A Short Review in Model Order Reduction Based on

Proper Generalized Decomposition

Francisco Chinesta, Pierre Ladevèze, Elías Cueto

To cite this version:

Francisco Chinesta, Pierre Ladevèze, Elías Cueto. A Short Review in Model Order Reduction Based
on Proper Generalized Decomposition. Archives of Computational Methods in Engineering, Springer
Verlag, 2011, 18 (4), pp.395-404. �10.1007/s11831-011-9064-7�. �hal-01004940�

HAL Id: hal-01004940

https://hal.archives-ouvertes.fr/hal-01004940
Submitted on 27 Mar 2017

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Public Domain
A Short Review on Model Order Reduction Based on Proper
Generalized Decomposition
Francisco Chinesta · Pierre Ladeveze · Elías Cueto

Abstract This paper revisits a new model reduction method- 1 Introduction

ology based on the use of separated representations, the
so called Proper Generalized Decomposition—PGD. Space Today many problems in science and engineering remain
and time separated representations generalize Proper Or- intractable, in spite of the impressive progresses attained
thogonal Decompositions—POD—avoiding any a priori in mechanical modeling, numerical analysis, discretization
knowledge on the solution in contrast to the vast majority of techniques and computer science during the last decade,
POD based model reduction technologies as well as reduced because their numerical complexity, or the restrictions im-
bases approaches. Moreover, PGD allows to treat efficiently posed by different requirements (real-time, for instance)
models defined in degenerated domains as well as the multi- make them unaffordable for nowadays technologies.
dimensional models arising from multidimensional physics We can enumerate different challenging scenarios for ef-
(quantum chemistry, kinetic theory descriptions, . . .) or from ficient numerical simulations:
the standard ones when some sources of variability are in-
troduced in the model as extra-coordinates. – The first one concerns models that are defined in high di-
mensional spaces, usually encountered in quantum chem-
istry and kinetic theory descriptions of complex fluids.
Models defined in high dimensional spaces suffer the so-
called curse of dimensionality. If one proceeds to the solu-
This work has been partially supported by the IUF—Institut tion of a model defined in a space of dimension N by us-
Universitaire de France—, the french ANR COSINUS SIMDREAM
ing a standard mesh based discretization technique, where
project and the Spanish Ministry of Science and Innovation, through
grants number CICYT-DPI2008-918 and DPI2011-27778-C02-01. M nodes are used for discretizing each space coordinate,
the resulting number of nodes reaches the astronomical
F. Chinesta ()
value of M N . With M = 1000 (a very coarse description
EADS Foundation Chair “Advanced Computational
Manufacturing Processes”, GEM, UMR CNRS - Centrale Nantes, in practice) and N = 30 (a very simple model) the numer-
Institut Universtaire de France, 1 rue de la Noe, BP 92101, 44321 ical complexity results 1090 . It is important to recall that
Nantes cedex 3, France 1080 is the presumed number of elementary particles in
e-mail: [email protected]
the universe!
P. Ladeveze Quantum mechanics and molecular modeling of com-
EADS Foundation Chair “Advanced Computational Structural plex fluids are not the only branches of science that suffer
Mechanics”, LMT-Cachan (ENS Cachan/CNRS/UPMC/PRES from the curse of dimensionality. Consider for example
UniverSud Paris), 61 Avenue du Président Wilson, 94235 Cachan
cedex, France a chemical process involving so few molecules of the re-
e-mail: [email protected] acting species that the use of the continuous concept of
concentration is not valid. This situation is often found in
E. Cueto
genetic processes such as expression of genes. The state
Aragón Institute of Engineering Research (I3A), Universidad de
Zaragoza, Maria de Luna, 3, 50018 Zaragoza, Spain of such a discrete system is given by a probability dis-
e-mail: [email protected] tribution for the number of individual molecules of each

1
 one of the coexisting species. The balance equation gov- usually require, when approached with standard tech-
erning the evolution of the system, the so-called chemical niques, the direct computation of a very large number
master equation, is again defined in a high-dimensional of solutions of the concerned model for particular values
space that precludes a direct solution by means of stan- of the problem parameters. Consider for example the op-
dard mesh-based techniques. timization of a process where optimal parameter values
There are of course alternative methods to address must be determined for process operating conditions (e.g.
these high-dimensional problems indirectly, stochastic speed, position and temperature of heaters) and material
simulations being one of the foremost approaches. Stochas- properties (e.g. thermal and rheological properties of the
tic techniques have their own challenges, however. While materials). Clearly, it would be useful to be able to simu-
variance reduction is always an issue, it is quite difficult late this process at once for all values of these parameters
within the stochastic framework to implement paramet- within a prescribed interval, and then perform data min-
ric or sensitivity studies that go beyond the brute force ing within this rather general solution to identify optimal
approach of computing a large number of expensive, in- values.
dividual simulations. – Traditionally, Simulation-based Engineering Sciences—
– A second category of problems involves multiscale prob- SBES—relied on the use of static data inputs to perform
lems not necessarily defined in high-dimensional spaces, the simulations. These data could be parameters of the
but whose spectrum of characteristic times or lengths is so model(s) or boundary conditions. The word static is in-
wide that standard incremental discretization techniques tended to mean here that these data could not be modi-
cannot be applied. In such time-multiscale problems, for fied during the simulation. A new paradigm in the field
instance, the time step is extremely small as a conse- of Applied Sciences and Engineering has emerged in the
quence of numerical stability requirements. Thus, simula- last decade. Dynamic Data-Driven Application Systems
tions over the much larger time interval of interest, which (DDDAS) constitute nowadays one of the most challeng-
typically requires the solution of a large linear algebraic ing applications of simulation-based Engineering Sci-
system at each time step, simply become impossible. ences. By DDDAS we mean a set of techniques that allow
Multiscale models involving a wide range of characteris- the linkage of simulation tools with measurement devices
tic times abound in many fields. Reaction-diffusion mod- for real-time control of simulations. DDDAS entails the
els of the degradation of plastic materials, for example, ability to dynamically incorporate additional data into an
describe chemical reactions occurring within microsec- executing application, and in reverse, the ability of an ap-
onds coupled to diffusion of chemical substances taking plication to dynamically steer the measurement process.
place over years. In processes involving microwaves, ul- In this context, real time simulators are needed in many
trasounds, etc. or materials exhibiting different relaxation applications. One of the most challenging situations is
times, the difficulty related to time integration is crucial. that of haptic surgery simulators, where forces acting on
The same scenario is found in solid mechanics where the the surgical tool must be translated to the peripheral de-
constitutive equations are strongly non linear and cou- vice at a rate of 500 Hz. Control, malfunctioning iden-
pled, involving many scales and different characteristic tification and reconfiguration of malfunctioning systems
times. also need to run in real time. All these problems can be
– Other challenging problems are defined in degenerated seen as typical examples of DDDAS.
geometrical domains. By this we mean that at least one – Augmented reality is another area in which efficient (fast
of the characteristic dimensions of the domain is smaller and accurate) simulation is urgently needed. The idea is
by several orders of magnitude than the others. This is the supplying in real time appropriate information to the real-
case of bar, plate or shell-like domains typical of struc- ity perceived by the user. Augmented reality could be an
tures or materials processing applications. In simple sit- excellent tool in many branches of science and engineer-
uations, such problems are readily transformed into re- ing.
duced, one or two-dimensional approximate theories (e.g. – Light computing platforms are appealing alternatives to
the classical elastic plate theory). When geometrical or heavy computing platforms that in general are expen-
material non-linearities are present, however, it is usually
sive and whose use requires technical knowledge. One
impossible to derive lower-dimensional models of suffi-
can imagine that the off-line parametric solution of many
cient validity. Standard mesh-based discretization meth-
models will make possible the on-line manipulation of
ods then quickly become impractical, in view of the com-
those general solutions by using very light computing
pulsory discretization of the small length scales that yield
platforms, as for example smartphones or tablets.
extremely fine meshes.
– Many problems in process control, parametric modeling, While the previous list is by no means exhaustive, it
inverse identification, and process or shape optimization, includes a set of problems with apparent no relationship

