A Component-Based Hybrid Reduced Basis/finite Element Method For Solid Mechanics With Local Nonlinearities

Available online at www.sciencedirect.
com
ScienceDirect
Comput. Methods Appl. Mech. Engrg. 329 (2018) 498–531

www.elsevier.com/locate/cma
A component-based hybrid reduced basis/finite element method for

solid mechanics with local nonlinearities✩
J. Ballani c , D.B.P. Huynh a , D.J. Knezevic a , ∗, L. Nguyen a , A.T. Patera b
a Akselos S.A., EPFL Innovation Park, Building D, 1015 Lausanne, Switzerland
b Department of Mechanical Engineering, MIT, Cambridge MA, 02139, USA
c Swiss Federal Institute of Technology in Lausanne, Station 8, 1015 Lausanne, Switzerland
Received 10 April 2017; received in revised form 10 August 2017; accepted 12 September 2017
Abstract
The SCRBE (Static-Condensation Reduced-Basis-Element) method is a component-to-system model order reduction approach
for efficient many-query and real-time treatment of linear partial differential equations characterized by many spatially distributed
constitutive, geometry, and topology parameters. In this paper we incorporate the SCRBE approach into a framework for analysis
of problems in solid mechanics which are largely linear with the exception of local nonlinearities. In particular, we exploit a
linear–nonlinear domain decomposition to develop a hybrid formulation: we consider a SCRBE approximation over the (assumed
predominant) part of the domain associated to a linear elasticity model; we revert to a full finite element (FE) approximation over
the part of the domain associated to the locally nonlinear model. We adapt the SCRBE port training procedures to anticipate the
behavior of the field over the linear–nonlinear interface and hence ensure an accurate solution over the entire domain. We choose
a globally conforming approximation which exploits the intrinsic port structure of the SCRBE method and yields a decoupled
“non-invasive” formulation of the respective linear and nonlinear blocks of the residual vector and (Newton) Jacobian matrix.
We present numerical results for several local nonlinearities – elastic contact with and without friction, plastic (contact) – which
demonstrate substantial savings relative to standard FE approximation over the entire domain, in particular in the regime in which
the linear part of the domain represents the majority of the (FE) degrees of freedom.
⃝c 2017 Elsevier B.V. All rights reserved.
Keywords: Partial differential equations; Reduced basis method; Port training; Components; Finite element method; Contact analysis; Plasticity
1. Introduction
Parametric model order reduction is an effective approach for the treatment of parametrized partial differential
equations (PDEs) in the real-time and many-query contexts. Several approaches to parametrized model order reduction
✩ Submitted to CMAME.
∗ Corresponding author.
E-mail addresses: jonas.ballani@epfl.ch (J. Ballani), phuong.huynh@akselos.com (D.B.P. Huynh), david.knezevic@akselos.com
(D.J. Knezevic), loi.nguyen@akselos.com (L. Nguyen), patera@mit.edu (A.T. Patera).
https://doi.org/10.1016/j.cma.2017.09.014
0045-7825/⃝ c 2017 Elsevier B.V. All rights reserved.
J. Ballani et al. / Comput. Methods Appl. Mech. Engrg. 329 (2018) 498–531 499
are in widespread use, including reduced basis (RB) methods [1–4], the Proper Orthogonal Decomposition (POD) [5],
and the Proper Generalized Decomposition (PGD) [6]. (We restrict attention here to time-independent problems; for
a more complete review of parametric model order reduction, with emphasis on dynamical systems, we refer to [7].)
In this paper we focus on RB methods due to the stronger norm in parameter, and also efficiency, associated with the
Weak Greedy sampling procedure [8,9], however there are certainly contexts in which the POD or PGD approaches
are preferred.
As typically practiced, the “single-domain” RB method suffers from several limitations which preclude application
to industrial-scale engineering problems:
1. A single-domain RB method only admits parametrizations which induce continuous dependence of the PDE
solution on parameter. Discontinuous parametrizations – for example to describe topology variations, often
crucial in engineering analysis and design – cannot be treated.
2. The single-domain RB method requires, at each parameter value proposed by the Weak Greedy procedure,
solution of a costly underlying finite element (FE) approximation. In many industrial contexts, for example
large structures which perforce exhibit a wide range of scales, even a single FE solution may be prohibitively
expensive.
3. The single-domain RB method is restricted to relatively few parameters. For problems with many parameters,
both the offline cost associated with an extensive training set, and the online cost associated with a rich reduced
basis space, largely eliminate any benefit of model order reduction.
We refer to the above as single-domain RB shortcomings.

We next recall the SCRBE (Static-Condensation Reduced-Basis-Element) method [10–13]. The SCRBE approach
comprises three principal ingredients: component-to-system synthesis, formulated as a Static-Condensation (SC) [14]
procedure; model order reduction, informed by evanescence arguments at component interfaces (port reduction by
spectral truncation) and low-dimensional parametric manifolds in component interiors (bubble reduction by RB
approximation); and offline–online computational decomposition strategies based on affine or approximate affine
expansions in parameter [15]. In practice, both the offline and online stages benefit from parallel implementation
(parallel computation can also effectively serve in the single-domain RB context [16]). We note that the SCRBE
model reduction, just as single-domain RB model reduction, is effected relative to – built upon – an underlying FE
approximation.
The SCRBE approach addresses the three single-domain RB shortcomings: (online) changes in topology are
readily accommodated through component instantiation and connection; both bubble and port training procedures are
typically performed on one-component and two-component systems, and hence we avoid large FE calculations even
in the offline stage; many applications are characterized by spatially distributed parameters, and hence the number of
parameters associated with any given component is relatively small—divide and conquer.
The SCRBE method may be viewed as a combination of the Component Mode Synthesis (CMS) method [17–19]
and the reduced basis method [1–4]: the CMS method provides the foundation of components, systems, and port
and bubble reduction; the RB method provides the framework for efficient parametric analysis. A synthesis of
CMS and RB approaches is first introduced in the Reduced Basis Element method (RBE) of [20]. The SCRBE
method may thus be interpreted as an RBE method for the particular choice of Static Condensation (SC) interface
treatment [10] and particular strategies for port training [11–13] and bubble training [8,9]. (For other applications of
domain decomposition to model order reduction we refer to [21–25].)
