ṙ2 −ṙ1 ):(ṙ2 −ṙ1 )
rn :ṙn (X)
 n 
( f˙ p,2 − f˙ p,1 )2
Rfp = 0.5 2 (XI)
f p,n
   
Rf = 0.5 ḟf
2 −ḟ1 
= 0.5 (ḟ2 −ḟ1 ):(ḟ2 −ḟ1 )
(XII)
n fn :f˙ n
6 Compare relative error Rn with tolerance level STOL . If Rn >STOL the substep has failed and a smaller
pseudo time step needs to be found by extrapolation. First compute:

= max{0.9 STOL /Rn , 0.1}
and then set:
Tn = max{ Tn , Tmin }, with Tmin = 10−3 (maximum number of substeps 1000) before returning to
step 3
7 The substep is successful, so update the stresses and the hardening parameters with values of step 4
8 Extrapolate to obtain the size of the next substep by computing:

= min{0.9 STOL /Rn , 1.1}
if the previous step failed, limit the step size growth further by enforcing:
= min{ , 1}
Compute new step size and update pseudo-time according to:
Tn = Tn
T = T + Tn

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1599

Table I. Continued.
Step Description
9 Ensure the next step size is not smaller than the minimum step size and check that integration does not
proceed beyond T = 1 by setting:
T = max{T, Tmin }
and then
T = min{T, 1− T }
10 Exit with stresses rn and hardening parameters f p,n and fn at the end of increment with T = 1

Figure 4. Schematic illustration of the two-step modified Euler stress integration scheme
presented in detail in Table I.

Then a temporary update of the stress state (rn−1 + ṙ1 ), the void ratio e and all hardening
parameters enable a second order modified Euler approximation of the stress increment ṙ2 , which
is calculated using constitutive relations (1) through (25), with all state (stress and void ratio)
dependent quantities and hardening parameters estimated at the temporary updated state. In this
manner, ṙ2 , ṙ2 , f˙p,2 and ḟ2 may be estimated by setting subscript j = 2 in equations (I), (II), (III)
and (IV) in step 3 of Table I in order to denote the use of the temporarily updated values of
all constitutive ingredients. Before proceeding, note that the estimation of ṙ1 and ṙ2 requires the
definition of the image point location on the bounding surface, which is estimated according to the
procedure described in Section 3.3, once for estimating ṙ1 and a second time for estimating ṙ2 .
The new stresses (rn and rn ) and hardening parameters ( f p,n and fn ) at the end of the increment
are computed and temporarily stored based on the average estimated increments according to
equations (V) through (VIII) in step 4 of Table I. This two-step averaging procedure employed by
the stress integration scheme is graphically presented in Figure 4.
The reason for highlighting rn , rn , f p,n and fn in Table I is that these are the quantities used for
estimating the local error associated with the stress integration scheme, in an attempt to control
the global integration error. Specifically, for each sub-increment, the local error measure R of each

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1600 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

of these quantities (denoted as R , Rr , R f p and R f , respectively) is found by taking the difference


between their second ( j = 2) and first ( j = 1) approximations, normalized by their respective
temporarily stored values in step 4 of Table I. The local error measures used are estimated according
to equations (IX) through (XII) in step 5 of Table I.
If the maximum of these error estimates Rn = max(R , Rr , Rfp , R f ) for this sub-increment
exceeds the predefined tolerance level STOL , then the sub-increment is rejected and a smaller
Tn is introduced following the procedure outlined in step 6 of Table I. This procedure uses
a reduction factor that takes into account how much was STOL exceeded, as well as a user-
defined maximum number of sub-increments introduced via Tmin (e.g. Tmin = 10−3 implies
that a maximum number of 1000 sub-steps may be used during a load step). Given this smaller
sub-increment, all computations starting from step 3 are repeated until Rn <STOL , the condition
for accepting a sub-increment (see step 7 in Table I) and allowing the final update of the stresses
and the hardening parameters with the values estimated at the last step 4. In the sequel, if a
sub-increment is accepted, then a growth factor is enforced for the next sub-increment, with a
maximum value of 10% increase relative to the previous sub-increment (see step 8 in Table I),
and this in an effort to reduce the total computational cost. The integration process continues until
T = 1, a condition that marks the end of the increment and which allows for exit of the UDM
routine with the final values of the stresses r and all state variables.

3.3. Image point location on the bounding surface


This section of the UDM routine is of key importance, as it facilitates the definition of the mapping
rule and consequently of the Lode angle  (Figure 3), the unit vector n and the distances from
the model surfaces d d , d b , dmax
d , db
max used for the computation of the plastic modulus K p and the
dilatancy function D. For this purpose, the Pegasus procedure of Dowel and Jarratt [47] is adopted,
which was used by Sloan et al. [24] to find the yield surface intersection point. This procedure is
unconditionally convergent in four or five iterations, does not require the use of derivatives and is
detailed in Table II.
Specifically, as shown graphically in Figure 3, the conjugate (image) point on the bounding
surface rib is found from the projection along the (r−rref ) direction, and hence:

rib (
) = rref +
(r−rref ) (27)

where
is a (non-negative) scalar variable, whose appropriate value locates the image point of
r on the bounding surface, rib . Based on constitutive Equations (12) and (13), the image point is
assumed to be on the bounding surface when |F(
)|FTOL with tolerance level set to FTOL = 10−5
and F(
) given by the following expression:
 
2 b
F(
) = M n−ri (
) : n
b
(28)
3 

In order to compute the appropriate value of scalar


, an iterative procedure is initiated by defining
a proper set of values
0 and
1 that locate trial image points inside (i.e. F(
0 )<0) and outside (i.e.
F(
1 )>0) of the bounding surface, and then trying to stringent the range by linear interpolation.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1601

Table II. Section of UDM routine that locates the image point rib of the current stress point r on the
bounding surface. This section is called twice, for j = 1 and j = 2 in step 3 of Table I.
Step Description
1 Set initial values of
0 = 0 and
1 = 1 and perform iterations of steps 2–3 until a proper set of initial
values is defined
2 Calculate:
rib (
0 ) = rref +
0 (r−rref ) (I)
rib (
1 ) =rref +
1 (r−rref )  (II)

F(
0 ) = 2 M b n−rb (
) : n
3  i 0 (III)
 
F(
1 ) = 2 M b n−rb (
) : n
3  i 1 (IV)
with n and Lode angle  estimated using Equations (11) and (6), on the basis of each rib value
3 If F(
0 )F(
1 )<0 and F(
1 )>0 then convergence (go to step 4)
If F(
0 )F(
1 )>0 and F(
1 )<0 then re-set values:
0 =
1 and
1 = 2
1 (go to step 2)
4 Calculate:

1 )(
1 −
0 )

=
1 − F(
F(
)−F(
) (V)
1 0
rib (
) = rref +
(r−rref ) (VI)
 
F(
) = 2 M b n−rb (
) : n
3  i (VII)
5 If |F(
)| FTOL = 10−5 then rib = rib (
) and exit, else continue
6 If F(
)F(
0 )<0 then
1 =
& F(
1 ) = F(
) & go to step 4
F(
)F(
)
If F(
)F(
0 )>0 then F(
1 ) = F(
1 )F(
0) &
0 =
& F(
0 ) = F(
) & go to step 4
1

4. ELEMENT RESPONSE UNDER CYCLIC LOADING AND LIQUEFACTION

In this section, the performance of the proposed integration scheme is evaluated at element level.
More specifically, the common case of a cyclic triaxial test under undrained conditions (i.e. under
constant volume) is simulated in axisymmetric mode, using a single finite-difference zone (element)
having height and radius dimensions equal to 1.0 m. Initial conditions for this element consist of
an isotropic mean effective stress po = 160 kPa (corresponding to equal axial and radial effective
stresses ao = ro = 160 kPa) and an initial void ratio eo = 0.72. The element is subjected to cyclic
axial strain ε
of variable amplitude, while the radial strain at every step εr is controlled so
that the volume of the element is kept constant, i.e. by explicitly applying εr = −0.5ε
. The
boundary conditions and the loading sequence for this test are presented in Figure 5. No flow and
gravitational acceleration are considered.
To evaluate the efficiency of the substepping technique, a series of analyses were performed and
compared. The first consisted of 120×103 increments with axial strain rate ε
= 10−7 and error
tolerance STOL = 10−5 , and is to be used as the reference solution for comparison. Then, a series of
analyses was performed using a three-fold larger axial strain rate ε
= 10−4 and only 120 solution
steps. What differentiated these latter analyses was the STOL value used. Specifically, in these
analyses STOL varied between 103 , 100 , 10−2 and 10−4 , with the first of these values practically
disallowing substepping and the remaining three providing three different levels of substepping
intensity. All the foregoing analyses were performed with the values of model constants that

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1602 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

Figure 5. Boundary conditions and loading sequence for element test of Figure 6.

Table III. Model constants: physical meaning, constitutive equations where they appear
and their values for Nevada sand.
No. Physical meaning Equation Value
Mcc Deviatoric stress ratio at critical state in triaxial compression (TC) (1) 1.25
c Ratio of deviatoric stress ratios at critical state in triaxial extension (TE) over TC (5) 0.72
cs Void ratio at critical state for p = 1 kPa (3) 0.910
 Slope of critical state line in the [e-lnp] space (3) 0.022
B Elastic shear modulus constant (15) 600
 Elastic Poisson’s ratio (17) 0.33
kcb Effect of  on peak deviatoric stress ratio in TC (1) 1.45
kcd Effect of  on dilatancy deviatoric stress ratio in TC (2) 0.30
1 Reference cyclic shear strain for non-linearity of ‘elastic’ shear modulus (16) 0.025%

1 Non-linearity of ‘elastic’ shear modulus (17) 0.6


Ao Dilatancy constant (25) 0.8
h o Plastic modulus constant (19) 15 000
No Fabric evolution constant (24) 40 000

are appropriate for the simulation of Nevada sand and are presented in Table III. Details on the
calibration procedure of these constants are beyond the scope of this paper and can be found in
Andrianopoulos [38] and Andrianopoulos et al. [39].
The results of the reference solution are shown in Figure 6, in terms of the effective stress
path and the corresponding stress–strain relation. Observe that the mean effective stress is quickly
reduced towards zero, while the respective stress–strain loops exhibit a highly non-linear, softening
response, showing an example of the model predictions for cyclic loading paths leading to initial
liquefaction. In the sequel, Figure 7 shows the relative error in computed stresses when a larger
strain increment ε
is used (10−4 as compared to 10−7 ) in connection with substepping with
various error tolerance levels of STOL . Note that for the present application, the relative error
(%) is defined as:

(r−r∗ ) : (r−r∗ )
(%) = ×100 (29)
r∗ : r∗

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1603

Figure 6. Effective stress path and stress–strain relation of an exemplary undrained cyclic loading sequence,
as predicted by the new constitutive model (initial conditions: ao = ro = 160 kPa, eo = 0.72).

where r is the computed effective stress and r∗ is the respective effective stress of the reference
solution. The variation of the mean effective stress p and the deviatoric stress q as a function of
the imposed steps and STOL are presented in Figures 7(a),(b) respectively, while the respective
variation of the relative error and the maximum number of substeps in every incremental step
are presented in Figures 7(c),(d). It becomes evident that, when the tolerance level is tightened
(small STOL values), the relative error decreases and the number of substeps per incremental
step increases. For no-substepping (STOL = 103 ), the relative error is significant, especially when
the mean effective stress tends to low values, approaching liquefaction. On the contrary, when
substepping is activated, the algorithm automatically reduces the relative error proportionally to
the prescribed tolerance level.
The performance of the integration algorithm is additionally evaluated with the aid of iso-error
maps, a procedure that has been employed in the past by a number of authors [16, 17, 48, 49], as a
means to assess the overall accuracy of stress integration algorithms at the element level. According
to this technique, three widely different stress states at yield are initially selected. Then, for each
selected stress state, a sequence of specified strain increments is applied, and the corresponding
stresses are computed by applying the algorithm. The results are reported in terms of the relative
error between the effective stress r computed in a single step and the effective stress r∗ of the
reference solution, according to Equation (29). In general, for moderate strain increments (in the
range of 10−4 to 10−3 ), good accuracy is considered to have been attained, if the relative error
remains below 5%.
In the present application, the three initial stress states chosen referred to yield in triaxial
compression (A), in simple shear (B) and in triaxial extension (C). In all cases, an initial void ratio
of eo = 0.72 with an initial isotropic stress condition of po = ao = ro = 160 kPa was used and the
simulations were performed using a single element mesh with the values of model constants that
are presented in Table III. More specifically, stress state A was achieved by setting y3 = y4 = 0
and x1 = x3 = 0 and by applying y1 = y2 = εa0 , with εa0 = 10−4 , for the single element mesh
of Figure 5. Similarly, stress state C was achieved using the same boundary conditions as above,
with the difference being in the sign of εa0 = −10−4 . Finally, stress state B was achieved by setting
y1 = y2 = y3 = y4 = 0 and x3 = x4 = 0 and by applying x1 = x2 = 0 = 10−4 . In all cases,

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1604 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

Figure 7. Effect of the tolerance level STOL on: (a) the effective stress path; (b) the stress–strain relation;
(c) the relative error (%); and (d) the number of substeps per incremental step, for the exemplary
undrained cyclic loading sequence of Figure 6.

εa0 and 0 were applied in 104 incremental steps. Then, a sequence of specified strain increments
was applied in a single step concurrently in the vertical and horizontal directions of the single
element at yield in order to construct the iso-error maps. In each case, the reference solution was
obtained by applying the same sequence of strain increments in 104 steps, with a tolerance level
of STOL = 10−5 .
For reasons of comparison, iso-error maps were constructed for all three initial stress states (A,
B and C) and for three alternative integration schemes: (a) the well-known one-step forward Euler
scheme (implemented by using only the j = 1 solution of step 3 and by omitting steps 4 through 9
in Table I), (b) the two-step modified Euler scheme without substepping (i.e. the proposed stress
integration scheme of Table I using STOL = 103 ) and (c) the two-step modified Euler scheme with
automatic substepping using an error tolerance STOL = 10−3 .
Figure 8 presents the iso-error maps for initial stress states A, B and C and using the one-step
forward Euler scheme, while Figures 9 and 10 present the pertinent iso-error maps for the two-step
modified Euler scheme without and with automatic substepping, respectively. Observe that for the
one-step forward Euler scheme (Figure 8) and moderate strain increments in the range of 10−4
to 10−3 , the relative error is greater than 5%. Furthermore, note that the accuracy deteriorates
remarkably when the direction of imposed strain increments diverges from the initially applied,
i.e. when the loading induces an abrupt change of Lode angle. The accuracy of the algorithm is
enhanced when the two-step modified Euler scheme is employed (Figure 9), with the relative error

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1605

Figure 8. Iso-error maps for points A, B and C using the one-step forward Euler scheme for the stress
integration of the new constitutive model.

Figure 9. Iso-error maps for points A, B and C using the two-step modified forward Euler scheme without
substepping for the stress integration of the new constitutive model.

Figure 10. Iso-error maps for points A, B and C using the two-step modified forward Euler scheme with
automatic error control and substepping for the stress integration of the new constitutive model.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1606 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

becoming less than 5% for a wider range of strain increments. As expected, further enforcement
of automatic substepping (Figure 10) minimizes the relative error below the prescribed tolerance
levels. It is also noteworthy that, for all initial stress states, the performance of the substepping
technique becomes even better for undrained conditions, i.e. when the volume remains constant
by applying an axial (vertical) strain increment εa equal to −2εr with εr being the radial
(horizontal) strain increment. This occurs because keeping the ratio between radial and axial strain
components constant is a condition that is inherently consistent with the underlying assumption of
the substepping technique for proportional variation of the strains over a solution increment [50].

5. SYSTEM RESPONSE IN BOUNDARY VALUE PROBLEMS

The performance of the proposed numerical algorithm in relation to boundary value problems is
evaluated next through the simulation of a centrifuge experiment. For this purpose, Model test No. 2
of the well-known VELACS project [51] is being used, which reproduces the two-dimensional (2D)
response of a mildly sloping liquefiable soil layer. In this way, the algorithm can be quantitatively
validated against some critical aspects of soil response, such as the generation of excess pore water
pressures towards liquefaction, the shear-induced dilation and softening, the effect of evolving
sand fabric anisotropy and the coupling between the soil skeleton and the pore fluid response.
The test arrangement and the instrumentation for Model test No. 2 is shown in Figure 11. In
prototype scale, the experiment refers to a 10 m deep Nevada Sand layer of Dr = 40% relative
density. The centrifuge model was built inside a laminar box, which was tested at 50g centrifugal
acceleration, and was tilted by 2◦ in order to simulate earthquake-induced lateral spreading condi-
tions. The excitation applied at the base of the laminar box had a peak horizontal acceleration
amplitude of 0.235 g, a strong motion duration of approximately 11 s and a predominant frequency
of 2 Hz. The response of the mildly sloping ground was monitored with six (6) accelerometers,
eight (8) pore pressure transducers and six (6) LVDTs for displacement measurements, as shown
in Figure 11.
The finite-difference grid used for the numerical analysis of the prototype sand layer consisted of
1.0×1.0 m2 square elements. The bottom of this grid was fixed in both directions and the VELACS
horizontal acceleration time-history was applied. The parasitic vertical acceleration during the

Figure 11. Schematic illustration of centrifuge testing configuration of VELACS Model No. 2.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1607

experiments was minor and consequently was not taken into account in the numerical simulation.
The lateral boundaries were tied to one-another in order to enforce the same horizontal and vertical
displacements of the two boundaries, as imposed by the laminar box device in the centrifuge test.
Hence, no stress boundaries were used for the bottom and lateral boundaries. On the contrary,
appropriate normal stress and related pore pressure variation was applied at the top boundary, to
simulate the mildly inclined layer of water (1 m thick, on average) existing on top of the sand layer.
Initially static equilibrium was achieved, in order to create the geostatic stress field, and then the
dynamic motion was applied. Coupled analysis was performed, taking into account the interaction
between the mechanical behavior and the diffusion process. The model permeability coefficient
equal to kmodel = 2.1×10−3 cm/s [51] was scaled to kprototype = 1.05×10−1 cm/s in order to take
account the faster diffusion rate occurring at Ng centrifugal acceleration, since water was used
in this experiment as pore fluid (kprototype = N kmodel ). The values of model constants used for the
simulation were those for Nevada sand, as presented in Table III. The tolerance level of the stress
integration scheme was set to STOL = 10−3 , and the maximum number of sub-steps during a load
step was set to 103 (Tmin = 10−3 ).
Figure 12 compares recorded and predicted time histories of the excess pore pressure ratio ru =
u/vo along the axis of the centrifuge model at various depths (−1.45 m, −2.6 m, −5.0 m, −7.5 m)
within the sand layer (u is the excess pore pressure and 
vo the initial vertical effective stress).
Observe that the recordings as well as the numerical simulations indicate that initial liquefaction
has occurred (ru = 0.9−1.0) in the upper 5 m of sand. Furthermore, very high values of ru develop
from the first seconds of the shaking in the upper 2.5 m, while deeper locations develop similarly
high values of ru at later stages of the shaking. The rates of excess pore pressure buildup and
dissipation are satisfactorily simulated as well, except for the first two loading cycles during which
the model predicts a larger rate of buildup than what is shown by the recordings. It is noteworthy
that the time histories of excess pore pressures show instant drops at the depths where initial
liquefaction has been attained. This trait illustrates enhanced dilative response following a condition
of ru ∼
= 1 and is attributed to the shear stress offset originating from the static equilibrium of a
mildly sloping ground.
Figure 13 compares recorded and predicted time histories of ground acceleration (in percent of g),
at the ground surface (AH3) and at the mid-depth (AH5) of the sand layer, along the axis of the
centrifuge model. It may be observed that both sets of data indicate a liquefaction-induced de-
amplification of the acceleration at the ground surface, with peak values of 0.10−0.12 g (neglecting
very high-frequency spikes), as compared to the 0.235 g applied at the base. They also indicate
that, despite the liquefaction of the upper sand layers, the ground-surface acceleration does not
nullify. This is attributed to the foregoing enhanced dilative response which results in lower excess
pore pressure values. Furthermore, Figure 14 compares the recorded and predicted (relative to the
base) lateral displacement profiles, along the axis of the centrifuge model at four different time
instances: t = 3, 6, 9 and 12 s. Similarly to what was measured in the centrifuge test, the simulation
shows increased lateral displacements in the upper 5.0 m, a fact attributed to the higher excess
pore pressures and the ensuing lower shear strength of that upper layer.
In view of the observed good agreement between data and simulations in Figures 12–14, a
parametric study has been performed in order to evaluate the effect of tolerance level STOL on
the efficiency and accuracy of the algorithm for the same boundary value problem. Hence, three
different analyses were conducted that simulated the response of a loose (Dr = 40%), medium
(Dr = 60%) and dense (Dr = 90%) Nevada sand layer having a thickness of 10 m. The boundary
conditions, initial stress field and applied motion were the same as in the VELACS Model test

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1608 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

Figure 12. Comparison of data to simulations for the time history of the excess pore pressure ratio
ru = u/vo developed at various depths along the axis of VELACS Model No. 2.

Figure 13. Comparison of data to simulations for the time history of horizontal ground acceleration
developed at various depths along the axis of VELACS Model No. 2.

No. 2, with the exception of the permeability coefficient, that was set equal to 2.1×10−3 cm/s, i.e.
the actual value for this sand [51]. Again, the values of model constants used in all these analyses
were those for Nevada sand, as presented in Table III. The analyses for each value of Dr were
performed first without substepping and then with sub-stepping and three different tolerance levels
STOL = 10−2 , 10−3 and 10−4 . The effects of STOL on the accuracy of numerical predictions and
the computational cost are explored in Figures 15 and 16.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1609

Figure 14. Comparison of data to simulations for the (relative to the base) lateral displacement profile of
the soil layer at various times of the shaking for the VELACS Model No. 2.

Specifically, the comparison in Figure 15 is made in terms of computed time histories of


horizontal displacements near surface (−2.5 m). Observe that there is essentially no difference
between the predictions with and without sub-stepping, when in the former the tolerance limit
STOL is set higher or equal to 10−2 . The same applies to the respective computational costs,
which are compared in Figure 16, implying that these values of STOL are higher than the relative
error produced in the great majority of solution steps when the sub-stepping is not activated.
On the contrary, as the tolerance level STOL is reduced below 10−2 , sub-stepping is increasingly
activated leading to different displacement predictions at the cost of larger computational time. The
observed increase in the CPU√time for a tolerance level tightened by a factor of 10 is not significant
(compared to the expected 10 growth of total successive substeps [24]), probably due to the
beneficial interaction between the stress integration process and the explicit time marching scheme
used by the numerical code FLAC [25]. However, taking into account that a 10-fold decrease of
STOL from 10−3 to 10−4 provides only a marginal change in predicted displacements at significant
expense in computational cost, it is concluded that the optimum tolerance level for the specific
analyses is approximately equal to 10−3 .
All foregoing results indicate that the proposed algorithm can simulate crucial aspects of soil
behavior with accuracy and efficiency. It should be underlined that due to the explicit time marching
scheme used in FLAC, small increments are a prerequisite for the stability of computations. Thus,
the majority of solution steps (e.g. for strain increments smaller than 10−4 ) lead to relative errors
that are below the usually adopted tolerance levels and hence no sub-stepping is needed. However,
given that explicit integration schemes are only conditionally stable and that the proposed model is
highly non-linear and strongly path-dependent, a more elaborate technique is needed, which will
ensure the stability of the computations. This paper shows that the explicit modified Euler method
with automatic error control and substepping works well for the constitutive model presented,
without affecting significantly the computational time, given that for the optimum value of tolerance
level (STOL = 10−3 ) substepping is only partially activated.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1610 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

Figure 15. Effect of tolerance level STOL on the predicted time histories of the (relative to the base) lateral
displacements, for the various parametric analyses of a laterally spreading liquefied sand layer.

6. CONCLUSIONS

This paper evaluates the performance of a recently proposed explicit scheme with automatic error
control and substepping for the stress integration of a new bounding surface plasticity model

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1611

Figure 16. Effect of tolerance level STOL of the substepping on the computational cost, as this is measured
in terms of the increase in CPU time due to the substepping, for the various parametric analyses of a
laterally spreading liquefied sand layer.

for the monotonic and cyclic behavior of non-cohesive soils. The efficiency of the algorithm is
evaluated at element and system level, after implementing the foregoing constitutive model in the
commercial numerical code FLAC [25], via its UDM capability. In conclusion, attention is drawn
to the following main findings:
(a) The refined explicit integration scheme with automatic error control and substepping adopted
herein can be used successfully for the efficient stress integration of advanced constitutive
models oriented towards cyclic loading and earthquake-induced liquefaction phenomena.
(b) The use of a two-step modified Euler scheme leads to integration errors lower than those
using the classical form of one-step forward Euler scheme. Enhancing the former scheme
with automatic error control and substepping efficiently minimizes the error below a preset
tolerance level (e.g. well below 5% for strain increments in the order of 10−4 to 10−3 ).
(c) For numerical codes using an explicit time marching scheme, the increase of computa-
tional cost in boundary value √ problems due to the substepping technique is not significant
(compared to the expected 10 growth of total successive substeps [24]), given that such
codes already operate with small timesteps and the substepping technique is only partially
activated. Hence, the substepping technique offers numerical stability to a stress integration
rule that is only conditionally stable, at a small computational cost.
(d) Overall, the presented algorithm (constitutive model and stress integration scheme) can be
used for the accurate simulation of the cyclic response of soils, especially for boundary
value problems involving earthquake-induced liquefaction phenomena. This finding is more
thoroughly presented in Andrianopoulos et al. [39], where different boundary value problems
involving the seismic excitation of Nevada sand are equally well simulated with a single set
of values of the model constants, as calibrated on the basis of element test results.
It is important to note that the conclusions of the paper referring to the efficiency of the stress
integration scheme apply to numerical codes that use an explicit scheme for the time integration
of the equation of motion. In the opposite case, i.e. if the numerical code at hand uses an implicit
time integration scheme, the reader is referred to Manzari and Prachathananukit [52] who compare

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1612 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

the performance of various stress integration techniques for similarly complex plasticity models
aimed at cyclic loading.

ACKNOWLEDGEMENTS
The research presented herein was funded by the General Secretariat for Research and Technology
(..E.E.) of Greece, through research project EAN−23 (‘X-SOILS’) and by the Basic Research
Fund ‘Lefkippos’ of N.T.U.A. (Grant No. 65/1506). These contributions are gratefully acknowledged.
Furthermore, the authors thank the editor and the anonymous reviewers for their constructive comments
that helped in enhancing the quality of the manuscript.

REFERENCES
1. Aubry D, Modaressi A. GEFDYN, Manuel Scientifique. Ecole Centrale Paris: LMSS-Mat, 1996.
2. Prevost JH. DYNAFLOW—a nonlinear transient finite element analysis program. Version 0.2 Technical Report,
Department of Civil and Environmental Engineering, Princeton University, Princeton, NJ, 2002.
3. Oka F, Yashima A, Shibata T, Kato M, Uzuoka R. FEM-FDM coupled liquefaction analysis of a porous soil
using an elasto-plastic model. Applied Scientific Research 1994; 52:209–245.
4. Li XS, Wang ZL, Shen CK, SUMDES: a nonlinear procedure for response analysis of horizontally-layered sites
subjected to multi-directional earthquake loading. User’s Manual, Department of Civil Engineering, UC Davis,
1992.
5. Chan AHC. DIANA SWANDYNE II. User’s Manual, Department of Civil Engineering, Glasgow University,
1989.
6. Yang Z, Lu J, Elgamal A. A web-based platform for computer simulation of seismic ground response. Advances
in Engineering Software 2004; 35(5):249–259.
7. Yang Z, Elgamal A, Parra E. Computational model for cyclic mobility and associated shear deformation. Journal
of Geotechnical and Geoenvironmental Engineering (ASCE) 2003; 129(12):1119–1127.
8. Manzari MT, Dafalias YF. A critical state two-surface plasticity model for sands. Geotechnique 1997; 47(2):
255–272.
9. Pastor M, Zienkiewicz OC, Chan AHC. Generalized plasticity and the modelling of soil behaviour. International
Journal for Numerical and Analytical Methods in Geomechanics 1990; 14:151–190.
10. Li XS. A sand model with state-dependent dilatancy. Geotechnique 2002; 52(3):173–186.
11. Ling H, Liu H. Pressure–level dependency and densification behaviour of sand through generalized plasticity
model. Journal of Engineering Mechanics 2003; 129(8):851–860.
12. Mroz Z, Norris VA, Zienkiewicz OC. An anisotropic critical state model for soils subject to cyclic loading.
Geotechnique 1981; 31(4):451–469.
13. Gajo A, Wood DM. A kinematic hardening constitutive model for sands: the multiaxial formulation. International
Journal for Numerical and Analytical Methods in Geomechanics 1999; 23:925–965.
14. Elgamal A, Yang Z, Parra E. Computational modelling of cyclic mobility and post-liquefaction site response.
Soil Dynamics and Earthquake Engineering 2002; 22(4):259–271.
15. Simo JC, Ortiz M. A unified approach to finite deformation elastoplasticity based on the use of hyperelastic
constitutive equations. Computer Methods in Applied Mechanics and Engineering 1985; 49:221–245.
16. Simo JC, Taylor RL. Return mapping algorithm for plane stress elastoplasticity. International Journal for
Numerical Methods in Engineering 1986; 22:649–670.
17. Ortiz M, Popov EP. Accuracy and stability of integration algorithms for elastoplastic constitutive equations.
International Journal for Numerical Methods in Engineering 1985; 21:1561–1576.
18. Rouainia M, Wood DM. Implicit numerical integration for a kinematic hardening soil plasticity model. International
Journal for Numerical and Analytical Methods in Geomechanics 2001; 25:1305–1325.
19. Tamagnini C, Castellanza R, Nova R. A generalized backward Euler algorithm for the numerical integration
of an isotropic hardening elastoplastic model for mechanical and chemical degradation of bonded geomaterials.
International Journal for Numerical and Analytical Methods in Geomechanics 2002; 26:963–1004.
20. Borja RI, Lin CH, Montans FJ. Cam-clay plasticity part IV: implicit integration of anisotropic bounding surface
model with nonlinear hyperelasticity and ellipsoidal loading function. Computer Methods in Applied Mechanics
and Engineering 2001; 190:3293–3323.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
EXPLICIT INTEGRATION OF BOUNDING SURFACE MODEL 1613

21. Borja RI, Sama KM, Sanz PF. On the numerical integration of three-invariant elastoplastic constitutive models.
Computer Methods in Applied Mechanics and Engineering 2003; 192:1227–1258.
22. Ortiz M, Simo JC. Analysis of a new class of integration algorithms for elastoplastic constitutive relations.
International Journal for Numerical Methods in Engineering 1986; 23:353–366.
23. Sloan SW. Substepping schemes for the numerical integration of elastoplastic stress—strain relations. International
Journal of Numerical Methods on Engineering 1987; 24:893–911.
24. Sloan SW, Abbo AJ, Sheng D. Refined explicit integration of elastoplastic models with automatic error control.
Engineering Computations 2001; 18(1/2):121–154.
25. Itasca. Fast Lagrangian Analysis of Continua. Itasca Consulting Group Inc., Minneapolis, Minnesota, 2005.
26. Papadimitriou AG, Bouckovalas GD, Dafalias YF. Plasticity model for sand under small and large cyclic strains.
Journal of Geotechnical and Geoenvironmental Engineering ASCE 2001; 127(11):973–983.
27. Papadimitriou AG, Bouckovalas GD. Plasticity model for sand under small and large cyclic strains: a multiaxial
formulation. Soil Dynamics and Earthquake Engineering 2002; 22:191–204.
28. Zhao J, Sheng D, Rouainia M, Sloan SW. Explicit stress integration of complex soil models. International
Journal for Numerical and Analytical Methods in Geomechanics 2005; 29:1209–1229.
29. Byrne PM, Park S-S, Beaty M, Sharp M, Gonzalez L, Abdoun T. Numerical modelling of liquefaction and
comparison with centrifuge tests. Canadian Geotechnical Journal 2004; 41(2):193–211.
30. Wang Z-L, Makdisi FI, Egan J. Practical applications of a nonlinear approach to analysis of earthquake-induced
liquefaction and deformation of earth structures. Soil Dynamics and Earthquake Engineering 2006; 26:231–252.
31. Dakoulas P, Gazetas G. Insight into seismic earth and water pressures against caisson quay walls. Geotechnique
2008; 58(2):95–111.
32. Dafalias YF. On cyclic and anisotropic plasticity: (i) a general model including material behavior under stress
reversals, (ii) anisotropic hardening for initially orthotropic materials. Ph.D. Thesis, University of California,
Berkeley, 1975.
33. Dafalias YF, Popov EP. A model of nonlinearly hardening materials for complex loadings. Acta Mechanica 1975;
21(3):173–192.
34. Vucetic M. Cyclic threshold shear strains in soils. Journal of Geotechnical and Geoenvironmental Engineering
(ASCE) 1994; 120(12):2208–2228.
35. Been K, Jefferies MG. A state parameter for sands. Geotechnique 1985; 35(2):99–112.
36. Loukidis D, Salgado R. Modeling sand response using two-surface plasticity. Computers and Geotechnics 2008;
36(1–2):166–186.
37. Lopez-Querol S, Blazquez R, Liquefaction and cyclic mobility model for saturated granular media. International
Journal for Numerical and Analytical Methods in Geomechanics 2006; 30:413–439.
38. Andrianopoulos KI. Numerical modeling of static and dynamic behavior of elastoplastic soils. Doctorate Thesis,
Department of Geotechnical Engineering, School of Civil Engineering, National Technical University of Athens
(in Greek), 2006.
39. Andrianopoulos KI, Papadimitriou AG, Bouckovalas GD. Bounding surface plasticity model for the seismic
liquefaction analysis of geostructures. Soil Dynamics and Earthquake Engineering 2009; under review.
40. Naylor DJ. A continuous plasticity version of the critical state model. International Journal for Numerical
Methods in Engineering 1985; 21:1187–1204.
41. Martin J, Cundall PA. Mixed discretisation procedure for accurate solution of plasticity problems. International
Journal for Numerical and Analytical Methods in Geomechanics 1982; 6:129–139.
42. Wang Z-L, Dafalias YF, Shen CK. Bounding surface hypoplasticity model for sand. Journal of Engineering
Mechanics (ASCE) 1990; 116(5):983–1001.
43. Dafalias YF, Papadimitriou AG, Li XS. Sand plasticity model accounting for inherent fabric anisotropy. Journal
of Engineering Mechanics (ASCE) 2004; 130(11):1319–1333.
44. Nemat-Nasser S, Tobita Y. Influence of fabric on liquefaction and densification potential of cohesionless sand.
Mechanics of Materials 1982; 1:43–62.
45. Dafalias YF, Manzari MT. Simple plasticity sand model accounting for fabric change effects. Journal of
Engineering Mechanics (ASCE) 2004; 130(6):622–634.
46. Li XX, Dafalias YF. A constitutive framework for anisotropic sand including non-proportional loading.
Geotechnique 2004; 54(1):41–55.
47. Dowell M, Jarratt P. The Pegasus method for computing the root of an equation. BIT 1972; 12:503–508.
48. Krieg RD, Krieg DB. Accuracies of numerical solution methods for the elastic-perfectly plastic model. Journal
of Pressure Vessel Technology 1977; 99:510–515.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag
1614 K. I. ANDRIANOPOULOS, A. G. PAPADIMITRIOU AND G. D. BOUCKOVALAS

49. Simo JC, Hughes TJR. Computational Inelasticity. Springer: New York, 2000; 392.
50. Potts DM, Ganedra D. An evaluation of substepping and implicit point algorithms. Computer Methods in Applied
Mechanics and Engineering 1994; 119:341–354.
51. Arulmoli K, Muraleetharan KK, Hossain MM, Fruth LS. VELACS verification of liquefaction analyses by
centrifuge studies, Laboratory Testing Program, Soil Data Report, Research Report, The Earth Technology
Corporation, 1992.
52. Manzari MT, Prachathananukit R. On integration of cyclic soil plasticity model. International Journal for
Numerical and Analytical Methods in Geomechanics 2001; 25(6):525–549.

Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Anal. Meth. Geomech. 2010; 34:1586–1614
DOI: 10.1002/nag

Das könnte Ihnen auch gefallen