2
among them that can be, however, treated in an unified man- ones) are still in progress, while many and mathematical
ner as will be shown in this paper. Their common ingredient foundations remain to be fully understood. See [37] for a
is our lack of capabilities (or knowledge) to solve them nu- recent review. In what follows we are describing the PGD
merically in a direct, traditional way. In order to obtain a state of the art.
solution, some kind of model order reduction is thus com-
pulsory. 2.1 Advanced Non-linear Solvers Making Use of
Many years ago, in the 80’s, Pierre Ladeveze proposed a Space-Time Separated Representations
separated representation of the space and time coordinates
Even if non-linear models can be solved by using standard

i=Q
linearization strategies in the context of separated represen-
u(x, t) ≈ Xi (x) · Ti (t) (1)
tations, similarly to [8] or [113], more robust strategies exist
i=1
for addressing efficiently history-dependent non-linearities,
for performing efficient solutions of complex non-linear making use of the natural separation of the local and non-
thermo-mechanical models. The radial approximation (1) linear thermo-mechanical behavior from the global and lin-
was one of the main blocks of the powerful non-incremental ear equilibrium in which a space-time separated representa-
and non-linear LArge Time INcrement (LATIN) solver. The tion constitutes an essential ingredient allowing spectacular
corpus of literature devoted to this technique is vast, as computing time savings.
proved in the book by Ladeveze on the topic [67], but Thus, space-time separated representations can be viewed
remained in the form of space-time separations for many as a part of the LArge Time INcrement method (LATIN)
years. which has been introduced by P. Ladeveze in 1985 [57, 58]
A more general separated representation (in this case in and was the core of two books [64, 67]. These books re-
the conformation space of complex fluids descriptions) was late the story of the PGD until their publication. Originally
more recently employed by A. Ammar and F. Chinesta [3] designed to deal with nonlinearities, such as plasticity or vis-
for approximating the solution of multi-dimensional partial coplasticity, in time dependent problems [19–21, 27, 40–42,
differential equations. A. Nouy considered also such sepa- 59–61, 68, 72], during the past three decades, many theo-
rated representations for solving stochastic equations were retical works have been done to enhance the strategy and to
the deterministic coordinates and the stochastic ones were deal with fundamental problems in computational mechan-
separated, very much like in the radial, space-time approxi- ics. The strategy has been applied in the context of large
mation. Proper Generalized Decomposition—PGD—is the displacements in [22, 64]. A domain decomposition tech-
common name recently coined for techniques using such nique has been proposed in [16, 17, 30, 62, 63] and multi-
separated representations. The nature of the problem, the in- scale features in space have been introduced in [66, 69, 73].
volved coordinates and the constructor of such approxima- These features have been extended to time in [74–77, 79, 80,
tions can be very different, as will be seen. 100] and the scalability of the strategy illustrated in [81, 95,
The general form of the separated representation involved 110]. A multiparametric strategy for taking into account the
in the PGD reads: variability parameters or to be used in identification tech-
niques has been introduced in [2, 23–25]. The strategy has

i=Q
been applied to multiphysics problems in [46, 93, 94]. Some
u(x1 , . . . , xN ) ≈ Fi1 (x1 ) · · · FiN (xN ) (2)
techniques specific to the manipulation of fields approxi-
i=1
mated by the PGD have been proposed in [65]. The problem
where xi denote a scalar or vector coordinate defined in a of the verification of model reduction methods within the
domain Ωi of moderate dimension Ωi ⊂ Rd , with d ≤ 3, in PGD was addressed in [82].
general. The strategy was also successfully applied to many stud-
ies, strongly related to industrial problems: the prediction of
damage in composites [1, 14, 70, 78, 118], the computation
2 Review on PGD Foundations and Applications of assemblies [31], the simulation of dynamic shocks [26,
71, 86], the simulation of porous media [47, 92] and the vir-
Proper Generalized Decomposition techniques allow to cir- tual testing of joints for the prediction of damping in space
cumvent most of the challenges mentioned in the introduc- launchers [28].
tion in a unified way. They also make it possible to solve
models never until now solved, or speeding up the solution 2.2 Multidimensional Models
of others up to a limit not yet reached by other techniques.
However, because the technique is still in its infancy, many As discussed in the introduction, some models are inher-
applications (possibly the most challenging and impressive ently defined in a configurational or conformation space

3
that is high-dimensional in nature. Its discretization leads to but moderate-dimensional) domains suggest the following
the well-known curse of dimensionality if traditional, mesh- separated representation
based techniques are applied. Separated representations, on
the contrary, provide with an efficient means of dealing with 
i=Q
u(x, y, z) ≈ Xi (x) · Yi (y) · Zi (z) (3)
these requirements. It is thus natural that the origin of PGD i=1
techniques, has one of its two legs in this framework, the
other being the space-time separation of the LATIN method. Thus, the 3D solution results in a sequence of 1D so-
Separated representations were applied for solving the lutions for computing functions Xi (x), Yi (y) and Zi (z).
multidimensional Fokker-Planck equation describing com- This kind of domains are very frequent when addressing
plex fluids in the kinetic theory framework. In [3], the au- homogenization problems in which the elementary volume
thors addressed the solution of the linear Fokker-Planck element is a cube. The interested reader can refer to [34].
A full decomposition was also efficiently applied for solv-
equation describing multi-bead-spring molecular models in
ing the Navier-Stokes equations in a cube in [45]. Fully sep-
the steady state and in homogeneous flows. In that case
arated representations in complex, non-hypercubic domains
the distribution function describing the molecular confor-
can be equally performed by using an appropriate X-FEM
mation only depends on the molecule configurational coor-
like strategy [54]. In that work the enforcement of non-
dinates. The solution procedure was extended to non-linear homogeneous Dirichlet boundary conditions was also anal-
kinetic theory descriptions of more complex molecular de- ysed, giving rise to a generalized form of the PGD approach.
scriptions in [90]. The transient solution was addressed in In the case of plates, shells or extruded geometries, one
[5] in which the time was added as an extra-coordinate. could consider the separated approximation [18] also advan-
Transient solutions making use of reduced bases in the con- tageously:
text of an adaptive proper orthogonal decomposition [116]
were considered in the case of low dimensional configura- 
i=Q

tion spaces: the FENE model was addressed in [4] and liquid u(x, y, z) ≈ Xi (x, y) · Zi (z) (4)
i=1
crystalline polymers in [6]. In [87], the authors considered
multi-bead-spring models but considered spectral approxi- In the case of models defined in plate geometries (x, y)
mation for representing all the functions involved in the fi- represent the in-plane coordinates and z the thickness direc-
nite sums decomposition. A deeper analysis of non-linear tion. In the case of extruded profiles (x, y) represents the
and transient models was considered in [8]. Complex fluid surface extruded in the z direction. This kind of representa-
models were coupled with complex flows in [112] and [91] tion makes possible fully 3D solutions with numerical com-
opening very encouraging perspectives and claiming the ne- plexities (and therefore computation times) characteristic of
cessity of defining efficient stabilizations. A first tentative 2D solutions, without any simplifying kinematic assump-
of convective stabilization was proposed in [43]. Finally, in tion.
[33] the PGD was applied for solving the stochastic equa-
tion within the Brownian Configuration Field framework. 2.4 Parametric Models: A Route to Efficient Optimization,
Inverse Identification and Real Time Simulation
The interested reader can refer to [39] and the references
therein for an exhaustive overview of PGD in computational
Previous multidimensional models may seem too far from
rheology.
the usual computational mechanics practice. However, usual
Multidimensional models encountered in the finer de-
computational mechanics models could be enriched by a
scriptions of matter (ranging from quantum chemistry to sta- PGD treatment. Thus, adding some new coordinates to mod-
tistical mechanics descriptions) were revisited in [35]. The els initially non high-dimensional, could lead to new, never
multidimensional chemical master equation was efficiently explored insights in the physics. Imagine for instance solv-
solved in [38]. The Langer’s equation governing phase tran- ing the heat (Fourier) equation and assume unknown the ma-
sitions was considered in [85]. Finally, models coming from terial thermal conductivity. This could happen because it has
financial mathematics were addressed in [48]. a stochastic nature, for instance, or simply because prior to
solve the thermal model it should be measured. You have
2.3 Separating the Physical Space three possibilities: (i) wait for the conductivity to be mea-
sured (a conservative solution, but impractical in many en-
gineering situations); (ii) solve the equation for many values
Other models are not inherently defined in a high-dimen- of the conductivity (a sort of brute force approach) to get an
sional space, but can nevertheless be treated efficiently in a overall idea of the behavior of the solution; or (iii) solve the
separated manner. Models defined in cubic (or hypercubic, heat equation only once for any value of the conductivity.

4
 Obviously the third alternative is the most exciting one, in any of the geometries generated by the parameters con-
but up to our knowledge, no result has been presented in sidered as extra-coordinates [88]. This strategy could be an
the literature so far that could eventually allow to do so. To alternative to the POD-based shape optimization considered
compute this general solution, however, it suffices to intro- in [117]. Inverse methods in the context of real-time simu-
duce the conductivity k as an extra-coordinate, playing the lations were addressed in [55] and were coupled with con-
same role as the standard space and time coordinates, even if trol strategies in [53] as a first step towards DDDAS (dy-
there are not derivatives concerning this extra-coordinate in namic data-driven application systems). Moreover, because
the heat equation. A possible separated approximation could the general parametric solution was precomputed off-line,
be: it can be used on-line under real time constraints and using
light computing platforms like smartphones [18, 53], that

i=Q
constitutes a first step towards the use of this kind of infor-
u(x, t, k) ≈ Xi (x) · Ti (i) · Ki (k) (5)
mation in augmented reality platforms.
i=1
As mentioned before, surgical simulators must operate at
This procedure works very well in practice, and can be frequencies higher than 500 Hz. The use of model reduc-
extended to introduce many other extra-coordinates: source tion seems to be an appealing alternative for reaching such
term, boundary conditions, initial condition, etc. The price performances. However, techniques based on the use of the
to pay is the increase of the dimensionality of the resulting POD, PODI, even combined with an asymptotic numerical
model that now contains the standard physical coordinates methods to avoid the computation of the tangent matrix, ex-
(space and time) plus all the other extra-coordinates that we hibit serious difficulties to fulfil such requirements as dis-
decided to introduce. However, the separated representation cussed in [96–99]. The use of parametric solutions in which
of the PGD allows to treat these models advantageously. the applied load and its point of application are considered
This kind of parametric modeling was addressed in [11, as extra-coordinates open an unimaginable field of applica-
18, 114] where material parameters were introduced as tions.
extra-coordinates. In [83] and [84] thermal conductivities,
macroscopic temperature and its time evolution were intro- 2.6 Uncertainty Quantification and Stochastic Parametric
duced as extra-coordinates for computing linear and non- Analyses
linear homogenization. In [18] the anisotropy direction of
plies involved in a composite laminate were considered as The PGD method was introduced in [101] in the context of
extra-coordinates. By assuming a certain uncertainty in the uncertainty quantification. In this context, the method was
real orientation of such plies, authors evaluate the envelope initially called Generalized Spectral Decomposition (GSD)
of the resulting distorted structures due to the thermome- and was considered to be a generalization of Karhunen-
chanical coupling. Loève decomposition for the a priori construction of sep-
The futurist concept of “virtual chart for dimensioning” arated representations of the solution of stochastic partial
consists of a PGD-reduced model describing the physics for differential equations (SPDE).
a family of structures. It allows engineers to proceed to com- The problem that defines the reduced bases of functions
plex designs in some seconds. This concept, fully compati- in the decomposition of the solution has been interpreted as
ble with the most recent design procedures, is being actively a pseudo eigenproblem. This interpretation has led to the
developed by the Ladeveze’s group. development of dedicated algorithms inspired from solution
techniques for classical eigenproblems [103]. In this context
2.5 Real-Time Simulation, DDDAS and More of stochastic problems, a major interest of PGD algorithms
is that they decouple the solution of deterministic problems
It is easy to understand that after performing this type of and algebraic stochastic equations, thus making the PGD a
calculations, in which parameters are considered advanta- promising alternative to traditional methods for uncertainty
geously as new coordinates of the model, a posteriori in- propagation [105]. PGD has also been extended to stochastic
verse identification or optimization can be easily handled. nonlinear problems with two alternative strategies: the use of
This new PGD framework allows to perform this type of nonlinear iterative solvers with the PGD method as a solver
calculations very efficiently, because in fact all the result for the succession of linearized stochastic problems [102],
have been previously computed in the form of a separated, or the direct construction of a separated representation of
high-dimensional solution so that they constitute a simple the solution of nonlinear problems [104]. More recently, the
post-processing of this general solution. Process optimiza- PGD has been successfully applied to the solution of high
tion was considered in [52], for instance. Shape optimiza- dimensional stochastic parametric problems, with the intro-
tion was performed by considering all the geometrical pa- duction of suitable hierarchical tensor representations and
rameters as extra-coordinates, leading to the model solution associated algorithms [106].

5
 PGD has also been exploited for the construction of ro- in reflexive tensor Banach spaces. The above results have led
bust reduced order models in structural solid mechanics to a generalization of the concept of Singular Value Decom-
[32], thus providing an alternative to traditional modal su- position (SVD). In particular, a constrained version of the
perposition approaches. In [109], PGD method was coupled SVD has been proposed in [108].
with fictitious domain methods in order to handle the solu-
tion of PDEs with geometrical uncertainties. 2.8 Miscellaneous

2.7 Numerical Analysis There is a variety of open questions when applying the PGD
based discretization techniques. We discussed previously the
From the point of view of the numerical analysis of sepa- question related to the convective stabilization [43]. The is-
rated representation and its associated constructors we can sue related to the stabilization of mixed formulation has not
cite some recent works. There exist several approaches to been addressed until now.
the numerical analysis of PGD. They combine the existence Other important question lies in the error estimation.
of a best approximation and a greedy algorithm. The idea A first attempt was considered in [10]. Coupling standard
of using greedy algorithms to construct successive approx- mesh based discretization techniques with reduced bases
imations was considered in [3] in the context of the nu- (POD and PGD) is an issue of major interest in order to
merical solution of multidimensional PDEs. One can note, represent localized behaviors (e.g. discontinuities, singular-
however, that the strategy considered in [3] can be inter- ities, boundary layers, . . .). Some coupling strategies were
preted as a greedy algorithm for the error minimization only proposed in [7] and [12].
for elliptic symmetric problems since they can be rewrit- Multi-scale and multi-physics non-linear coupled mod-
ten in terms of minimization of an energy functional. In [9] els involving different characteristic times were efficiently
the analysis of the convergence of the greedy rank-one up- coupled in [36] where a globalization of local problems to
date algorithm for solving full-rank linear systems was ad- speed-up the simulation in a separated representation frame-
dressed. work has been proposed. Later, in [13] the time axis was
A more complete theoretical study was addressed in transformed in a two-dimensional domain to account for
[89] for symmetric elliptic problems by exploiting its con- two very different time scales. The coupling of other multi-
nection to greedy algorithms from nonlinear approxima- physic models in the context of composites manufacturing
tion theory explored, for example, in [44]. Hence, the processes was performed in [113]. A monolithic approach of
variational version of the algorithm, based on the mini- coupled thermo-mechanical models was employed in [15].
mization of a sequence of Dirichlet energies, was shown In [111] the solution of the electromagnetic equations was
to converge. In [51] the authors extend the convergence considered.
analysis of the pure greedy and orthogonal greedy algo- In [56] the PGD was introduced in the boundary ele-
rithms considered in [89] to the technically more com- ment method framework for solving transient models where
plicated case where the Laplace operator is replaced by few works concerning the use of reduced bases exist [115].
a high-dimensional Ornstein-Uhlenbeck operator with un- Because the use of a space-time separated representation
bounded drift. This case bears many similarities with that the time-dependent kernel is no more needed, and only
appearing in Fokker-Planck equations arising in bead-spring the steady state kernel functions are needed. The space-
chain type kinetic polymer models with finitely exten- time separation implies significant computational time sav-
sible non-linear elastic potentials. These are frequently ings.
posed on a high-dimensional Cartesian product configura- In [107], alternative definitions of PGD and dedicated
tion space D = D1 × · · · × DN contained in RN d , where algorithms are provided in the context of time-dependent
each set Di , i = 1, . . . , N , is a bounded open ball in Rd , PDEs. In particular, a new definition of PGD called Mini-
d = 2, 3. max PGD is introduced for handling non symmetric varia-
General elliptic problems were considered in [49]. The tional problems.
extension of these results to non-linear symmetric coercive
problems is the subject of the paper [29] that also con-
tains specific results when the algorithm is applied to finite- 3 Discussion
dimensional problems. Current research tracks include for
example the extension of the convergence analysis to non- As can be noticed from the pages before, the Proper Gen-
symmetric problems and the analysis of the rate of conver- eralized Decomposition technique has been advantageously
gence for non-linear equations. In [50], it is provided a math- applied to a plethora of different problems and in different
ematical analysis of a family of progressive and updated ways. Ranging from the original space-time separated rep-
PGDs for a particular class of convex optimization problems resentations to more recent applications in which models are

6
efficiently cast into a high-dimensional framework to over- 14. Aubard X, Cluzel C, Guitard L, Ladevèze P (2000) Damage
come real-time requirements by just making real-time post- modeling at two scales for 4D carbon/carbon composites. Com-
put Struct 78(1–3):83–91
processing, the potential field of applications seems not to 15. Beringhier M, Gueguen M, Grandidier JC (2010) Solution of
have been completely explored. strongly coupled multiphysics problems using space-time sepa-
In fact, some crucial aspects at the very root of the rated representations: application to thermoviscoelasticity. Arch
method continue to be not fully understood. For instance, Comput Methods Eng 17(4):393–401
16. Blanzé C, Danwe R, Ladevèze P, Moreau J-P (1993) Une méth-
the relationship among POD/SVD and PGD techniques, has ode pour l’étude d’assemblage de structures massives. In: Col-
not been analysed to the whole extent. Optimality conditions loque National en Calcul des Structures, Hermès, pp 913–919
of the separated representation, for instance, are intimately 17. Blanzé C, Champaney L, Cognard J-Y, Ladevèze P (1996)
related to this question and have not been completely solved A modular approach to structure assembly computations—
application to contact problems. Eng Comput 13(1):15
so far. 18. Bognet B, Leygue A, Chinesta F, Poitou A, Bordeu F (2011)
We believe, therefore, that there is a lot of room in the Advanced simulation of models defined in plate geometries: 3D
PGD field to perform research, both at the theoretical level solutions with 2D computational complexity. Comput Methods
and from the point of view of applications, whose range Appl Mech Eng. doi:10.1016/j.cma.2011.08.025
19. Boisse P, Ladevèze P, Rougée P (1989) A large time incre-
seems to be unlimited. ment method for elastoplastic problems. Eur J Mech A, Solids
8(4):257–275
20. Boisse P, Bussy P, Ladevèze P (1990) A new approach in non-
References linear mechanics—the large time increment method. Int J Numer
Methods Eng 29(3):647–663
21. Boisse P, Ladevèze P, Poss M, Rougée P (1991) A new large time
1. Allix O, Ladevèze P, Gilleta D, Ohayon R (1989) A damage pre- increment algorithm for anisotropic plasticity. Int J Plast 7(1–
diction method for composite structures. Int J Numer Methods 2):65–77
Eng 27(2):271–283 22. Boucard PA, Ladevèze P, Poss M, Rougée P (1997) A non-
2. Allix O, Vidal P (2002) A new multi-solution approach suitable incremental approach for large displacement problems. Comput
for structural identification problems. Comput Methods Appl Struct 64:499–508
Mech Eng 191:2727–2758 23. Boucard PA, Ladevèze P (1999) A multiple solution method for
3. Ammar A, Mokdad B, Chinesta F, Keunings R (2006) A new non-linear structural mechanics. Mech Eng 50(5):317–328
family of solvers for some classes of multidimensional partial 24. Boucard PA, Ladevèze P (1999) Une application de la méthode
differential equations encountered in kinetic theory modeling of latin au calcul multirésolution de structures non linéaires. In: Re-
complex fluids. J Non-Newton Fluid Mech 139:153–176 vue Européenne des Eléments Finis, pp 903–920
4. Ammar A, Ryckelynck D, Chinesta F, Keunings R (2006) On the 25. Boucard PA (2001) Application of the LATIN method to the cal-
reduction of kinetic theory models related to finitely extensible culation of response surfaces. In: 1st MIT conference on compu-
dumbbells. J Non-Newton Fluid Mech 134:136–147 tational fluid and solid mechanics, vol 1, pp 78–81
5. Ammar A, Mokdad B, Chinesta F, Keunings R (2007) A new 26. Boucard PA, Derumaux M, Ladevèze P (2003) Macro-meso
family of solvers for some classes of multidimensional partial models for joints submitted to pyrotechnic shock. In: Compu-
differential equations encountered in kinetic theory modeling of tational fluid and solid mechanics, vol 1–2, pp 139–142.
complex fluids. Part II: Transient simulation using space-time 27. Bussy P, Rougée P, Vauchez P (1990) The large time increment
separated representation. J Non-Newton Fluid Mech 144:98–121 method for numerical simulation of metal forming processes. In:
6. Ammar A, Pruliere E, Chinesta F, Laso M (2009) Reduced NUMETA. Elsevier, Amsterdam, pp 102–109
numerical modeling of flows involving liquid-crystalline poly- 28. Caignot A, Ladevèze P, Néron D, Durand JF (2010) Virtual test-
meres. J Non-Newton Fluid Mech 160:140–156 ing for the prediction of damping in joints. Eng Comput 27(5–
7. Ammar A, Pruliere E, Ferec J, Chinesta F, Cueto E (2009) Cou- 6):621–644
pling finite elements and reduced approximation bases. Eur J 29. Cancès E, Ehrlacher V, Lelièvre T (2011) Convergence
Comput Mech 18(5–6):445–463 of a greedy algorithm for high-dimensional convex nonlin-
8. Ammar A, Normandin M, Daim F, Gonzalez D, Cueto E, ear problems. Math Models Methods Appl Sci. doi:10.1142/
Chinesta F (2010) Non-incremental strategies based on separated S0218202511005799
representations: applications in computational rheology. Com- 30. Champaney L, Cognard J-Y, Dureisseix D, Ladevèze P (1997)
mun Math Sci 8(3):671–695 Large scale applications on parallel computers of a mixed do-
9. Ammar A, Chinesta F, Falco A (2010) On the convergence of a main decomposition method. Comput Mech 19(4):253–263
greedy rank-one update algorithm for a class of linear systems. 31. Champaney L, Cognard J-Y, Ladevèze P (1999) Modular analy-
Arch Comput Methods Eng 17(4):473–486 sis of assemblages of three-dimensional structures with unilateral
10. Ammar A, Chinesta F, Diez P, Huerta A (2010) An error esti- contact conditions. Comput Struct 73(1–5):249–266
mator for separated representations of highly multidimensional 32. Chevreuil M, Nouy A (2011) Model order reduction based on
models. Comput Methods Appl Mech Eng 199:1872–1880 proper generalized decomposition for the propagation of uncer-
11. Ammar A, Normandin M, Chinesta F (2010) Solving parametric tainties in structural dynamics. Int J Numer Methods Eng. doi:10.
complex fluids models in rheometric flows. J Non-Newton Fluid 1002/nme.3249
Mech 165:1588–1601 33. Chinesta F, Ammar A, Falco A, Laso M (2007) On the reduc-
12. Ammar A, Chinesta F, Cueto E (2011) Coupling finite elements tion of stochastic kinetic theory models of complex fluids. Model
and proper generalized decompositions. Int J Multiscale Comput Simul Mater Sci Eng 15:639–652
Eng 9(1):17–33 34. Chinesta F, Ammar A, Lemarchand F, Beauchene P, Boust F
13. Ammar A, Chinesta F, Cueto E, Doblare M (2011) Proper gen- (2008) Alleviating mesh constraints: model reduction, parallel
eralized decomposition of time-multiscale models. Int J Numer time integration and high resolution homogenization. Comput
Methods Eng. doi:10.1002/nme.3331 Methods Appl Mech Eng 197(5):400–413

7
35. Ammar A, Chinesta F, Joyot P (2008) The nanometric and mi- 56. Bonithon G, Joyot P, Chinesta F, Villon P (2011) Non-
crometric scales of the structure and mechanics of materials re- incremental boundary element discretization of parabolic mod-
visited: an introduction to the challenges of fully deterministic els based on the use of proper generalized decompositions. Eng
numerical descriptions. Int J Multiscale Comput Eng 6(3):191– Anal Bound Elem 35(1):2–17
213 57. Ladevèze P (1985) New algorithms: mechanical framework and
36. Chinesta F, Ammar A, Cueto E (2010) Proper generalized de- development (in french). Technical Report 57, LMT-Cachan
composition of multiscale models. Int J Numer Methods Eng
58. Ladevèze P (1985) On a family of algorithms for structural me-
83(8–9):1114–1132
chanics. CR Acad Sci Paris 300(2):41–44 (in french)
37. Chinesta F, Ammar A, Cueto E (2010) Recent advances and
new challenges in the use of the Proper Generalized Decompo- 59. Ladevèze P, Rougée P (1985) Viscoplasticity under cyclic
sition for solving multidimensional models. Arch Comput Meth- loadings: properties of the homogenized cycle. CR Acad Sci
ods Eng 17(4):327–350 301:891–894
38. Chinesta F, Ammar A, Cueto E (2010) On the use of proper gen- 60. Ladevèze P (1989) The large time increment method for the an-
eralized decompositions for solving the multidimensional chem- alyze of structures with nonlinear constitutive relation described
ical master equation. Eur J Comput Mech 19:53–64 by internal variables. CR Acad Sci Paris, 309:1095–1099
39. Chinesta F, Ammar A, Leygue A, Keunings R (2011) An 61. Ladevèze P (1991) New advances in the large time increment
overview of the proper generalized decomposition with appli- method. In: Ladevèze P, Zienkiewicz OC (eds) New advances in
cations in computational rheology. J Non-Newton Fluid Mech computational structural mechanics. Elsevier, Amsterdam, pp 3–
166:578–592 21
40. Cognard J-Y (1990) Le traitement des problèmes nonlinéaires à 62. Ladevèze P, Lorong Ph (1992) A large time increment approach
grand nombre de degrés de liberté par la méthode à grand incré-
with domain decomposition technique for mechanical non linear
ment de temps. In: Fouet J-M et al (eds) Calcul des structures et
problems. In: Computing methods in applied sciences and engi-
intelligence artificielle, Pluralis, pp 211–222
neering INRIA, pp. 569–578
41. Cognard J-Y, Ladevèze P (1993) A large time increment ap-
proach for cyclic viscoplasticity. Int J Plast 9:141–157 63. Ladevèze P, Lorong Ph (1993) Formulation et stratégies “par-
42. Cognard J-Y, Ladevèze P, Talbot P (1999) A large time incre- allèles” pour l’analyse non linéaire des structures. In: Colloque
ment approach for thermo-mechanical problems. Adv Eng Softw national en calcul des structures. Hermès, Paris, pp 910–919
30(9–11):583–593 64. Ladevèze P (1996) Mécanique non linéaire des structures. Her-
43. Gonzalez D, Cueto E, Chinesta F, Debeugny L, Diez P, Huerta A mès, Paris
(2010) Int J Mater Form 3(1):883–886 65. Ladevèze P (1997) A computational technique for the inte-
44. DeVore RA, Temlyakov VN (1996) Some remarks on greedy al- grals over the time-space domain in connection with the LATIN
gorithms. Adv Comput Math 5:173–187 method (in french). Technical Report 193, LMT-Cachan
45. Dumon A, Allery C, Ammar A (2011) Proper general decom-
66. Ladevèze P, Dureisseix D (1998) A 2-level and mixed domain
position (PGD) for the resolution of Navier-Stokes equations.
decomposition approach for structural analysis. Contemp Math
J Comput Phys 230(4):1387–1407
218:246–253
46. Dureisseix D, Ladevèze P, Néron D, Schrefler BA (2003)
A multi-time-scale strategy for multiphysics problems: applica- 67. Ladevèze P (1999) Nonlinear computationnal structural
tion to poroelasticity. Int J Multiscale Comput Eng 1(4):387–400 mechanics—new approaches and non-incremental methods of
47. Dureisseix D, Ladevèze P, Schrefler BA (2003) A latin computa- calculation. Springer, Berlin
tional strategy for multiphysics problems: application to poroe- 68. Ladevèze P, Cognard J-Y, Talbot P (1999) A non-incremental
lasticity. Int J Numer Methods Eng 56(10):1489–1510 and adaptive computational approach in thermo-viscoplasticity.
48. Falco A (2010) Algorithms and numerical methods for high di- In: Bruhns OT, Stein E (eds) IUTAM symposium on micro- and
mensional financial market models. Rev Econ Financ, 20:51–68 macrostructural aspects of the thermoplasticity, pp 281–291
49. Falcó A, Nouy A (2011) A proper generalized decomposition 69. Ladevèze P, Dureisseix D (1999) A new micro-macro computa-
for the solution of elliptic problems in abstract form by using a tional strategy for structural analysis. CR Acad Sci, Ser Ii, Fasci-
functional Eckart-Young approach. J Math Anal Appl 376:469– cule, B—Mec Phys Astron, 327(12):1237–1244
480
70. Ladevèze P, Guitard L, Champaney L, Aubard X (2000) Debond
50. Falco A, Nouy A Proper generalized decomposition for nonlinear
modeling for multidirectional composites. Comput Methods
convex problems in tensor Banach spaces. arXiv:1106.4424v1
51. Figueroa L, Süli E (2011) Greedy approximation of high- Appl Mech Eng 185(2–4):109–122
dimensional Ornstein-Uhlenbeck operators with unbounded 71. Ladevèze P, Lemoussu H, Boucard PA (2000) A modular ap-
drift. arXiv:1103.0726 proach to 3-d impact computation with frictional contact. Com-
52. Ghnatios Ch, Chinesta F, Cueto E, Leygue A, Breitkopf P, Vil- put Struct 78(1–3):45–51
lon P (2011) Methodological approach to efficient modeling and 72. Ladevèze P, Perego U (2000) Duality preserving discretization of
optimization of thermal processes taking place in a die: applica- the large time increment methods. Comput Methods Appl Mech
tion to pultrusion. Composites, Part A, Appl Sci Manuf 42:1169– Eng 189(1):205–232
1178 73. Ladevèze P, Loiseau O, Dureisseix D (2001) A micro-macro
53. Ghnatios Ch, Masson F, Huerta A, Cueto E, Leygue A, Chinesta and parallel computational strategy for high heterogeneous struc-
F (2011) Proper generalized decomposition based dynamic data- tures. Int J Numer Methods Eng, 52(1–2):121–138
driven control of thermal processes. Comput Methods Appl 74. Ladevèze P, Nouy A (2002) A multiscale computational method
Mech Eng. Submitted
with time and space homogenization. CR Mec, 330(10):683–689
54. Gonzalez D, Ammar A, Chinesta F, Cueto E (2010) Recent
advances on the use of separated representations. Int J Numer 75. Ladevèze P, Nouy A (2002) Une stratégie de calcul multiéchelle
Methods Eng 81(5):637–659 avec homogénéisation en espace et en temps. CR Mec, 330:683–
55. Gonzalez D, Masson F, Poulhaon F, Leygue A, Cueto E, Chinesta 689
F (2011) Proper generalized decomposition based dynamic data- 76. Ladevèze P, Nouy A, Loiseau O (2002) A multiscale compu-
driven inverse identification. Mathematics and Computers in tational approach for contact problems. Comput Methods Appl
Simulation, Submitted, 2011 Mech Eng, 191(43):4869–4891

8
77. Ladevèze P, Nouy A (2003) On a multiscale computational strat- 97. Niroomandi S, Alfaro I, Cueto E, Chinesta F (2010) Model order
egy with time and space homogenization for structural mechan- reduction for hyperelastic materials. Int J Numer Methods Eng
ics. Comput Methods Appl Mech Eng, 192(28–30):3061–3087 81(9):1180–1206
78. Ladevèze P (2004) Multiscale modeling and computational 98. Niroomandi S, Alfaro I, Cueto E, Chinesta F (2011) Account-
strategies for composites. Int J Numer Methods Eng, 60(1):233– ing for large deformations in real-time simulations of soft tis-
253 sues based on reduced order models. Comput Methods Program
79. Ladevèze P, Néron D, Gosselet P (2007) On a mixed and mul- Biomed. doi:10.1016/j.cmpb.2010.06.012
tiscale domain decomposition method. Comput Methods Appl 99. Niroomandi S, Alfaro I, Gonzalez D, Cueto E, Chinesta F (2011)
Mech Eng 96:1526–1540 Real time simulation of surgery by reduced order modeling and
80. Ladevèze P, Néron D, Passieux J-C (2009) On multiscale com- X-FEM techniques. Int J Numer Methods Biomed Eng In press
putational mechanics with time-space homogenization. In: Fish J 100. Nouy A, Ladevèze P (2004) Multiscale computational strategy
(ed) Multiscale methods—bridging the scales in science and en- with time and space homogenization: a radial type approximation
gineering. Oxford University Press, Oxford, pp 247–282. chapter technique for solving micro problems. Int J Multiscale Comput
Space Time Scale Bridging methods Eng 170(2):557–574
81. Ladevèze P, Passieux J-C, Néron D (2010) The latin multiscale 101. Nouy A (2007) A generalized spectral decomposition technique
computational method and the proper generalized decomposi- to solve a class of linear stochastic partial differential equations.
tion. Comput Methods Appl Mech Eng, 199(21–22):1287–1296 Comput Methods Appl Mech Eng 196:4521–4537
82. Ladevèze P, Chamoin L (2011) On the verification of model re- 102. Nouy A (2007) Méthode de construction de bases spectrales
duction methods based on the proper generalized decomposition. généralisées pour l’approximation de problèmes stochastiques.
Comput Methods Appl Mech Eng 200:2032–2047 Mec Ind 8(3):283–288
83. Lamari H, Chinesta F, Ammar A, Cueto E (2009) Non- 103. Nouy A (2008) Generalized spectral decomposition method for
conventional numerical strategies in the advanced simulation of solving stochastic finite element equations: invariant subspace
materials and processes. Int J Mod Manuf Technol, 1:49–56 problem and dedicated algorithms. Comput Methods Appl Mech
84. Lamari H, Ammar A, Cartraud P, Legrain G, Jacquemin F, Eng 197:4718–4736
Chinesta F (2010) Routes for efficient computational homoge- 104. Nouy A, Le Maître O (2009) Generalized spectral decomposi-
nization of non-linear materials using the proper generalized de- tion method for stochastic non linear problems. J Comput Phys,
composition. Arch Comput Methods Eng, 17(4):373–391 228(1):202–235
85. Lamari H, Ammar A, Leygue A, Chinesta F On the solution of 105. Nouy A (2009) Recent developments in spectral stochastic meth-
the multidimensional Langerõs equation by using the proper gen- ods for the numerical solution of stochastic partial differential
eralized decomposition method for modeling phase transitions.
equations. Arch Comput Methods Eng, 16(3):251–285
Model Simul Mater Sci Eng. Submitted
106. Nouy A (2010) Proper generalized decompositions and separated
86. Lemoussu H, Boucard P-A, Ladevèze P (2002) A 3d shock com-
representations for the numerical solution of high dimensional
putational strategy for real assembly and shock attenuator. Adv
stochastic problems. Arch Comput Methods Eng, 17:403–434
Eng Softw 33(7–10):517–526
107. Nouy A (2010) A priori model reduction through proper general-
87. Leonenko GM, Phillips TN (2009) On the solution of the Fokker-
ized decomposition for solving time-dependent partial differen-
Planck equation using a high-order reduced basis approximation.
tial equations. Comput Methods Appl Mech Eng 199:1603–1626
Comput Methods Appl Mech Eng 199(1–4):158–168
108. Nouy A, Falco A Constrained tensor product approximations
88. Leygue A, Verron E (2010) A first step towards the use of proper
general decomposition method for structural optimization. Arch based on penalized best approximations. Linear Algebra Appl,
Comput Methods Eng 17(4):I465–472 oai:hal.archives-ouvertes.fr:hal-00577942
89. Le Bris C, Lelièvre T, Maday Y (2009) Results and ques- 109. Nouy A, Chevreuil M, Safatly E (2011) Fictitious domain
tions on a nonlinear approximation approach for solving method and separated representations for the solution of bound-
high-dimensional partial differential equations. Constr Approx ary value problems on uncertain parameterized domains. Comput
30:621–651 Methods Appl Mech Eng. doi:10.1016/j.cma.2011.07.002
90. Mokdad B, Pruliere E, Ammar A, Chinesta F (2007) On the 110. Passieux J-C, Ladevèze P, Néron D (2010) A scalable time-space
simulation of kinetic theory models of complex fluids using the multiscale domain decomposition method: adaptive time scale
Fokker-Planck approach. Appl Rheol, 17(2):26494, 1–4 separation. Comput Mech 46(4):621–633
91. Mokdad B, Ammar A, Normandin M, Chinesta F, Clermont JR 111. Pineda M, Chinesta F, Roger J, Riera M, Perez J, Daim F (2010)
(2010) A fully deterministic micro-macro simulation of complex Simulation of skin effect via separated representations. Int J
flows involving reversible network fluid models. Math Comput Comput Math Electr Electron Eng, 29(4):919–929
Simul 80:1936–1961 112. Pruliere E, Ammar A, El Kissi N, Chinesta F (2009) Recirculat-
92. Néron D, Ladevèze P, Dureisseix D, Schrefler BA (2004) Ac- ing flows involving short fiber suspensions: numerical difficulties
counting for nonlinear aspects in multiphysics problems: appli- and efficient advanced micro-macro solvers. Arch Comput Meth-
cation to poroelasticity. In: Lecture notes in computer science, ods Eng, 16:1–30
vol 3039, pp 612–620 113. Pruliere E, Ferec J, Chinesta F, Ammar A (2010) An efficient re-
93. Néron D, Dureisseix D (2008) A computational strategy for duced simulation of residual stresses in composites forming pro-
poroelastic problems with a time interface between coupled cesses. Int J Mater Form, 3(2):1339–1350
physics. Int J Numer Methods Eng 73(6):783–804 114. Pruliere E, Chinesta F, Ammar A (2010) On the deterministic
94. Néron D, Dureisseix D (2008) A computational strategy for solution of multidimensional parametric models by using the
thermo-poroelastic structures with a time-space interface cou- proper generalized decomposition. Math Comput Simul 81:791–
pling. Int J Numer Methods Eng 75(9):1053–1084 810
95. Néron D, Ladevèze P (2010) Proper generalized decomposition 115. Ryckelynck D, Hermanns L, Chinesta F, Alarcon E (2005) An
for multiscale and multiphysics problems. Arch Comput Meth- efficient a priori model reduction for boundary element models.
ods Eng 17(4):351–372 Eng Anal Bound Elem 29:796–801
96. Niroomandi S, Alfaro I, Cueto E, Chinesta F (2008) Real-time 116. Ryckelynck D, Chinesta F, Cueto E, Ammar A (2006) On the a
deformable models of non-linear tissues by model reduction priori model reduction: overview and recent developments. Arch
techniques. Comput Methods Programs Biomed 91:223–231 Comput Methods Eng, 13(1):91–128

9
117. Schmidt F, Pirc N, Mongeau M, Chinesta F (2011) Efficient 118. Violeau D, Ladevèze P, Lubineau G (2009) Micromodel-based
mould cooling optimization by using model reduction. Int J simulations for laminated composites. Compos Sci Technol,
Mater Form, 4(1):71–82 69(9):1364–1371

10