The SCRBE method, and in particular the Static Condensation elimination of component-interior degrees of
freedom, is intrinsically limited to linear problems. However, we can nevertheless incorporate the SCRBE method
into nonlinear analyses. In this paper we shall consider a class of problems in which the spatial domain Ω can
be decomposed as Ω = Ω LIN ∪ Ω NLIN such that the field is acted upon by a linear operator over ΩLIN – in our
case, linear elasticity – and a nonlinear operator over ΩNLIN – in our case contact or plasticity. The linear–nonlinear
domain decomposition now admits a hybrid formulation: we consider a “linear” SCRBE approximation over ΩLIN ;
we consider a “nonlinear” approximation (more precisely, an approximation suitable for nonlinear operators) over
ΩNLIN . We shall be interested in local nonlinearities, for which the linear region ΩLIN is as large as, and preferably
substantially larger than, the nonlinear region ΩNLIN ; “linear predominance” is often realistic for contact and plasticity,
in particular for structures in service (though not necessarily for forming processes).
500 J. Ballani et al. / Comput. Methods Appl. Mech. Engrg. 329 (2018) 498–531
There are (at least) two choices for the nonlinear approximation. In the first choice, and most efficiently, we can
consider model order reduction. Reduced basis methods are well developed for quadratic nonlinearities [26] but can
also treat higher-order smooth nonlinearities [27]. Even more recently, reduced basis methods can now accommodate
non-smooth nonlinearities associated with variational inequalities and in particular contact problems [28,29]. In the
second choice, and less efficiently, we may appeal to our assumption of linear predominance to revert to a full FE
approximation in ΩNLIN . We pursue in the current paper the second choice, motivated by ease of implementation
and generality (as regards the form of the local nonlinearity), and justified by the requirement – to be confirmed on
a case-by-case basis – that the nonlinearity acts over only a relatively small region of the domain. We shall thus
denote this hybrid approximation as SCRBE/FE. We can now also better quantify linear predominance: the number
of degrees of freedom associated with the underlying FE approximation over ΩLIN – relative to which we perform
the SCRBE model order reduction – should be larger than, and ideally much larger than, the number of degrees of
freedom associated with the (unreduced) FE approximation over ΩNLIN .
The remainder of this paper is organized as follows. In Section 2 we provide an overview of the SCRBE
methodology in the context of linear elasticity, and in particular we introduce the quantities that will subsequently
serve in the linear–nonlinear hybrid formulation. In Section 3 we describe our particular (conforming) hybrid
formulation, SCRBE/FE, with emphasis on decoupled, or “non-invasive”, treatment of the respective linear and
nonlinear portions of the problem; we also discuss the extension of the SCRBE port training procedures to anticipate
behavior of the field over the linear–nonlinear interface. In Section 4 we introduce the nonlinear PDE models that we
shall consider in this paper, in particular elastic contact without friction, elastic contact with friction, and plasticity
(including plastic contact). Finally, in Section 5, we illustrate for a number of examples the efficiency and flexibility
of the hybrid approach.
2. The SCRBE approximation

In this section we present the SCRBE method in preparation for the hybrid linear–nonlinear formulation,
SCRBE/FE.
2.1. Expositional problem statement
We define here a linear elasticity model which will serve to describe the SCRBE method. We shall affix to certain
quantities a subscript “LIN” to emphasize that the model to which the quantity pertains is linear; this subscript,
appropriately re-interpreted, will then permit us to directly apply the results of the current section to the linear–
nonlinear hybrid approximation of the next section.
We first define model parameter P-tuple µ in parameter domain P ⊂ R P . We then introduce our parametrized
spatial domain ΩLIN (µ) ⊂ R3 with (non-empty) Dirichlet boundary ∂ΩLIN,D (µ), associated function space X LIN (µ) ≡
{v ∈ (H 1 ( ΩLIN (µ) ))3 | v|∂ΩLIN,D (µ) = 0}, and parametrized bilinear and linear forms a(·, ·; µ) : X LIN (µ)× X LIN (µ) →
R and f (·; µ) : X LIN (µ) → R, respectively. The bilinear form shall be given by
∂wi elas ∂vk
∫
a(w, v; µ) ≡ E i jkℓ (µ) d x; (1)
ΩLIN (µ) ∂ x j ∂ xℓ
jkℓ (µ) = λ1 (µ)δi j δkℓ + λ2 (µ)(δik δ jℓ + δiℓ δ jk ), where δi j is the Kronecker-delta symbol and λ1 and λ2 are
here E ielas
respectively the first and second Lamé constants. It is readily demonstrated that a is continuous and coercive over
X LIN (µ). The linear form f may be any bounded linear functional over X LIN (µ), and will typically include both body
forces (such as self-weight) and surface tractions. Note we assume summation over repeated (coordinate) indices.
We may then state our parametrized model: Given model parameter µ ∈ P, find u LIN (µ) ∈ X LIN (µ) such that
a(u LIN (µ), vLIN ; µ) − f (vLIN ; µ) = 0, ∀vLIN ∈ X LIN (µ). (2)

Note the parameter µ will include geometric parameters, related to ΩLIN (µ), and also non-geometric parameters, for
example related to the Young’s modulus, E, and Poisson’s ratio, ν Poisson , which appear in the Lamé constants. The
parametrized model is defined by µ in parameter domain P, the parametrized domain Ω (µ) (here ΩLIN (µ)), and the
parametrized weak form, here (2).
2.2. Components and systems
We define a library as (i) a set of n arch parametrized archetype components which share a common partial
differential equation, and (ii) a set of n fid fiducial ports. We describe first the former and then the latter.
Each archetype component α, 1 ≤ α ≤ n arch , is defined by several attributes:
α
1. A parameter P α -tuple ν α in parameter domain V α ⊂ R P .
2. A physical spatial domain, Dνα . Note in quantities in which α already appears as label we will typically suppress
the superscript α in any associated ν α . The elements of the parameter vector ν α that characterize the physical
spatial domain are referred to as “geometric parameters”; the remaining parameters, for example related to
material properties such as Young’s modulus, Poisson’s ratio, and density, are referred to as “non-geometric”
parameters.
3. A reference spatial domain D̂ α = Dν̂α and associated parametrized geometric mapping Tνα : R3 → R3 such
that the archetype component physical domain Dνα is related to the archetype component reference domain
by Tνα ( D̂ α ); here ν̂ α ∈ V α is the reference parameter value such that Tν̂α is the identity. We shall denote a
point in Dνα (respectively, D̂ α ) by x ≡ (x1 , x2 , x3 ) (respectively x̂ ≡ (x̂1 , x̂2 , x̂3 )). For our purposes here, we
presume that Tνα = Tνrot α Tνdef α , where Tνrot α is a pure rotation independent of x̂, and Tνdef α is any deformation
which does not rotate the ports relatively. Note that the mapping Tνα will in fact depend only on the geometric
parameters.
4. The partial differential equation of interest – in our case here, isotropic linear elasticity – expressed as
parametrized bilinear and linear forms, âνα and fˆνα , respectively, over the reference spatial domain D̂ α . We
recall the standard mapping: given the bilinear form over the component physical spatial domain,
∂wi elas ∂vk
∫
aνα (w, v) ≡ E i jkℓ (ν) d x, (3)
Dνα ∂ x j ∂ xℓ
we effect the change of variable (w, v) = (Tνrot α ŵ, Tνrot α v̂) ◦ (Tνα )−1 , x = Tνα (x̂), to obtain the bilinear form
over the component reference domain,
∂ ŵi def α −1 elas ∂ v̂k
∫
α
âν (ŵ, v̂) ≡ (Jν ) j j ′ E i j ′ kℓ′ (ν) (Jνdef α )−1 |J def α | d x̂, (4)
D̂ α ∂ x̂ j
ℓℓ′
∂ x̂ℓ ν
such that aνα (w, v) = âνα (ŵ, v̂). Here Jνdef α is the Jacobian associated with Tνdef α , and ◦ denotes composition.
A similar procedure relates the physical-domain f να to the reference-domain fˆνα . Note that âνα and fˆνα depend
on both the non-geometric parameters and, through the mapping functions, the geometric parameters as well.
The reference-domain formulation is crucial to the accuracy and efficiency of the SCRBE formulation.
5. A set of I reference-domain “ports”, γ̂ α,i , 1 ≤ i ≤ I , in ∂ D̂ α \ γ̂Dα , and corresponding physical-domain ports
γνα,i = Tνα (γ̂ α,i ); here γ̂Dα is the portion of the reference domain over which Dirichlet conditions are applied (on
all components of the field). The ports shall serve to connect instantiated archetype components. For purposes
of exposition we shall assume that I = 2; the more general case, which permits I = 0, I = 1, and I > 2,
presents no new complications. We shall also assume that the ports are mutually disjoint; the more general case,
in which ports may share a vertex or edge, does present new complications, but may be accommodated through
a wire-basket approach which comprises “vertex” ports and “edge” ports as well as “face” ports. We shall refer
to component ports by “number” or by domain, whichever is most convenient.
6. An underlying finite element mesh and associated finite element approximation space X̂ h;α defined over the
reference spatial domain D̂ α ; note X̂ h;α ⊂ {v ∈ (H 1 ( D̂ α ))3 | v|γ̂Dα = 0}. We further introduce associated bubble
h;α
spaces, X̂ bub ≡ {v ∈ X̂ h;α | v|γ̂ α,i = 0, i = 1, 2}. We denote the dimensions of X̂ h;α and X̂ bub h;α
by N h;α and
h;α
Nbub , respectively.
We shall elaborate further on these ingredients in the development below.

We next describe the set of n fid fiducial ports γkfid , 1 ≤ k ≤ n fid . In any given archetype component α, each
reference port γ̂ α,i must be the image of some parent fiducial port γk(α,i) fid
through a parameter-independent map τ α,i :
γ̂ α,i = τ α,i (γk(α,i)
fid
); typically, τ α,i is a rotation and translation and uniform dilation. Many archetype component
reference-domain ports, and in particular reference-domain ports associated with different archetype components,
will share the same parent fiducial port. For simplicity of presentation, we shall henceforth assume that n fid = 1,
however the general case is readily treated, and indeed illustrated in our examples. We associate to our (now, single)
fiducial port γ fid ≡ γ1fid a set of linearly independent finite element functions χ h;fid
j , 1 ≤ j ≤ Mh , such that, for
each archetype component α in the library, the restriction of any function in X̂ h;α
to γ̂ α,i can be expressed as a
α,i −1
unique linear combination of the χ j ◦ (τ ) , 1 ≤ j ≤ M ; in practice, this constraint is operationalized as a
h;fid h
requirement on the archetype component mesh. Note that, under our assumption that each component has two ports,
h;α
N h;α = Nbub + 2Mh .
We now indicate how we assemble components into a system associated to our parametrized model (2). For any
given µ ∈ P, we associate to our model an assembly – or system – of n inst (µ) instantiated (archetype) components
connected at compatible pairs of ports.
Instantiation. Each instantiated component ℓ corresponds to some archetype component α(µ, ℓ) in our library
for some admissible value of archetype component parameter ν α (µ, ℓ) ∈ V α ; instantiated components which
originate in the same archetype component can (and typically will) be assigned different component parameter
values. We note that, for our distributed systems, ν α (µ, ℓ) for a given instantiated component will often depend
on only a few elements of the model parameter µ ∈ P. We may also define, for instantiated component ℓ,
α(µ,ℓ),i
instantiated component ports, γν(µ,ℓ) , for i = 1 (port 1) and i = 2 (port 2). We emphasize that the model
parameter may include not only variables which relate to geometry but also variables which relate to topology:
different values of µ ∈ P may correspond to different sets of instantiated components; hence the notation
n inst (µ) and α(µ, ℓ) rather than simply n inst and α(ℓ), respectively.
Connection. The system may be described by a set E connect (µ) which comprises compatible pairs of instantiated
(component, component port) couples: for example, the pair {(1, 1), (3, 2)} refers to the connection of
instantiated port 1 of instantiated component 1 to instantiated port 2 of instantiated component 3; note we
shall refer to component ports by “number” as well as by domain, whichever is more convenient. The
compatibility requirement is simply that the two instantiated ports must coincide: for pair {(ℓ, i), (ℓ′ , i ′ )},
α(µ,ℓ′ ),i ′ α(µ,ℓ),i
γν(µ,ℓ′ ) = γν (µ,ℓ) ; in general, two instantiated ports in any given connection pair will share the same
fiducial port. We then introduce n glob (µ) global ports Γ1 , . . . , Γn glob (µ) : each global port is the coalescence
of two connected component ports associated with a pair in E connect (µ). Note that different parameter values
may and often will correspond to different sets of connections and global ports; hence the explicit reference to
µ in E connect (µ) and n glob (µ). Finally, we note that instantiated ports may reside on ∂ΩLIN (µ), in which case by
convention the second member of the connection pair shall be null and there is no compatibility requirement.1
We now illustrate these ingredients graphically. The numerical test cases later in this paper are either rather simple
geometrically or very complicated geometrically, and we thus choose here a distinct example specifically designed
to illustrate archetype components, instantiation and connection, and systems. In particular, we consider a library
for analysis of shafts with n arch = 6 archetype components and n fid = 1 fiducial ports; note for the purposes of this
paragraph we provide the archetype components and fiducial ports with names rather than numbers. We show in Fig. 1
the n arch = 6 archetype components, with two or three views to reveal the (mapped) fiducial ( D ISK) ports, depicted
in blue. (The green surface corresponds to a traction boundary condition, not a port.) In Fig. 2 we show in detail
a particular archetype component, α = G ROOVE, which has two geometric parameters (as well as translation and
rotation, and also property, parameters): ν G ROOVE ≡ Shaft Radius/Library Radius, Groove Radius/Library Radius, ∈
V G ROOVE ≡ [0.4, 0.8] × [2.0, 6.0] (Length/Library Radius = 4 is fixed); note that the Library Radius is the
characteristic length scale for all archetype components in the library. We now consider a particular parametrized
model which corresponds to shaft with groove as well as keyhole: the model parameter µ includes the groove and
keyhole radii as well as the three distances which define respectively the distance before, between, and after the
two stress-concentration features. We show in Fig. 3 a system corresponding to a particular value of µ: we require
three instantiations of the archetype component C YLINDER, one instantiation of the archetype component L OAD
and one instantiation each of the archetype components G ROOVE and K EYHOLE; the component parameters in
each of these instantiated components is dictated by the model parameter µ. Note that Fig. 3 shows on the left the
instantiated components prior to connection (but in position), and on the right the connected instantiated components;
the transparent rendering on the right depicts (in red) the global ports formed through connection.
1 The algorithm extends readily to any number of archetype component ports, I . In the case in which I = 0, the entire system is a single
instantiated component, and SCRBE reduces to single-domain RB.
Fig. 1. Shaft library archetype components: (from top to bottom, and left to right) C YLINDER, L OAD, U NDERCUT, F ILLET, G ROOVE, K EYHOLE.
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Fig. 2. Archetype component G ROOVE and associated geometric component parameters: Shaft Radius, Groove Radius (Length is fixed).
Fig. 3. A shaft system: (left) instantiated components positioned but not yet connected, with component ports indicated in blue and load surface
indicated in green; (right) fully assembled components, with global ports indicated in red. (Note we would apply homogeneous displacement
conditions to the final global port; in actual practice, though not in our exposition here, component ports may be part of the Dirichlet boundary.)
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
2.3. Finite element approximation and static condensation
As already indicated, our model reduction approach is built upon an underlying finite element approximation.
h h
Our finite element approximation space, X LIN (µ), is constructed from our system description. In particular, X LIN (µ)
is formed by “stitching together” the finite element spaces associated with each instantiated component; the latter, in
turn, are induced by the finite element spaces associated with archetype component reference domains. More precisely,
h
X LIN (µ) ≡
rot α(µ,ℓ) α(µ,ℓ) (5)
{ v ∈ X LIN (µ) | ((Tν(µ,ℓ) )−1 v)| Dα(µ,ℓ) ◦ Tν(µ,ℓ) ∈ X̂ h;α(µ,ℓ) , 1 ≤ ℓ ≤ n inst (µ) }.
ν(µ,ℓ)
Note that the conforming condition is very easily satisfied thanks to the requirements on (i) archetype component
reference-domain finite element approximations, in particular related to a (common) fiducial port mesh, and
h h
(ii) connection pairs, in particular related to geometric compatibility. In fact, the dimension of X LIN (µ), NLIN (µ),
∑n inst (µ) h;α(µ,ℓ) glob h
is simply ℓ=1 Nbub + n (µ)M .
We can now also be more precise in our definition of a(w, v; µ) in particular related to the parameter dependence
implicit in (1). By construction,
a(w, v; µ) ≡
n inst (µ)
∑ α(µ,ℓ) rot α(µ,ℓ) α(µ,ℓ) rot α(µ,ℓ) α(µ,ℓ) (6)
âν(µ,ℓ) (((Tν(µ,ℓ) )−1 w)| Dα(µ,ℓ) ◦ Tν(µ,ℓ) , ((Tν(µ,ℓ) )−1 )v| Dα(µ,ℓ) ◦ Tν(µ,ℓ) );
ν(µ,ℓ) ν(µ,ℓ)
ℓ=1
note that we invoke only the archetype component reference-domain bilinear form. We may view (6) as an abstract
form of direct stiffness summation applied at the component level. A similar construction exists for the synthesis of
f (v; µ) of (1) from the linear forms associated to our archetype components.
We may now state our finite element approximation to (2): Given model parameter µ ∈ P, find u LIN
h h
(µ) ∈ X LIN (µ)
such that
h
a(u LIN (µ), vLIN ; µ) − f (vLIN ; µ) = 0, ∀vLIN ∈ X LIN
h
(µ). (7)
This underlying finite element approximation shall serve as the foundation on which we build our SCRBE
approximation . We will thus ultimately need to confirm both that our SCRBE approximation is sufficiently close
h h
to u LIN (µ) and furthermore that u LIN (µ) is sufficiently close to u LIN (µ).
We now perform static condensation to obtain a bubble-eliminated finite element space,
Y h (µ) = span{Φmh (µ), 1 ≤ m ≤ J h (µ)}, (8)
h glob h
where J (µ) ≡ n (µ)M is the total number of (global) port degrees of freedom associated with our finite element
approximation. Note each function Φ·h is associated to a particular global port, say Γk , shared by two instantiated
components, say ℓ and ℓ′ , and a particular fiducial port mode, say χ h;fid
j : Φ·h replicates χ h;fid
j (appropriately mapped)
on Γk , vanishes at all global ports other than Γk , is non-zero only over instantiated components ℓ and ℓ′ , and satisfies
the finite element equations for all bubble test functions in instantiated components ℓ and ℓ′ . We may then write our
finite element approximation, for given µ ∈ P, as
h h;hom h;inhom
u LIN (µ) = u LIN (µ) + u LIN (µ), (9)
h;hom
for u LIN (µ)∈ Y h (µ) solution of
( )
h;hom
a(u LIN (µ), v; µ) − f (v; µ) − a(u LIN
h;inhom
(µ), v; µ) = 0, ∀v ∈ Y h (µ). (10)
h;inhom
Here u LIN (µ), which comprises a bubble function for each instantiated component, addresses inhomogeneities
h;inhom
(loads) interior to the components; u LIN (µ) may be calculated independently on each instantiated component, and
thus poses little computational difficulty.
To arrive at the algebraic representation we insert the expansion
h (µ)
J∑
h;hom h h
u LIN (µ) = ULIN m ′ (µ) Φm ′ (µ) (11)
m ′ =1
into (10) and test on Φmh (µ), 1 ≤ m ≤ J h (µ), to obtain our (statically condensed) linear system of equations: Given
h
µ ∈ P, find ULIN
h
(µ) ∈ R J (µ) solution of
h h h
SLIN (µ)ULIN (µ) = FLIN (µ). (12)
h h (µ)×J h (µ)
Here SLIN (µ) ∈ R J is given by
h
(SLIN (µ))mm ′ = a(Φmh ′ (µ), Φmh (µ); µ), 1 ≤ m, m ′ ≤ J h (µ); (13)
h h J h (µ)
the matrix SLIN (µ) is denoted the Schur complement. The vector FLIN (µ)
∈R may be expressed in terms of f ,
h;inhom
a, the Φ·h , and (known) u LIN (µ). Note that linearity, a requirement for the static condensation procedure , ensures
that the sum (11) satisfies the finite element equations for all bubble test functions in all instantiated components.
To motivate the model reduction, we note two computational difficulties associated with the classical Static
Condensation approach. First, the number of finite element degrees of freedom on the ports, Mh , will be large,
resulting in a Schur complement with a large – 2Mh × 2Mh – dense block for each instantiated component in the
system. Thus the solution of the Schur complement system (12) will be costly. Second, the calculation of each of
the J h (µ) ≡ n glob (µ) Mh functions Φ·h – which recall satisfy the finite element equations for bubble test functions
h;α(µ,ℓ)
in instantiated components ℓ and ℓ′ – will be expensive: solution of two (sparse) linear systems of size Nbub ×
h;α(µ,ℓ) h;α(µ,ℓ′ ) h;α(µ,ℓ′ )
Nbub and Nbub × Nbub , respectively. Thus the formation of the Schur complement will be costly.
2.4. Reduced order approximation
2.4.1. Formulation
There are two aspects to the model order reduction: port reduction (PR) – we retain just the first M port modes,
for M ≪ Mh ; reduced basis (RB), or bubble, reduction – we replace, for 1 ≤ α ≤ n arch (µ), archetype component
h;α h;α
(bubble) finite element space X̂ bub of dimension Nbub with 2M +1 reduced basis approximation spaces of (maximum)
α α h;α
dimension N , with N ≪ Nbub . As regards port reduction, we note that our simple truncation of port modes
presumes that the port modes are ordered in an appropriate hierarchical fashion. As regards bubble reduction, we note
that we create a tailored space for each port and each port mode as well as a bubble space for any loads applied to
the component. (With respect to the latter, we assume for convenience that no loads are applied over the component
ports.)
We may then propose the reduced bubble-eliminated space
Y h,M,N (µ) = span{Φmh,M,N (µ), 1 ≤ m ≤ J M (µ)}, (14)
M glob
where J (µ) ≡ n (µ)M; note h refers to the underlying finite element approximation space, M refers to the
number of port modes retained, and N refers to the reduced basis approximation of Φ·h . We may then write our
SCRBE approximation, for given µ ∈ P, as
h,M,N h,M,N ;hom h,M,N ;inhom
u LIN (µ) = u LIN (µ) + u LIN (µ), (15)
h,M,N ;hom
for u LIN (µ) ∈ Y h,M,N (µ) solution of
( )
h,M,N ;hom
a(u LIN (µ), v; µ) − f (v; µ) − a(u LIN
h,M,N ;inhom
(µ), v; µ) = 0, ∀v ∈ Y h,M,N (µ). (16)
h,M,N ;inhom h;inhom
Here u LIN (µ) is the reduced basis approximation of u LIN (µ); the former, just as the latter, may be calculated
independently on each instantiated component. The approximation (16) will enjoy the usual optimality properties
associated with Galerkin projection.
We now proceed to the linear algebraic representation. In particular, we insert the expansion
M (µ)
J∑
h,M,N ;hom h,M,N h,M,N
u LIN (µ) = ULIN m ′ (µ) Φm ′ (µ) (17)
m ′ =1
into (16) and test on Φmh,M,N (µ), 1 ≤ m ≤ J M (µ), to obtain the SCRBE system: Given µ ∈ P, find ULIN
h,M,N
(µ) ∈
J M (µ)
R solution of
h,M,N h,M,N h,M,N
SLIN (µ)ULIN (µ) = FLIN (µ), (18)
h,M,N J M (µ)×J M (µ)
where SLIN (µ) ∈ R , the reduced Schur complement, is given by
h,M,N
(SLIN (µ))mm ′ = a(Φmh,M,N
′ (µ), Φmh,M,N (µ); µ), 1 ≤ m, m ′ ≤ J M (µ). (19)
h J M (µ) h,M,N ;inhom
The vector FLIN (µ) ∈ R may be expressed in terms of f , a, the Φ·h,M,N , and (known) u LIN .
In actual practice, we typically pursue a Petrov–Galerkin variant of the Galerkin procedure. In particular, we test
not on functions in Y h,M,N (µ), but rather on functions in Ytest h,M,N
(µ) = span{Ψmh,M,N , 1 ≤ m ≤ J M (µ)}. Here
h,M,N
Ψ· , associated to global port Γk , is a parameter-independent function which replicates a particular fiducial port
mode (appropriately mapped) on Γk , vanishes at all global ports other than Γk , and is suitably smooth over the
instantiated components ℓ and ℓ′ which share Γk . The Petrov–Galerkin approach is computationally more efficient
than the Galerkin approach since the Ψ·h,M,N are parameter-independent (more precisely, the Ψ·h,M,N are the images
of parameter-independent functions defined over the parent archetype component reference domain). As regards the
resulting approximation, the Petrov–Galerkin variant affects only stability and not best approximation, and in fact for
sufficiently accurate RB approximation spaces the difference between the Petrov–Galerkin and Galerkin formulations
is small. Justified by the latter, we also choose to symmetrize the Petrov–Galerkin Schur complement for purposes of
more stable and also more efficient solution of the linear algebraic system. In our exposition we continue to consider
the simpler Galerkin projection, in particular since the hybrid approach is applicable to both Galerkin and Petrov–
Galerkin approximation; we return to the latter point in the next section.
The reduced order approximation eliminates the shortcomings of FE static condensation. Let us assume,
optimistically, that M ≪ Mh and N α ≪ Nbub h,α
, 1 ≤ α ≤ n arch (µ). In that case we successfully address the two
computational challenges associated with Static Condensation. The first computational issue, the large number of port
modes Mh , is addressed by truncation of our port mode expansion in M ≪ Mh terms. The reduction Mh → M
h
decreases the cost to form the Schur complement and the cost to solve the Schur complement: SLIN (µ), which is
h,M,N
block-sparse n (µ)M ×n (µ)M , is replaced by SLIN (µ), which is block-sparse n (µ)M ×n glob (µ)M. The
glob h glob h glob
second computational issue, the cost of the bubble contributions, is addressed by the reduced basis approximation.
The reduction Nbub h;α
→ N α decreases the cost to form the Schur complement: the (sparse) Nbub h,α h,α
× Nbub FE matrices
h α α
required to calculate Φ· are replaced by the (dense) N × N matrices – henceforth “RB matrices” – required to
calculate Φ·h,M,N .
2.4.2. Port reduction

The SCRBE bubble reduction [10] is a standard extension of single-component reduced basis best practices [8,9]
which furthermore requires no modification for the hybrid formulation of the current paper: bubble reduction by
construction considers homogeneous Dirichlet data on component ports, and may thus be pursued without reference
to ΩNLIN . We note here only that the training of the reduced basis (bubble) spaces may proceed independently on
each archetype component and furthermore without reference to any particular system or parametrized model. We
also emphasize that we construct an RB space for each port and port mode, which effectively removes the port modes
from the parametrization for any given RB space, and thereby ensures low-dimensional RB spaces.
On the other hand, the SCRBE port reduction procedure must be adjusted to accommodate the hybrid formulation
of the current paper. We thus briefly describe the standard SCRBE port training procedures. To begin, we discuss
the rationale for reduction in the number of port degrees of freedom. It is well known, from simple separation-of-
variable arguments, that (for elliptic partial differential equations) higher-wavenumber profiles over a cross-section
will decay exponentially rapidly in the direction normal to the cross-section; in acoustics, the phenomenon is known as
“evanescence” [30]. This then provides the motivation, or justification, for port reduction: even if within instantiated
components many higher modes may be (locally) excited by geometric variations, these evanescent modes will largely
decay before reaching the corresponding instantiated ports. (We thus also identify an archetype component design best
practice: to the extent possible, we should separate ports from regions of rapid or discontinuous geometric or property
variation.)
It remains to exploit evanescence in a constructive algorithm. There are several options for the identification of
optimal port modes in the context of parametrized partial differential equations [12,13]. All follow a similar procedure:
define training systems of instantiated components; collect data (traces) on training system instantiated ports; map
instantiated port data to parent fiducial port; apply POD to obtain fiducial port modes χ h;fid
j , 1 ≤ j ≤ M (for each
fiducial port). Note for the training systems we can identify “trainee” ports – on which we collect the data to be
reduced – and “trainer” ports – on which we impose representative (say) Dirichlet data.
We summarize two port training procedures:
• Method I: the training systems corresponds to models (or parts of models) of interest – each typically with
many trainee ports – for (i) plausible Dirichlet data on trainer ports, and (ii) a large set of samples of model
parameters. This approach is general, readily implemented, easily extended to a wide variety of problems, and
often yields a small (but sufficient) number of port modes, M. However, the method requires construction and
evaluation of potentially very large systems, the port modes identified are only reliable for the particular models
sampled, and furthermore there is no theory for convergence.
• Method II: the training systems correspond to two-component systems instantiated from all possible pairs of
archetype components in a given library for (i) optimal Dirichlet data on trainer ports, as identified by a transfer
eigenproblem [31,32,13], and (ii) a large set of samples of component parameters. This method does not require
the construction or evaluation of any (large) models, yet nevertheless provides port spaces relevant to a large
family of (possible) models; furthermore, the method is supported by a theoretical framework, and in certain
instances yields provably optimal port spaces. However, the training procedure is expensive, and the number of
port modes can be excessive for a library with many archetype components.
In this paper, we exercise both Method I and Method II, the former for small models, the latter for large models.
2.4.3. Computational procedure

We now turn to the SCRBE computational procedures. There are two stages to the SCRBE approach, an offline
stage, and an online stage. In the offline stage we (i) train the port modes and reduced basis bubble functions, and
(ii) store, in an Online Dataset, parameter-independent finite element inner products required to form the entries of
h,M,N
the RB matrices and also the reduced Schur complement SLIN (µ). Note the latter requires an affine-in-parameter
expansion of the bilinear and linear forms, typically effected (for nontrivial geometric variations) by the Empirical
Interpolation Method, or EIM [15,4]. In the online stage, given a particular model and model parameter µ, we invoke
the Online Dataset to form and subsequently solve the reduced Schur complement system, (18); the latter is typically
effected by direct techniques in particular given that the reduced Schur complement is small and block sparse. The
essential point is that the operation count (and storage) of the online stage, apart from any field visualization, is
h
independent of the resolution of the underlying finite element approximation, NLIN (µ).
Finally, we comment on the important role of components in the computational efficiency of the SCRBE approach.
First, as regards parametrization, we reduce a single large problem – the full parametrized model – with many
parameters to many small problems with just a few parameters – the instantiated components; we thus address, by
divide and conquer, the curse of dimensionality which often plagues reduced basis techniques. Second, in the offline
stage (at least for port training Method II), we solve finite element problems only over pairs of components or single
components; we rarely solve – except as possible in final verification studies – finite element problems associated
with potentially very large systems. Third, we may amortize our offline effort not just over many queries (for different
model parameter values) to a particular parametrized model, but also over all parametrized models which we may
synthesize from our library of archetype components; we can thus tolerate higher offline costs. Fourth, we facilitate
the development of new models: a common impediment to model order reduction without components is the relative
inflexibility imposed by a priori specification of a particular model parameter domain, and the upfront “undiversified”
risk associated with substantial offline investment for a single problem; both issues are mitigated by a component
strategy.
3. Two-domain formulation of linear–nonlinear model

3.1. Problem statement
We describe here the general linear–nonlinear parametrized model. We first define the model parameter P-tuple µ
in parameter domain P ⊂ R P . Note that the parametrized linear–nonlinear model of this section is different from the
parametrized linear model of Section 2, and hence the corresponding interpretations of µ, P, and P are also different.
We then introduce our parametrized model spatial domain Ω (µ) ⊂ R3 with (non-empty) Dirichlet boundary
∂ΩD (µ). We shall decompose Ω (µ) into two open subdomains ΩLIN (µ) and ΩNLIN (µ) such that
⋃ ⋂
Ω LIN (µ) Ω NLIN (µ) = Ω (µ), ΩLIN (µ) ΩNLIN (µ) = ∅. (20)
We further define the interface between the two subdomains as
⋂
ΓINT (µ) ≡ Ω LIN (µ) Ω NLIN (µ) . (21)
We shall assume that ∂ΩD (µ) is the union of disjoint ∂ΩLIN,D (µ) and ∂ΩNLIN,D (µ); we further assume, for simplicity
of exposition, that all Dirichlet (displacement) conditions are homogeneous, and that furthermore the Dirichlet
boundaries are disjoint from ΓINT . The subdomains ΩLIN (µ) and ΩNLIN (µ) are distinguished in two related ways:
the field over ΩLIN (µ) shall be acted upon only by linear operators, and in particular the linear elasticity operator,
whereas the field over ΩNLIN (µ) may be acted upon by nonlinear operators; the field over ΩLIN (µ) shall be represented
by a SCRBE approximation (fundamentally restricted to linear models), whereas the field over ΩNLIN (µ) shall be
represented by a full finite element approximation. In the current section we focus on the former. In what follows, for
any function w defined over Ω (µ), wLIN (respectively, wNLIN ) shall refer to the restriction of w to ΩLIN (µ), w|ΩLIN (µ)
(respectively, the restriction of w to ΩNLIN (µ), w|ΩNLIN (µ) ).
We introduce the space X LIN (µ) ≡ {v ∈ (H 1 (ΩLIN (µ)))3 | v|∂ΩLIN,D = 0} and also the space X NLIN (µ) ≡ {v ∈
(H 1 (ΩNLIN (µ)))3 | v|∂ΩNLIN,D = 0}. We may then define the space over Ω (µ) as X (µ) ≡ {v ∈ (H 1 (Ω (µ)))3 | vLIN ∈
X LIN , vNLIN ∈ X NLIN }. We next introduce forms f (·; µ) : X LIN (µ) → R, a(·, ·; µ) : X LIN (µ) × X LIN (µ) → R, and
g(·, ·; µ) : X NLIN (µ) × X NLIN (µ) → R which are respectively linear, bilinear, and linear in the second argument;
we shall also require the Gateaux derivative of g about w̄ ∈ X NLIN , which we represent by the bilinear form
(w ∈ X NLIN (µ), v ∈ X NLIN (µ)) → dg(w, v; w̄; µ).
We may now state the weak form of our partial differential equation: Given µ ∈ P, find u(µ) ∈ X (µ) such that
a(u LIN (µ), vLIN ; µ) − f (vLIN ; µ) + g(u NLIN , vNLIN ; µ) = 0, ∀v ∈ X (µ); (22)
note that g may include inhomogeneous terms, and in particular loads within ΩNLIN (µ). We emphasize that the flux
continuity conditions on the field on the interface ΓINT are incorporated (weakly) directly into (22), in particular
associated with test functions v which do not vanish on ΓINT , per standard practice for conforming variational
approximations. We shall suppose for our purposes here that, for the parameter domain prescribed, our forms satisfy
the necessary continuity and stability conditions such that (22) admits a unique solution. In fact, our approach can
also be applied to problems with multiple solution branches and even bifurcations by appropriate enhancement of the
SCRBE training procedures – to include behavior from different branches – and associated solution algorithms – to
incorporate continuation information.
The linear model – the model over ΩLIN (µ), corresponding to the first two terms in (22) – is precisely the linear
model considered in Section 2: weak statement (2) for bilinear form (1) (with the interpretation of µ, P, and P suitably
adapted to the linear–nonlinear model). The presence of the third (nonlinear) term in (22) serves, from the perspective
of the linear model, solely to provide the appropriate inhomogeneous Neumann term. (The latter, of course, only
emerges from the solution of the coupled linear–nonlinear model.) Although our linear model is universal to all the
problems that we shall consider in this paper, the nonlinear model – the model over ΩNLIN (µ) – will vary from problem
to problem. The weak form g will be specified (in the final sections of this paper) in the context of each family of
nonlinearities considered.
3.2. Hybrid method: SCRBE/FE
3.2.1. Formulation
We consider here the SCRBE/FE hybrid approximation [33]: SCRBE for the linear model over ΩLIN (µ), and
full finite element (FE) over ΩNLIN (µ). There are of course many ways in which we might couple two different
discretizations, including both nonconforming approaches and conforming approaches. In general, the interface
conditions are treated as constraints which may be imposed either through Lagrange multipliers and penalty terms
or by explicit elimination. In our work here we choose the simplest route: conforming approximations and explicit
constraint elimination, both facilitated by the fiducial port structure associated with the (conforming) SCRBE method.
As already noted, the linear part of the linear–nonlinear model is identical to the linear model of Section 2. We
may thus directly apply our SCRBE formulation of the previous section to the linear model in the current section.
We require only that the SCRBE approximation is constructed such that ΓINT corresponds to a single global port,
which for convenience, and without loss of generality, we choose as Γ1 . We shall further assume, to unencumber the
already hopelessly encumbered notation, that Γ1 is independent of µ. (In actual practice, and indeed in our numerical
examples, ΓINT may correspond to any number of global ports, all of which may be parameter-dependent.) We choose
our enumeration of components and ports to ensure that global port Γ1 corresponds to instantiated port 1 of instantiated
component 1 derived from archetype component 1. We further choose our index mapping such that the SCRBE degrees
of freedom associated with Γ1 correspond to the first M basis functions of Y h,M,N (µ), Φmh,M,N (µ), 1 ≤ m ≤ M.
Before proceeding to the SCRBE/FE approximation we identify the global finite element approximation over
Ω (µ). We recall that the component synthesis of the system associated to our model induces, for any given µ ∈ P,
h h
an underlying finite element approximation, X LIN (µ) of dimension NLIN (µ), given by (5). We next introduce the
h h
finite element space for the nonlinear model over ΩNLIN (µ), X NLIN (µ) ⊂ X NLIN (µ) of dimension NNLIN (µ). We shall
require that X LIN (µ) and X NLIN (µ) strongly respect the interface ΓINT (µ): for any v1 ∈ X LIN (µ) (respectively, for any
h h h
v2 ∈ X NLIN
h
(µ)) there exists a v2 ∈ X NLIN
h
(µ) (respectively, a v1 ∈ X LIN
h
(µ)) such that v1 |ΓINT (µ) = v2 |ΓINT (µ) . The latter
h h
of course imposes a constraint on the meshes associated with X LIN (µ) and X NLIN (µ) : coincident same-order element
faces on ΓINT . We shall later discuss a simple strategy by which to enforce this constraint, in particular through fiducial
h
ports. Finally, we shall also need the space X NLIN,0 h
(µ) ≡ {v ∈ X NLIN (µ) | v|ΓINT (µ) = 0}; X NLIN,0
h
(µ) is of dimension
h h h h h
NNLIN,0 (µ) = NNLIN (µ) − M (given that ΓINT is a single port). We then associate to X NLIN,0 (µ) and X NLIN (µ) nodal
bases {κi (µ)}i=1,...,N h and {κi (µ)}i=1,...,N h , respectively; note that, for convenience in the subsequent algebraic
NLIN,0 NLIN
h
representation, we enumerate last the nodal degrees of freedom associated with X NLIN (µ) on the interface, ΓINT .
The global finite element space over Ω (µ), X (µ) of dimension N (µ), is thus given by X h (µ) ≡ {v ∈
h h
X (µ) | vLIN ∈ X LIN

h
(µ), vNLIN ∈ X NLIN
h
(µ)}; under our assumptions on the constituent finite element spaces,
h h
N (µ) = NLIN (µ) + NNLIN (µ) − Mh , where the last term corrects for the duplication of degrees of freedom on
h
(the single-port interface) ΓINT (µ). We can then state our underlying global finite element approximation: Given
µ ∈ P, find u h (µ) ∈ X h (µ) such that
h
a(u LIN (µ), vLIN ; µ) − f (vLIN ; µ) + g(u NLIN
h
, vNLIN ; µ) = 0, ∀v ∈ X h (µ). (23)
h
Our SCRBE/FE approximation is built upon the global finite element approximation (23): X NLIN (µ) is the finite
h
element space for the full (unreduced) FE approximation over ΩNLIN (µ); X LIN (µ) is the underlying finite element
space for the SCRBE approximation over ΩLIN (µ). We must thus be certain both that our SCRBE/FE approximation
is sufficiently close to u h (µ), and that u h (µ) is sufficiently close to u(µ). The global finite element approximation
u h (µ) will also serve as a reference with respect to which we can assess the computational cost of the SCRBE/FE
approach.
We now turn to the SCRBE/FE approximation. We first introduce the SCRBE space over ