Computers and Structures 193 (2017) 155–171

Contents lists available at ScienceDirect

Computers and Structures

journal homepage: www.elsevier.com/locate/compstruc

An enhanced numerical procedure for the shakedown analysis in

multidimensional loading domains
K.V. Spiliopoulos a,⇑, K.D. Panagiotou b,c
a
Institute of Structural Analysis & Antiseismic Research, Department of Civil Engineering, National Technical University of Athens, Zografou Campus, 157-80 Athens, Greece
b
AICES, RWTH Aachen University, Schinkelstraße 2, 52062 Aachen, Germany
c
Institute of Applied Mechanics, RWTH Aachen University, Mies-van-der-Rohe-Straße 1, 52074 Aachen, Germany

a r t i c l e i n f o a b s t r a c t

Article history: The Residual Stress Decomposition Method for Shakedown (RSDM-S) is a new iterative direct method to
Received 6 June 2017 estimate the shakedown load in a 2-dimensional (2D) loading domain. It may be implemented to any
Accepted 15 August 2017 existing finite element code, without the need to use a mathematical programming algorithm. An
improved and enhanced RSDM-S is proposed herein. A new convergence criterion is presented that
makes the procedure almost double as fast. At the same time, the procedure is formulated in a
Keywords: 3-dimensional (3D) polyhedral loading domain, consisting of independently varying mechanical and
Numerical algorithms
thermal loads. Using a cyclic loading program that follows the outline of this domain, it is shown that
Direct methods
Cyclic loading
there is hardly any increase in the computational time when passing from a 2D to a 3D domain.
Plasticity Finally, keeping the efficiency, using an alternative cyclic loading program, an automation of the
Shakedown approach to any n-dimensional loading domain is presented. Examples of application are included.
Residual stresses Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction method [4]. Approaches based also on a series of elastic analyses

produced by modifying, iteratively, the modulus of elasticity, form
A major task in civil and mechanical engineering is the estima- another class of direct methods. Among them one should mention
tion of the load carrying capacity of a structure or a component the Linear Matching Method (LMM) [5,6]. An incremental-iterative
under variable loadings. Structures, like buildings, bridges, procedure, that appears to work well in cases of alternative plastic-
pavements, nuclear reactors, aircraft propulsion engines, etc., dur- ity but not for cases of ratcheting, was proposed in [7] and has been
ing their lifetime, are subjected to loads (live load, heavy traffic, implemented in a commercial code. Very recently, a numerical
seismic action, internal pressure, thermo-mechanical loads, etc.) scheme is presented, based on the conditions of the asymptotic
acting in a varying manner. This type of cyclic mechanical and state linked with a specific trial and projection operation, to esti-
thermal loading leads often these structures beyond the elastic mate the plastic strain increments [8].
limit, resulting to plastic straining. A direct method, which is known as the Residual Stress Decom-
The asymptotic cyclic behavior of an elastic-perfectly plastic position Method (RSDM), was presented in [9,10]. The method can
structure under cyclic loading may be determined by time con- predict the long-term cyclic state, either it is shakedown or reverse
suming incremental time-stepping calculations. Direct methods, plasticity or incremental collapse, of an elastic perfectly-plastic
alternatively, have a big computational advantage as they attempt structure when subjected to a given cyclic loading history. The
to find directly this cyclic asymptotic state. Such states are guaran- approach is based on physical arguments that have to do with
teed for structures made of stable material [1]. the expected cyclic nature of the residual stresses. The residual
There are a few direct methods, proposed in the literature, stresses are decomposed into Fourier series with respect to time
among which one may mention the work presented in [2,3] which and the coefficients of these series are calculated iteratively by sat-
forms a sequence of elastic solutions using as a modified loading an isfying equilibrium and compatibility at time points inside the
update of initial strains computed through an update of internal cycle.
variables. This method is the basis of a recently presented direct When, on the other hand, the loading history is unknown, for a
structure to be safe and serviceable, safety margins, e.g.
shakedown limits, have to be estimated so that it fails neither
⇑ Corresponding author. due to incremental collapse (often referred to as ratcheting) nor
E-mail address: [email protected] (K.V. Spiliopoulos). due to reverse plasticity that leads to low cycle fatigue. A direct

http://dx.doi.org/10.1016/j.compstruc.2017.08.008
0045-7949/Ó 2017 Elsevier Ltd. All rights reserved.
156 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171

shakedown analysis is the only way to provide this information. converges to a solution where the constant term is the only non-
For small displacements and elastic-perfectly plastic solids the zero term of the Fourier series.
shakedown analysis is based on two different approaches, the The efficiency of the RSDM-S and RSDM to provide shakedown
lower bound [11] or the upper bound [12] shakedown theorems. boundaries as well as to unveil unsafe conditions in 2-dimensional
The extensions of these two theorems to cover thermal loadings loading domains was recently demonstrated in [54].
were given in [13,14], respectively. In the present work, the RSDM-S method is enhanced by a dif-
Attempts to consider geometric nonlinearities appeared in the ferent convergence criterion, inside the inner loop, that makes the
literature (e.g. [15,16]). Conditions to extend the static theorem method run faster, even more than 40%. Moreover, the method is
to elastic-perfectly plastic cracked bodies have been presented in formulated for a 3-dimensional loading domain consisting of two
[17]. Limited kinematic (e.g. [18]) and nonlinear kinematic harden- mechanical and a thermal load. It is shown that the extension from
ing has been also addressed (e.g. [19]). Recent developments on the a 2-D to a 3-D loading domain hardly influences the amount of
subject have appeared in [20,21] in the framework of the bipoten- computational time to estimate the shakedown load factor as
tial theory. Non-associated plasticity has also been discussed (e.g. opposed to the IPM algorithms where the time is shown to double
[22,23]). Polizzotto, has discussed the shakedown theorems in [55]. Finally, it is shown how the method may be automated to
the context of gradient plasticity theory [24,25]. cater for any n-dimensional loading domain.
The two shakedown theorems form the basis of the big majority The paper is organized in the following way: In Section 2 a proof
of the existing numerical procedures to estimate the shakedown of an existing theorem makes possible to realize the arbitrariness
load. They are formulated as mathematical programming (MP) of the cyclic loading program that passes through the vertices of
problems whose scope is to find the minimum or maximum value the convex loading domain; in Section 3 the enhanced RSDM-S
of an objective function (normally the loading factor) which is sub- procedure, in the form of a flow chart, with the new convergence
jected to various static or kinematic constraints. Linearization, criterion, formulated in a 3-D thermomechanical loading domain
mainly of the yield surface, has led to some early solutions using and assuming a von Mises yield criterion, is presented. The signif-
linear programming algorithms (e.g. [26,27]). More recent contri- icant faster convergence of the enhanced approach is demon-
butions have appeared along the same line (e.g. [28–30]). If the strated through examples of 2-D loading domains in Section 4. In
constraints are not linearized and are kept in their original form Section 5 the method is applied to a 3-D polyhedral loading
(nonlinear), the problem can be formulated as a nonlinear (NLP) domain using a cyclic loading program that passes consecutively
programming problem. The discretization of the continuum by a from all its vertices. Finally, in Section 6 an alternative cyclic load-
large number of finite elements and the big number of constraints ing path combined with a combinatorial algorithm shows how the
often lead to the solution of large size optimization problems. Var- whole procedure may be automated for an n-dimensional domain.
ious numerical techniques have been developed to solve these
problems. Among these one could mention the reduced basis tech- 2. Theoretical considerations
nique [19,31] or algorithms based on Newton iterations [32]. The
evolution of the interior point algorithms (IPM) to solve large scale Let us suppose a structure is subjected to independently varying
optimization problems led to the extensive formulation and solu- cyclic loads that have the same period T. Although the theory may
tion of limit and shakedown analysis problems using these algo- be applied to any number of loads, for reasons of visualization, a
rithms or related techniques (e.g. [33–43]). maximum of three loading (3-D) domain that consists of two
One may also find some alternative approaches in the litera- mechanical and a thermal load will be demonstrated (Fig. 1(a)).
ture for the evaluation of the shakedown load. Such an approach Such a cyclic loading may be represented in the loading space
is based on the work presented in [2], whose application using as a closed loop (Fig. 1(b)). Let us further suppose that each load
the finite element method (FEM) may be found in subsequent has a minimum and a maximum value of variation. Without any
publications (e.g. [44]). Reverse plasticity and collapse load solu- loss of generality, the minimum of all the loads will be considered
tions have been shown to provide upper bounds to the shake- zero. The maximum of each of the loads, denoted by starred quan-
down load [45]. The LMM has also been used to estimate the tities, together with the origin may define a convex (hyper-) cuboid
shakedown load of a structure (e.g. [5,46]). In [47] a solution (Fig. 1(b)). Thus, the cyclic loading will be contained inside this
is proposed, based on the LMM, to estimate a possible shake- cuboid.
down load when friction slip occurs between a rigid surface in In response to this loading the structure that consists of an
contact with an elastic body, subjected to cyclic loading. A quite elastic-perfectly plastic material will develop a stress that may
involved strategy, equivalent to a fictitious incremental strain be decomposed into two parts; an elastic part assuming purely
driven elastoplastic problem, and applied for a von Mises type elastic material behavior and a residual stress part to account for
of material, has been presented in [48]. The numerical perfor- plasticity:
mance of this approach is compared against the IPMs in [49].
An analogous methodology, involving more general yield criteria, rðsÞ ¼ rel ðsÞ þ qðsÞ ð1Þ
was proposed in [50].
where s ¼ t=T denotes a time point inside the cycle.
A numerical approach, which was called RSDM-S has appeared
The structure is discretized, following a standard procedure,
recently [51–53]. It may be used for the evaluation of the shake-
into a finite number of elements that are interconnected at a dis-
down load of elastic-perfectly plastic structural elements under
crete number of nodal points situated on their boundaries. Bold
cyclic thermo-mechanical loading. The basis of the method, both
letters are herein used for vectors and matrices. The stress and
from the conceptual as well as the implementation points of view,
strain vectors are evaluated at the Gauss points (GPs) of the finite
is the RSDM. Since, now, only the variation intervals of the loads
elements (FE).
are known, the problem is converted to an equivalent prescribed
The strain rates, on the other hand, may be decomposed into the
loading problem, drawing any time curve crossing these intervals.
following parts:
The RSDM-S consists of two different iteration loops, one inside the
other and has been formulated for two loads that may vary either e_ ðsÞ ¼ e_ el ðsÞ þ e_ h ðsÞ þ e_ elr ðsÞ þ e_ pl ðsÞ ð2Þ
proportionally or independently. Starting from a high load factor, a
descending sequence of loading factors is established and the _ el
where e ðsÞ is the elastic straining due to both the mechanical and
shakedown load factor is calculated when the iterative procedure the thermal loading [52]. e_ h ðsÞ denotes thermal strain rates that
 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171 157

P2
P2
P2*
P1

(P1* , P2* , *
)
θ
V
Ω
S 0
θ
P1*
(P1* , *
)
P1
(a) (b)
Fig. 1. (a) Structure with applied thermomechanical loads, (b) 3-D loading domain.

may be calculated from the temperature distribution using the coef- The following theorem is useful in the numerical investigation
ficient of thermal expansion. The last two parts of (2) come from the for the shakedown threshold [60]:
inelastic behavior with the very last one being the plastic strain rate
vector. ‘‘If a structure shakes down under a cyclic loading containing all
The following relations may be written between strains and the vertices of the convex loading domain X then it shakes
stresses: down for any loading path contained in X”.

r_ el ðsÞ ¼ D  e_ el ðsÞ A proof, which highlights some important issues, is presented

next:
q_ ðsÞ ¼ D  e_ elr ðsÞ ð3Þ Let us consider a stress point at some cyclic loading path inside
e_ pl ðsÞ ¼ k_ @r@fðsÞ the convex domain. We may write the elastic solution at this point
as a linear combination [61] of the elastic solutions at its vertices:
with f being a convex yield surface, containing the stresses’ origin,
and k_ the plastic multiplier and D the material matrix. rel ¼ b1 rel1 þ b2 rel2 þ    þ bs rels ð6Þ
A stress state may be either inside or on the yield surface. A where b1 þ b2 þ    þ bs ¼ 1 and bi P 0; with i ¼ 1; . . . s, where s is
state of stress r inside the yield surface is a safe stress for which
the number of vertices of the convex loading domain, and reli are
f ðrÞ < 0, whereas if it is either on the yield surface or inside is
the elastic stresses at each vertex of this domain.
an allowable stress state for which f ðrÞ 6 0. Since by the hypothesis, a cyclic loading that passes through the
For two stress states r; q one may write the following inequality vertices of the loading domain at different time points causes the
due to the convexity of the yield surface [56]: structure to shake down, according to Melan’s theorem, a residual
f ðbr þ ð1  bÞqÞ 6 bf ðrÞ þ ð1  bÞfðqÞ ð4Þ stress q
 constant in time, and therefore common at all the vertices,
will exist, such that the total stress at each vertex will be a safe
where 0 6 b 6 1. state of stress:
Since q = 0 is a safe state of stress one may write:
f ðreli þ q
 Þ < 0; i ¼ 1; . . . s ð7Þ
f ðbrÞ 6 bf ðrÞ ð5Þ
where for a homogeneous yield function f, as, for example, a von This residual stress may be combined with rel to give:
Mises yield criterion, equality holds.
f ðrel þ q
 Þ ¼ f ðb1 rel1 þ b2 rel2 þ    þ bs rels þ q
Þ
Also, related to the convexity of the yield surface is the Druck-
er’s postulate of stability [1]. In [57] it is proved that a structure ¼ f ½b1 rel1 þ b2 rel2 þ    þ bs rels þ ðb1 þ b2 þ    þ bs Þq

whose material obeys Drucker’s postulate will reach, after many ð8Þ
cycles of loading, an asymptotic state in which the stresses and
strain rates gradually stabilize and become also cyclic having the Ps 
using the first property of the coefficients bi i¼1 bi ¼ 1 .
same period with the applied loads [58]. By grouping them, one can write the above expression as:
Shakedown is a favorable asymptotic state such that, provided
the load margins are below a threshold, the structure adapts itself, f ðrel þ q
 Þ ¼ f ½b1 ðrel1 þ q
 Þ þ b2 ðrel2 þ q
 Þ þ    þ bs ðrels þ q
 Þ
ð9Þ
after some initial plastic straining, to a purely elastic behavior. 6 b1 f ðrel1 þ q
 Þ þ b2 f ðrel2 þ q
 Þ þ    þ bs f ðrels þ q
Þ
Conditions for shakedown are given by Melan’s theorem [11].
The theorem is composed of the following two statements [59]: The above inequality is justified through the use of the general-
ized form of the convexity inequality (4), often called Jensen’s
(a) The structure will shake down under a cyclic loading if there inequality, which may be proved by induction. Thus, we conclude:
exists a time-independent distribution of residual stresses q 
) f ðrel þ q
Þ < 0 Q :E:D: ð10Þ
such that, under any combination of loads inside prescribed
limits, its superposition with the elastic stresses rel , i.e. where use of the inequalities (7) and the non-negative nature of the
rel þ q , results in a total safe stress state at any point of bi was made.
the structure, A 2-D dimensional loading cuboid is a rectangle, i.e. s = 4 (22).
(b) Shakedown never takes place unless a time-independent dis- For a 3-D loading domain (Fig. 1(b)), s = 8 (23). For an n-
tribution of residual stresses can be found such that, under all dimensional loading domain (n-cuboid) one can envisage that the
the possible load combinations, the sum of the residual and number of vertices will be equal to the combination of the loads
elastic stresses constitutes an allowable stress state. in all possible ways. Thus, one may write:
158 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171

        n  
X
n n n n n (Fig. 1(b)), and in effect the cyclic loading program (Fig. 2), is mul-
s¼ þ þ  þ  ¼ ¼ 2n ð11Þ
0 1 k n k tiplied by a loading factor cðlÞ . This may be accounted for in the
k¼0
purely elastic part of eqns. (1) and (2), which may now be written
 
n n! as:
where ¼ is the binomial coefficient and the expres-
k k!ðn  kÞ!
sion above is a standard binomial formula that may be found in any rðsÞ ¼ cðlÞ rel ðsÞ þ qðsÞ ð15Þ
mathematics handbook (e.g. [62]).
The construction of a prescribed cyclic loading program that e_ ðsÞ ¼ cðlÞ ðe_ el ðsÞ þ e_ h ðsÞÞ þ e_ elr ðsÞ þ e_ pl ðsÞ ð16Þ
may pass through the vertices of the 3-D loading domain may be
With an initial high value, the load factor is sequentially
seen in Fig. 2. We may see inside one period T the time variation
decreased by shrinking the loading domain in a continuous way,
of the three loads that follow the consecutive movements from
until the conditions of the limit cycle are reached [51].
each vertex to the next (Fig. 1(b)): 0 ! P1 ! ðP 1 ;
      The decomposition of the residual stresses in Fourier series of
h Þ ! ðP1 ; P2 ; h Þ ! ðP1 ; P2 Þ ! P 2 ! ðP 2 ; h Þ ! h ! 0. The loading
  
Eq. (13) provides a natural way to implement the transition to
vector may thus be expressed, through its maximum values:
the limit cycle. Thus, the ending phase of the procedure is when,
8 9 8  9
in the course of iterations, the only remaining terms of the Fourier
< P1 ðsÞ >
> < P1  a1 ðsÞ >
= > =
series are the constant terms a0 :
PðsÞ ¼ P2 ðsÞ ¼ P2  a2 ðsÞ ð12Þ
>
: ; >
> :  >
; As a starting point for the procedure one can use the cycle time
hðsÞ h  a3 ðsÞ point, where one of the loads attains its maximum value, whereas
where the ai denote time functions. the others are zero (e.g. s in Fig. 2). Having assumed a von Mises
type of material, the effective stresses of the elastic stresses corre-
sponding to this load value (P 1 ) at all the GPs may be calculated;
3. An enhanced RSDM-S procedure the ratio of the uniaxial yield stress rY to the minimum non-zero
of these effective stresses may serve as an initial load factor cð1Þ
The RSDM-S [51] has its roots in the RSDM [9] which is a direct
since this selection, guarantees, at least for this cycle time, that
method that can be used to determine the kind of asymptotic cyclic
all the elements of the structure will be plastic, rendering cð1Þ not
state (ratcheting, or reverse plasticity, or shakedown) for a given
only far above the shakedown but even the limit load factor [51].
cyclic history, without following cumbersome time-stepping
The flow chart of the suggested procedure may be seen in Fig. 3.
calculations.
As one may realize, it consists of two iterative loops one inside the
The main idea of the RSDM is to decompose the sought cyclic
other. The Greek letters j and l are used to denote an iteration of
residual stresses in the asymptotic cycle to Fourier series (13),
the inner and the outer loop respectively. They both have 1 as the
1 X
1 initial values.
qðsÞ ¼ a0 þ fcosð2kpsÞ  ak þ sinð2kpsÞ  bk g ð13Þ At an iteration j a total stress is found that consists of an exist-
2 k¼1
ing estimate of the residual stresses together with the multiplied
Differentiating (13) with respect to time, one may get an by the current loading factor elastic stresses, which are evaluated
expression for their derivatives: from the contributions of the two mechanical and the thermal
( ) load. If more loads act on the structure they could be easily accom-
X1
modated here. For the thermal load, in particular, the stresses are
q_ ðsÞ ¼ 2p ðk sinð2kpsÞ  ak þ k cosð2kpsÞ  bk ð14Þ
k¼1
calculated using (17)

It is proved that the Fourier coefficients ak ; bk ; a0 may be found relh ¼ D  B  rh  D  eh ð17Þ
in an iterative way through the evaluation of these derivatives.
This evaluation is performed by satisfying, at cycle time points, where rh is the vector of nodal displacements due to the peak of the
equilibrium, with zero loads, and compatibility [9]. thermal load that may be calculated from (18), with B being the
Focusing now on shakedown, the two statements of Melan’s strain-displacement compatibility matrix in a FE environment and
theorem define, for a prescribed cyclic loading, the limit cycle K being the standard stiffness matrix:
which is a transition cycle between one with plastic straining Z


and one without plastic straining. It may be proved [59] that the K  rh ¼ BT  D  eh dV ð18Þ
V
residual stress distribution of this cycle is unique, being indepen-

dent of the preceding deformation history. where the thermal strains eh , in (17) and (18), are calculated pro-
The numerical procedure RSDM-S is actually an iterative transi- portional to the temperature h , through the coefficient of thermal
tion process to this cycle. In every iteration (l) the loading domain expansion.

P1, P2, θ

P1*

P2*
*
θ

τ* 1
0
τ1 τ2 τ3 τ4 τ5 τ6 τ8 τ
τ7
1

Fig. 2. Individual cyclic load variation over one time period.

 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171 159

START

(for each GP)

(for each cycle point)
(Next iteration (κ+1)
(Next iteration (μ+1)

STOP

Fig. 3. Flow chart of the enhanced RSDM-S.

The equivalent stress r  ðjÞ is used to check whether the total amount of plastic straining is equivalenced [9] with the plastic
! !
stress vector rpl which is part of the radial vector OC and its mag-
ðjÞ
stress (vector OC in Fig. 4) exceeds the yield surface. Since no exact
knowledge of the plastic strain rate e_ pl !
ðjÞ (Fig. 4) is needed, the nitude is determined from the magnitude of OC , proportionally
160 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171

C
vector :

vector :

A
B
D

Fig. 4. Estimation of the plastic straining inside an iteration.

not accepted area

γ(μ)
(μ+1)
γ
γsh
φ(γsh)>tol
φ(γ(μ+1))<tol
scaled by ω
no. of iteration
γ(μ+1)<0

Fig. 5. Convergence scheme of the RSDM-S.

through the ratio n (Fig. 3). After applying this procedure to every ðjþ1Þ ðjÞ ðjþ1Þ ðjÞ
ak ! ak ; bk ! bk which, in view of (14), guarantees that
Gauss point (GP) an equivalent plastic nodal vector is added to the the residual stress rate solution has been stabilized, for the current
nodal forces due to the mechanical and thermal loads to give R_ 0 ðsÞ loading factor. This is a new convergence criterion which, keeping
(whose expression may be seen in Fig. 3), which is then used to the same accuracy, leads to a substantial reduction of the inner
provide an estimate of the rate of the residual stresses q_ ðsÞ at loop iterations as compared to the one adopted in the original pub-
the cycle point s. Both these two expressions may be derived by lication [51] where the converged solution was judged on the
combining Eqs. (16) and (3) and the fact that the residual stress updates of the full expressions of the residual stresses, including
rates are self-equilibrated (see, for example, [9,51]). Here again the constant terms.
these two expressions would be augmented if more loads are The value of u, being always positive, is then used to decrease
applied. the loading factor and thus perform the outer loop iterations which
The calculations are carried over for every cycle point and, by stop when the value of u becomes virtually zero. This may be
performing numerical time integration over all the cycle points, expressed through the following double inequality, which is also
an update of the Fourier coefficients leading to an update of the a different criterion to the one used in the original RSDM-S [51]:
residual stresses is obtained.
An update of the function u which is the sum of the norms of tol < uðcðlÞ Þ < tol þ d ð19Þ
the coefficients ak and bk in front of the trigonometric parts of
the Fourier series is then calculated. This is checked against the One obtains accurate results for a value of tol  103 with d
previous iterate of u and if it is found to be within a specified tol- being an error tolerance, e.g. 104 .
erance, we exit the inner loop [63]. The fact that two consecutive A convergence parameter x is needed (Fig. 3) so that the proce-
values of u are virtually the same, leads to the conclusion that dure always guarantees to converge. The numerical strategy is to
 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171 161

Fig. 6. Geometry, loading and finite element mesh of the frame.

Table 1 overshooting may result to even negative value of cðlþ1Þ (Fig. 5);
Material properties of the frame.
the same strategy by halving x is followed till we get a positive
Young’s modulus E = 200 GPa value of cðlþ1Þ which also satisfies uðcðlþ1Þ Þ > tol.
Poisson’s ratio m = 0.3 It should be noted here that when the conditions for shakedown
Yield stress rY = 100 MPa
have been reached, the plastic stress vector, as expected, also
approaches zero, within some tolerance, at all the cycle points [51].

4. 2-D loading domain examples

start the procedure with the convergence parameter x = 1. For The numerical efficiency of the enhanced RSDM-S is demon-
many numerical examples, this normally leads to a monotonic con- strated in a couple of examples consisted of a frame and a contin-
vergence, from above, to the shakedown load. uous beam subjected to two independent loads. Both the examples
It could happen however, especially when starting from a high are modelled using quadrilateral continuum elements. Two differ-
initial value that an overshooting of the shakedown factor occurs. ent loading domains are considered that show the versatility of the
This means that the procedure bypasses the predefined tolerance approach.
tol for u. To deal with the overshooting, a numerical convergence
scheme is followed, which is depicted in Fig. 5. Thus, a loading fac- 4.1. Frame example
tor cðlþ1Þ is evaluated at the current iteration l + 1 for which
uðcðlþ1Þ Þ < tol; this cannot be accepted (inequality (19)) and the The first example is the frame of Fig. 6. The frame is assumed
convergence factor is continuously halved until we get a loading homogeneous, isotropic, elastic-perfectly plastic, having the mate-
factor for which uðcðlþ1Þ Þ > tol. In some other case an even bigger rial data shown in Table 1. The finite element mesh discretization

P1
α1 P2
α2
1

P2
0.8
-16τ2+8τ -16τ2+16τ-3
0.6 3
α1,2 (τ)

P 2* (P1*, P2*)
-16τ2+16τ-3 -16τ2+24τ-8
0.4
4 2

0.2
0
1 P 1* P1
0
0 0.25 0.5 0.75 1
Time τ
(a) (b)
Fig. 7. Independent cyclic load variation over one time period in (a) time domain, (b) load domain.
162 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171

12

11

10

Load factor γ 8
improved convergence
7 criterion
6 old convergence
criterion
5

2
0 100 200 300 400
no. of iterations
Fig. 8. Comparison of the two convergence criteria for the frame example.

of the frame, shown also in Fig. 6, consists of 400 eight-noded, iso- A prescribed loading in time domain that passes through the
parametric elements with 3  3 Gauss integration points. The four vertices of the rectangle, in a consecutive manner, may be
frame is subjected to two uniformly distributed loads P 1 ðsÞ and defined using the following equations:
P2 ðsÞ, applied on the external faces of AB and BC, respectively.    
P 1 ð sÞ P1 a1 ðsÞ
PðsÞ ¼ ¼ where the time functions a1 ðsÞ; a2 ðsÞ are :
P 2 ð sÞ P2 a2 ðsÞ

4.1.1. Standard loading domain

A rectangular loading domain is considered with the two loads a1 ðsÞ ¼ 9:6s2 þ 4:8s þ 0:4; a2 ðsÞ ¼ 0:4; s 2 ½0; 1=4
varying independently (Fig. 7), having maximum values P1 ¼ 3 and a1 ðsÞ ¼ 1; a2 ðsÞ ¼ 9:6s2 þ 9:6s  1:4; s 2 ð1=4; 1=2
P2 ¼ 1, respectively, while their minimum values are equal to zero
ð20Þ
a1 ðsÞ ¼ 9:6s2 þ 9:6s  1:4; a2 ðsÞ ¼ 1; s 2 ð1=2; 3=4
[51].
This example has been investigated in [39] using an edge-based
a1 ðsÞ ¼ 0:4; a2 ðsÞ ¼ 9:6s2 þ 14:4s  4:4; s 2 ð3=4; 1
smoothed finite element method (ES-FEM) and a primal-dual In this case P1 ¼ 3, P 2 ¼ 1 and 0:4 6 a1 ðsÞ; a2 ðsÞ 6 1 (see
shakedown algorithm, and in [49] using a strain driven strategy. Fig. 10).
In those applications, however, the loading domains have one or For this example, the initial convergence parameter x, in the
both the minimum values different to zero unlike the loading process of the iterations, had to be halved twice, for the RSDM-S
domain used here, where the minimum values are both assumed to converge to the final shakedown limit which was found equal
zero. This more general kind of loading domain will be studied in to 3.91.
the next section. The comparison between the enhanced RSDM-S solution and
A prescribed loading in the time domain that passes through those of different analysis methods in the literature, are shown
the four vertices of the rectangle, in a consecutive manner (Fig. 7 in Table 2. It may be seen that they match quite well.
(b)), may be defined using the time functions a1 ðsÞ; a2 ðsÞ of Fig. 7
(a).
4.2. Symmetric continuous beam under distributed load
For this problem, the initial convergence parameter x, in the
process of the iterations, had to be halved twice, for the procedure
A symmetric three-span continuous beam under uniform loads
to converge to the final shakedown factor which was found equal
is considered next. Due to symmetry, only half of the beam is ana-
to 2.47.
A number of 175 iterations were required to converge. The
amount of CPU time needed to solve this problem, for an Intel Core 1
i7 at 2.93 GHz with 4096 MB RAM, was around 160 s.
On the other hand, using the old convergence criterion, conver- 0.8
gence was achieved after 280 s (CPU time), with 354 iterations [51].
Thus, it is realized that using the proposed enhanced procedure
0.6
P2

there is a significant acceleration (about 43%).

The convergence of the two procedures, the old and the new
one, may be seen in Fig. 8. For better illustration reasons, the com- 0.4
mon 10 first iterations are not plotted.
0.2

4.1.2. Loading domain with different than zero origin 0

0 0.6 1.2 1.8 2.4 3
A rectangular loading domain is considered herein with the two P1
loads P 1 ðsÞ and P2 ðsÞ varying independently, between the values
½1:2; 3 and ½0:4; 1 respectively (Fig. 9). Fig. 9. Loading domain of the frame example.
 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171 163

α1 α2
1

α1(τ)=-9.6τ2+4.8τ+0.4 α2(τ)=-9.6τ2+14.4τ-4.4
0.8
α2(τ)=-9.6τ2+9.6τ-1.4

0.6 α1(τ)=-9.6τ2+9.6τ-1.4
α1,2(τ)
0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
Time τ
Fig. 10. Time functions variation, over one period corresponding to the load domain of Fig. 9.

Table 2 Table 3
Comparison of numerical results of the frame. Material properties of the symmetric con-
tinuous beam example.
Author Shakedown factor
Garcea et al. [48] 3.925 Young’s modulus E = 180 GPa
Tran et al. [39] 4.006 Poisson’s ratio m = 0.3
Pham [64] 4.015 Yield stress rY = 100 MPa
Present 3.91

The beam is assumed homogeneous, isotropic, elastic-perfectly

plastic, having the material data of Table 3.

4.2.1. Standard loading domain

The rectangular loading domain of Fig. 7 was considered first.
The loads vary in the domain P 1 2 ½0; P 1  and P 2 2 ½0; P2  where
P1 ¼ 1 and P 2 ¼ 2 [51].
This example has also been treated in [48,64] using a different
loading domain, that has its origin different to zero, to the one
employed here. This particular load domain will be studied in
the next section.
A prescribed loading in the time domain that passes through
the four vertices of the rectangle, may be defined by using the
same time functions a1 ðsÞ; a2 ðsÞ of Fig. 7.
For this example, the initial convergence parameter x, in the
process of the iterations, had to be halved three times, for the pro-
cedure to converge to the final shakedown factor which was found
equal to 0.191. Once again although the starting point was quite
high as compared to the final result, the initial descent was rapid
in just 12 iterations, as shown in Fig. 12.
The convergence of the enhanced RSDM-S, based on the new
criterion, and its comparison with the old one, may be seen in
Fig. 12. For a better illustration of the comparison, the first 10 iter-
ations are not plotted.
A number of 65 iterations were required to converge, whereas
100 iterations were needed with the old criterion [51]. The amount
of CPU time needed to solve this problem was around 220 s which
is around 40% faster.

4.2.2. Loading domain with different than zero origin

Fig. 11. Geometry, loading and finite element mesh of the continuous beam. A rectangular loading domain is considered (Fig. 13) with the
two loads varying independently, as P 1 2 ½1:2; 2; P 2 2 ½0; 1.
A prescribed loading in the time domain that passes through
lyzed (Fig. 11). This is a relatively large scale problem since its the four vertices of the rectangle, in a consecutive manner, may
finite element discretization consists of 800 eight-noded, iso- be defined by using the following time functions a1 ðsÞ; a2 ðsÞ
parametric elements with 3  3 Gauss integration points (Fig. 11). (Fig. 14):
164 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171

2.25

1.75

1.5

Load factor γ
1.25
old convergence
1 criterion
0.75 improved convergence
criterion
0.5

0.25

0
0 20 40 60 80 100
no. of iterations

Fig. 12. Comparison of the two convergence criteria for the beam example.

0.8

0.6
P2

0.4

0.2

0
0 0.4 0.8 1.2 1.6 2
P1

Fig. 13. Loading domain of the beam example.

α1 α2
1

α1(τ)=-6.4τ2+3.2τ+0.6
0.8 α1(τ)=-6.4τ2+6.4τ-0.6
α1,2(τ)

0.6
α2(τ)=-16τ2+16τ-3
0.4
α2(τ)=-16τ2+24τ-8
0.2

0
0 0.2 0.4 0.6 0.8 1
Time τ

Fig. 14. Time functions variation, over one period corresponding to the load domain of Fig. 13.

a1 ðsÞ ¼ 6:4s2 þ 3:2s þ 0:6; a2 ðsÞ ¼ 0; s 2 ½0; 1=4 In this case P 1 ¼ 2, P 2 ¼ 1 and 0:6 6 a1 ðsÞ 6 1; 0 6 a2 ðsÞ 6 1
a1 ðsÞ ¼ 1; a2 ðsÞ ¼ 16s2 þ 16s  3; s 2 ð1=4; 1=2 (Fig. 14).
ð21Þ For this example, the initial convergence parameter x, in the
a1 ðsÞ ¼ 6:4s2 þ 6:4s  0:6; a2 ðsÞ ¼ 1; s 2 ð1=2; 3=4 process of the iterations, had to be halved three times, for the
a1 ðsÞ ¼ 0:6; a2 ðsÞ ¼ 16s2 þ 24s  8; s 2 ð3=4; 1 RSDM-S to converge to the final shakedown limit which was found
 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171 165

Fig. 15. Convergence of the RSDM-S towards the shakedown factor for the continuous beam problem.

Table 5
Table 4 Material properties of the plate.
Comparison of numerical results of the sym-
metric continuous beam. Young’s modulus E = 208 GPa
Poisson’s ratio m = 0.3
Author Shakedown limit Yield stress rY = 360 MPa
Garcea et al. [48] 3.244 Coefficient of thermal expansion 5  105 °C1
Tran et al. [39] 3.377
Pham [64] 3.264
Present 3.177 5. 3-D loading domain example

Let us consider the holed square plate of Fig. 16 subjected to a

three-dimensional loading consisting of a thermal load, i.e. a tem-
equal to 3.177. Less than 50 iterations were required for this prob- perature difference DhðsÞ between the edge of the hole and the
lem to converge (Fig. 15). In the insert of the figure, one may see, edge of the plate, and two uniformly distributed mechanical loads
after the initial descent, the last iterates towards the shakedown P1 ðsÞ and P2 ðsÞ. The plate is assumed homogeneous, isotropic,
value. elastic-perfectly plastic with the material data of Table 5.
The shakedown factor obtained by the enhanced RSDM-S, and Concerning the geometrical characteristics of the plate, the ratio
its comparison with the results of different analysis methods, is between the diameter D of the hole and the length L of the plate is
shown in Table 4. It may be seen that there is a good equal to 0.2 and the ratio of the thickness d of the plate to its length
agreement. is equal to 0.05. For this study L = 20 cm has been chosen.

P2(τ)
θ0

L
θ1 P1(τ)

Fig. 16. Geometry, and finite element discretization of the plate subjected to mechanical and thermal loading.
166 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171

Table 6
Numerical results of shakedown analysis in three-dimensional loading space.

(P*1, P*2, Dh*) P1/rY P2/rY rt/2rY

(1, 0, 0) 0.7 0 0
(0, 1, 0) 0 0.7 0
(0, 0, 1) 0 0 1.002
(1, 1, 0) 0.522 0.522 0
(1, 0, 1) 0.566 0 0.196
(1, 1, 1) 0.448 0.448 0.156
(1, 0.5, 1) 0.507 0.254 0.176
(1, 0.5, 0.5) 0.547 0.274 0.048
(1, 1, 0.5) 0.484 0.484 0.042
(0.5, 0.5, 1) 0.388 0.388 0.269
(0.5, 1, 0.5) 0.275 0.549 0.048

Fig. 17. Three - dimensional load domain.

Due to the symmetry of the structure and the loading, only one
quarter of the plate is analyzed. The finite element mesh dis-
cretization of the plate is also shown in Fig. 16. Ninety-eight,
eight-noded, iso-parametric elements with 3  3 Gauss integration
points were used.
The variation of the temperature with radius r has the same log-
arithmic distribution as in [52]:

DhðsÞ  ln 5D=2
r
hðr; sÞ ¼ h0 þ ð22Þ
ln 5
The above relation describes the temperature variation inside Fig. 19. Shakedown domain in three-dimensional loading space.
the plate giving a value of h1 ðsÞ ¼ h0 þ DhðsÞ around the edge of
the hole ðr ¼ D=2Þ and h1 ¼ h0 at the outer edges of the plate
ðr ¼ 5D=2Þ. The temperature h0 is assumed to be equal to zero. It In the same figure one can see, starting from the origin, the suc-
should be noted that, in the results, rt denotes the maximum cessive movement that passes through the eight vertices of the
effective thermal elastic stress due to the fluctuating temperature. loading domain that may define a prescribed loading. This loading
The loads vary independently in the three-dimensional loading may be realized in the time domain by the following expression:
8  9
domain of Fig. 17 having the following variations: < P1 a1 ðsÞ =

PðsÞ ¼ P a ðsÞ , where the time functions
P1 2 ½0; P1 ; P2 2 ½0; P2 ; Dh 2 ½0; Dh  : 2 2 ;
Dh a3 ðsÞ
where the maximum values are P1 ¼ P 2 ¼ Dh . a1 ðsÞ; a2 ðsÞ; a3 ðsÞ are (Fig. 18):

θ∇
P2 P1
1

-64τ2+16τ -64τ2+64τ-15 -64τ2+112τ-48

0.8
-64τ2+32τ-3 -64τ2+96τ-35
0.6
α1,2,3(τ)

-64τ2+48τ-8 -64τ2+96τ-35
0.4

-64τ2+48τ-8
0.2

0
0 1/8 2/8 3/8 4/8 5/8 6/8 7/8 1
Time (τ)
Fig. 18. Time functions variation, over one period corresponding to the three-dimensional load domain of Fig. 17.
 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171 167

1 the total 3D loading domain. In Fig. 24 one may see a typical con-
vergence behavior of the RSDM-S i.e. for the fixed ratios
0.8 P1 =P2 =Dh ¼ 1. The initial value of x ¼ 1 was sufficient for
convergence.
0.6 An important observation is that both the CPU time and the
σt/2σY

iterations needed for the RSDM-S to converge in the case of 3D

P1/P2=1 loading domain, are of the same order with the ones required for
0.4
P1/P2=0 the case of the 2D loading domain [52]. On the contrary, for the
0.2
problem solved through the use of an IPM algorithm, there is a sig-
nificant increase of the running time between the two cases, since
the number of variables and the number of constraints, in the
0
0 0.2 0.4 0.6 0.8 1 three-dimensional loading case are virtually twice the number in
the two-dimensional case [55].

Fig. 20. Shakedown domains in planes for fixed ratios (P*1/P*2).

6. Generalization for n-dimensional loading domain
a1 ðsÞ ¼ 0; a2 ðsÞ ¼ 64s2 þ 16s; a3 ðsÞ ¼ 0; s 2 ½0;1=8
As it may also be seen by the proof in Section 2, the theorem of
a1 ðsÞ ¼ 0; a2 ðsÞ ¼ 1; a3 ðsÞ ¼ 64s2 þ 32s  3; s 2 ð1=8; 2=8 König and Kleiber [60] is not associated with any specific form of a
a1 ðsÞ ¼ 64s2 þ 48s  8; a2 ðsÞ ¼ 1; a3 ðsÞ ¼ 1; s 2 ð2=8;3=8 cyclic loading. Thus, one may use any prescribed cyclic loading,
a1 ðsÞ ¼ 1; a2 ðsÞ ¼ 1; a3 ðsÞ ¼ 64s2 þ 48s  8; s 2 ð3=8;4=8 which passes through the vertices that define the loading domain,
as long as it is on or within the boundaries of this domain. One
a1 ðsÞ ¼ 1; a2 ðsÞ ¼ 64s2 þ 64s  15; a3 ðsÞ ¼ 0; s 2 ð4=8;5=8 would expect, of course, a non-uniqueness of the residual stresses
a1 ðsÞ ¼ 1; a2 ðsÞ ¼ 0; a3 ðsÞ ¼ 64s2 þ 96s  35; s 2 ð5=8;6=8 of the limit cycle for the different prescribed loadings.
a1 ðsÞ ¼ 64s2 þ 96s  35; a2 ðsÞ ¼ 0; a3 ðsÞ ¼ 1; s 2 ð6=8;7=8 The cyclic loading used so far, follows the boundaries of the
domain and it may be automated for 2D and 3D loading domains.
a1 ðsÞ ¼ 0; a2 ðsÞ ¼ 0; a3 ðsÞ ¼ 64s2 þ 112s  48; s 2 ð7=8;1 In cases, however, of more than three independent loads, this
ð23Þ automation becomes more difficult.
We enter the discussion towards this automation, by consider-
In Table 6 one may see the resulted shakedown factor for some
ing an alternative cyclic loading program for a 2D loading domain,
specific ratios of P1 =P2 =Dh .
which may be seen in Fig. 25(a). This loading comprises of a move-
Finally, the 3D shakedown domain of the problem is shown in
ment going from the origin to each vertex and coming back, before
Fig. 19. It should be mentioned that each solution for different
moving to the next vertex. The time functions that may describe
fixed ratios P1 =P2 , represents a fixed angle u in the P1  P 2 -plane.
such movement, inside a period, can simply be sine functions
Thus, the results may be presented as a sequence of two dimen-
(Fig. 25(b)). For s1 ¼ s2 ¼ s3 ¼ 1=3, these functions may be written
sional plots. In Figs. 20, 21, two-dimensional plots of the results
as:
obtained by RSDM-S, for different fixed ratios P 1 =P 2 , are presented.
8
Plotting the results in the three planes, the horizontal plane < sinð3psÞ; s 2 ½0; 1=3
>
(Fig. 22) or either of the two (due to the symmetry) vertical ones aðsÞ ¼ sinð3ps  pÞ; s 2 ð1=3; 2=3 ð24Þ
(Fig. 23), one may see the results of the RSDM-S for a two- >
:
sinð3ps  2pÞ; s 2 ð2=3; 1
dimensional loading case consisted of a mechanical and a thermal
or of two mechanical loads. These results are in perfect agreement To reach all the three vertices of the loading domain, each of
with the results of the method that was formulated for 2- these time functions are multiplied by either P 1 or P 2 , separately,
dimensional loading [52]. or simultaneously (Fig. 25(b)).
The CPU time needed for the RSDM-S to converge, for a typical Equivalently moving from the origin and back to each of the
case of P1 ¼ P2 ¼ Dh , was about 50 s on the same, as above, proces- seven vertices in a 3D loading domain (Fig. 26(a)), an alternative
sor. A total number of 40 time points proved enough to describe cyclic loading program to the one proposed previously, may be

0.8

0.6
σt/2σY

0.4 P1/P2=1/0.5
P1/P2=1/0.2

0.2

0
0 0.2 0.4 0.6 0.8 1

Fig. 21. Shakedown domains in planes for fixed ratios (P1*/P2*).

168 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171

Fig. 22. Shakedown domain produced by the RSDM-S for a two dimensional loading case (independent variation of thermal and mechanical load, rectangular loading
domain) [52].

seen in Fig. 26(b). The sine time functions that divide, equally and
consecutively a period, in this case, take the form:
1 8
>
> sinð7psÞ; s 2 ½0; 1=7
0.9 >
>
>
> sinð7ps  pÞ; s 2 ð1=7; 2=7
>
>
>
>
< sinð7ps  2pÞ; s 2 ð2=7; 3=7
>
0.8

0.7 aðsÞ ¼ sinð7ps  3pÞ; s 2 ð3=7; 4=7 ð25Þ

>
>
>
> sinð7ps  4pÞ; s 2 ð4=7; 5=7
0.6 >
>
>
>
P2/σY

> sinð7ps  5pÞ; s 2 ð5=7; 6=7

>
0.5 >
:
RSDM-S sinð7ps  6pÞ; s 2 ð6=7; 1
0.4
In order to reach the seven vertices of the 3D loading domain,
0.3 we may see from Fig. 26(b) that the first three time functions are
0.2 multiplied by the maximum values of the three loads, the next
three by their combination under pairs and the last one by all three
0.1 of them.
0 The use of the sine functions for both the 2D and 3D loading
0 0.2 0.4 0.6 0.8 1 domain contributes, obviously, towards the automation of the
P1/σy loading program. The above alternative loadings paths are applied
for the example of the holed squared plate of Fig. 16. A 2D loading
Fig. 23. Shakedown domain produced by RSDM-S for a two-dimensional loading case for P1 ¼ P2 (Fig. 23), and a 3D loading case for P1 ¼ P 2 ¼ Dh ;
case (without temperature, rectangular loading domain) [51].
obtained, using the same number of time points, virtually the same

3
P1/σY

0
0 20 40 60 80 100 120
no. of iterations

Fig. 24. Convergence of the RSDM-S towards the shakedown factor for the three-dimensional loading case.
 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171 169

P2 P1, P2

P2 * (P1*, P2*)
P1 *
3
P2 *
2 Ω
0 0
* τ
1 P1 P1 τ1 τ2 τ3

(a) (b)
Fig. 25. Alternative cyclic loading program for the 2D case.

P2

P2 *

2 6
Ω 7
5
3 θ*
1 θ
P1 * 4

P1 (a)

P1, P2, θ

P 1*
θ*
P2*

0
τ4
τ
τ1 τ2 τ3 τ5 τ6 τ7
(b)
Fig. 26. Alternative cyclic loading program for the 3D case.

shakedown load of 0:52rY and 0:45rY respectively, that was its origin and m ¼ 2n  1 vertices, one may write the m sine time
obtained with the original load paths of Figs. 7 and 17. For the functions for the equal splitting of the period:
3D case, in particular, one may see the convergence behavior of
l1 l
the two different loading paths, using the same number of 40 time al ðsÞ ¼ sinðmps  ðl  1ÞpÞ; 6s6 ; l ¼ 1; . . . ; m ð26Þ
points, in Fig. 27. As it may be seen, the penalty to pay towards the
m m
automation of the loading program is that the alternative load path Since each vertex of the n-cuboid, except for the origin, corre-
requires a relatively bigger amount of iterations to converge to the sponds to a combination of the maximum values of each load, it
 
shakedown load. n
may be seen from (11) that the first time functions should
The discussion above paves the way to generalize for any num- 1  
ber of loads. Thus, for an n-dimensional loading domain, defined by n
be multiplied by each of these values, the next time functions
2
170 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171

4.5

3.5

P1/σΥ
2.5 alternative
loading program
2
original loading
1.5
program
1

0.5

0
0 50 100 150
no. of iterations
Fig. 27. Convergence for the two different cyclic loading programs (3D case).

 
n cient with many loads, since, irrespective of their number, the
by their combination every two, the next by their combina-
3 loads participate just through their corresponding elastic stres-
tion every three, and so on. These combinations may be retrieved ses, which are calculated, once and for all, at the beginning of
using any combinatorial algorithm (e.g. [65]). Thus, the whole pro- the algorithm.
cedure in any n-dimensional loading space may be fully auto-
mated. This effective and robust formulation is of particular
importance since, when designing structures, there will be cases
References
where the number of independent loads would be much higher
than three (e.g. [66,67]). It should also be underlined that the pro- [1] Drucker DC. A definition of stable material. ASME J Appl Mech 1959;26:101–6.
cedure may be embedded in any existing FE code. [2] Zarka J, Engel JJ, Inglebert G. On a simplified inelastic analysis of structures.
Nucl Eng Des 1980;57:333–68.
[3] Zarka J. Direct analysis of elastic–plastic structures with overlay materials
7. Conclusions during cyclic loading. Int J Numer Meth Eng 1980;15:225–35.
[4] Hübel H. Simplified theory of plastic zones for cyclic loading and multilinear
hardening. Int J Press Vess Pip 2015;129:19–31.
An efficient novel iterative numerical procedure, called RSDM-S,
[5] Ponter ARS, Carter KF. Shakedown state simulation techniques based on linear
for the shakedown analysis of elastoplastic structures loaded by elastic solutions. Comput Methods Appl Mech Eng 1997;140:259–79.
independently varying cyclic loads, was presented. The method is [6] Ponter ARS, Chen HF. A minimum theorem for cyclic loading in excess of
shakedown, with applications to the evaluation of a ratchet limit. Eur J Mech A/
improved and enhanced over the original approach that appeared
Solids 2001;20:539–54.
recently in the literature and concerned a 2D loading domain. [7] Maitournam MH, Pommier B, Thomas JJ. Détermination de la réponse
The following improvements and enhancements have been pre- asymptotique d’une structure anélastique sous chargement
sented herein: thermomécanique cyclique. CR Mec 2002;330:703–8.
[8] Zouain N, SantAnna R. Computational formulation for the asymptotic response
of elastoplastic solids under cyclic loads. Eur J Mech A/Solids 2017;61:267–78.
 Both approaches, in the course of iterations use two iteration [9] Spiliopoulos KV, Panagiotou KD. A direct method to predict cyclic steady states
loops, one inside the other. In the present approach, a new con- of elastoplastic structures. Comput Methods Appl Mech Eng 2012;223:186–98.
[10] Spiliopoulos KV, Panagiotou KD. The residual stress decomposition method
vergence criterion is proposed which reduces the inner loop (RSDM): a novel direct method to predict cyclic elastoplastic states. In:
iterations with an effect to make the current version twice as Spiliopoulos K, Weichert D, editors. Direct methods for limit states in
fast. structures and materials. Dordrecht: Springer Science + Business Media;
2014. p. 139–55.
 The present approach is formulated in a 3D thermomechanical [11] Melan E. Zur Plastizität des raümlichen Kontinuums. Ing Arch 1938;9:116–26.
loading domain, using time functions that follow the outline [12] Koiter WT. General theorems for elastic-plastic structures. In: Sneddon IN, Hill
of this domain. It has to be noted that very few shakedown R, editors. Progress in solid mechanics. Amsterdam: North-Holland Publ Co;
1960. p. 165–221.
results exist in the literature, for such a domain, although it
[13] Prager W. Shakedown in elastic-plastic media subjected to cycles of load and
may often occur in practice. Extending the dimensions of the temperature. In: Proc Symp Plasticita nella Scienza delle Construzioni,
loading domain from 2D to 3D hardly increases the computa- Bologna. p. 239–44.
[14] de Donato O. Second shakedown theorem allowing for cycles of both loads and
tional time, which is not the case with any procedure that uses
temperature. 1st Lomb Sci Lett (A) 1970;104:265–77.
a MP algorithm, where the computational time more than [15] Weichert D. On the influence of geometrical nonlinearities on the shakedown
doubles. of elastic–plastic structures. Int J Plast 1986;2:135–48.
 Next, a generalization of the procedure is proposed, so that it [16] Gross-Weege J. A unified formulation of statical shakedown criteria for
geometrically nonlinear problems. Int J Plast 1990;6:433–47.
applies to any multiple loading domain. As a result of this [17] Belouchrani MA, Weichert D. An extension of the static shakedown theorem to
domain, a convex hull that consists of vertices that mark all inelastic cracked structures. Int J Mech Sci 1999;41:163–77.
the possible combinations of the load peaks may be visualized. [18] Pham D. Shakedown theory for elastic plastic kinematic hardening bodies. Int J
Plast 2007;23:1240–59.
A cyclic loading program which visits each vertex and comes [19] Stein E, Zhang G, König JA. Shakedown with nonlinear strain-hardening
back to the origin is suggested. The use of sine-type time func- including structural computation using finite element method. Int J Plast
tions that split the period of the cyclic loading, combined with a 1992;8:1–31.
[20] Simon JW. Direct evaluation of the limit states of engineering structures
combinatorial algorithm represent these visits effectively and exhibiting limited, nonlinear kinematical hardening. Int J Plast
automate the procedure. The approach appears also very effi- 2013;42:141–67.
 K.V. Spiliopoulos, K.D. Panagiotou / Computers and Structures 193 (2017) 155–171 171

[21] Bouby C, Kondo D, de Saxcé G. A comparative analysis of two formulations for [44] Cecot W. Application of h-adaptive FEM and Zarka’s approach to analysis of
nonlinear hardening plasticity models: application to shakedown analysis. Eur shakedown problems. Int J Numer Meth Eng 2004;61:2139–58.
J Mech A/Solids 2015;53:48–61. [45] Krabbenhøft K, Lyamin AV, Sloan SW. Bounds to shakedown loads for a class of
[22] Pycko S, Maier G. Shakedown theorems for some classes of non-associative deviatoric plasticity models. Comput Mech 2007;39:879–88.
hardening elastic–plastic material models. Int J Plast 1995;11:367–95. [46] Chen HF, Ponter ARS. Shakedown and limit analyses for 3-D structures using
[23] Bousshine L, Chaaba A, de Saxcé G. A new approach to shakedown analysis for the linear matching method. Int J Press Vess Pip 2001;78:443–51.
non-standard elastoplastic material by the bipotential. Int J Plast [47] Ponter ARS. Shakedown limit theorems for frictional contact on a linear elastic
2003;19:583–98. body. Eur J Mech A/Solids 2016;60:17–27.
[24] Polizzotto C. Shakedown theorems for elastic–plastic solids in the framework [48] Garcea G, Armentano G, Petrolo S, Casciaro R. Finite element shakedown
of gradient plasticity. Int J Plast 2008;24:218–41. analysis of two-dimensional structures. Int J Numer Meth Eng
[25] Polizzotto C. Shakedown analysis within the framework of strain gradient 2005;63:1174–202.
plasticity. In: Fuschi P, Pisano AA, Weichert D, editors. Direct methods for limit [49] Garcea G, Leonetti L. A unified mathematical programming formulation of
and shakedown analysis of structures. Switzerland: Springer International strain driven and interior point algorithms for shakedown and limit analysis.
Publishing; 2015. p. 221–52. Int J Numer Meth Eng 2011;88:1085–111.
[26] Maier G. Shakedown theory in perfect elastoplasticity with associated and [50] Hjiaj M, Krabbenhøft K, Lyamin AV. Direct computation of shakedown loads
nonassociated flow-laws: a finite element, linear programming approach. via incremental elastoplastic analysis. Fin Elem Anal Des 2016;122:39–48.
Meccanica 1969;4:1–11. [51] Spiliopoulos KV, Panagiotou KD. A residual stress decomposition based
[27] Cohn MZ, Maier G. Engineering plasticity by mathematical programming. In: method for the shakedown analysis of structures. Comput Methods Appl
Proceedings of NATO advanced study institute. New York: Pergamon Press; Mech Eng 2014;276:410–30.
1977. [52] Spiliopoulos KV, Panagiotou KD. A numerical procedure for the shakedown
[28] Zhang T, Raad L. An eigen-mode method in kinematic shakedown analysis. Int J analysis of structures under thermomechanical loading. Arch Appl Mech
Plast 2002;18:71–90. 2015;85:1499–511.
[29] Ngo NS, Tin-Loi F. Shakedown analysis using the p-adaptive finite element [53] Spiliopoulos KV, Panagiotou KD. RSDM-S: a method for the evaluation of the
method and linear programming. Eng Struct 2007;29:46–56. shakedown load of elastoplastic structures. In: Fuschi P, Pisano AA, Weichert
[30] Ardito R, Cocchetti G, Maier G. On structural safety assessment by load factor D, editors. Direct methods for limit and shakedown analysis of
maximization in piecewise linear plasticity. Eur J Mech A/Solids structures. Springer International Publishing; 2015. p. 159–75.
2008;27:859–81. [54] Panagiotou KD, Spiliopoulos KV. Assessment of the cyclic behavior of
[31] Heitzer M, Pop G, Staat M. Basis reduction for the shakedown problem for structural components using novel approaches. ASME J Press Vess Tech
bounded kinematic hardening material. J Glob Opt 2000;17:185–200. 2016;138. 0412-01–041201-10.
[32] Zouain N, Borges L, Silveira JL. An algorithm for shakedown analysis with [55] Simon JW, Weichert D. Shakedown analysis with multidimensional loading
nonlinear yield function. Comput Methods Appl Mech Eng 2002;191:2463–81. spaces. Comput Mech 2012;49:477–85.
[33] Andersen KD, Christiansen E, Cohn AR, Overton ML. An efficient primal-dual [56] Luenberger DG, Ye Y. Linear and nonlinear programming. New York: Springer;
interior point method for minimizing a sum of Euclidian norms. SIAM J Sci 2008.
Comput 2000;22:243–62. [57] Frederick CO, Armstrong PJ. Convergent internal stresses and steady cyclic
[34] Vu DK, Yan AM, Nguyen-Dang H. A primal-dual algorithm for shakedown states of stress. J Strain Anal 1966;1:154–69.
analysis of structures. Comput Methods Appl Mech Eng 2004;193:4663–74. [58] Polizzotto C. Variational methods for the steady state response of elastic-
[35] Magoariec H, Bourgeois S, Débordes O. Elastic plastic shakedown of 3D plastic solids subjected to cyclic loads. Int J Solids Struct 2003;40:2673–97.
periodic heterogeneous media: a direct numerical approach. Int J Plast [59] Gokhfeld DA, Cherniavsky OF. Limit analysis of structures at thermal
2004;20:1655–75. cycling. Netherlands: Sijthoff & Noordhoff; 1980.
[36] Bisbos CD, Makrodimopoulos A, Pardalos P. Second-order cone programming [60] König JA, Kleiber M. On a new method of shakedown analysis. Bull Acad Polon
approaches to static shakedown analysis in steel plasticity. Optim Meth Soft Sci Ser Sci Tech 1978;26:165–71.
2005;20:25–52. [61] Boyd S, Vandenberghe L. Convex optimization. Cambridge: Cambridge
[37] Nguyen AD, Hachemi A, Weichert D. Application of the interior point method University Press; 2004.
to shakedown analysis of pavements. Int J Numer Meth Eng 2008;4:414–39. [62] Bronshtein IN, Semendyayev KA, Musiol G, Mühlig H. Handbook of
[38] Hachemi A, Mouhtamid S, Nguyen A, Weichert D. Application of shakedown mathematics. 4th ed. Berlin: Springer-Verlag; 2004.
analysis to large-scale problems with selective algorithm. In: Weichert D, Ponter [63] Panagiotou KD. Limit state numerical procedure for cyclically loaded
A, editors. Limit states of materials and structures. Springer; 2009. p. 289–305. elastoplastic structures [Ph.D. Dissertation]. National Technical University of
[39] Tran TN, Liu GR, Nguyen-Xuan H, Nguyen-Thoi T. An edge-based smoothed Athens; 2015.
finite element method for primal-dual shakedown analysis of structures. Int J [64] Pham PT. Upper bound limit and shakedown analysis of elastic-plastic
Numer Meth Eng 2010;82:917–38. bounded linearly kinematic hardening structure [Ph.D. Dissertation]. RWTH
[40] Simon JW, Weichert D. Numerical lower bound shakedown analysis of Aachen University; 2011.
engineering structures. Comput Methods Appl Mech Eng 2011;200:2828–39. [65] Nijenhuis A, Wilf HS. Combinatorial algorithms. New York: Academic Press;
[41] Zhou S, Liu Y, Wang D, Wang K, Yu S. Upper bound shakedown analysis with 1978.
the nodal natural element method. Comput Mech 2014;54:1111–28. [66] Eurocode 1. Actions on structures. European Union EN; 1991.
[42] Le CV, Tran TD, Pham DC. Rotating plasticity and nonshakedown collapse [67] Leonetti L, Casciaro R, Garcea G. Effective treatment of complex statical and
modes for elastic–plastic bodies under cyclic loads. Int J Mech Sci dynamical load combinations within shakedown analysis of 3D frames.
2016;111:55–64. Comput Struct 2015;158:124–39.
[43] Yamaguchi T, Kanno Y. Ellipsoidal load-domain shakedown analysis with von
Mises yield criterion: a robust optimization approach. Int J Numer Meth Eng
2016;107:1136–44.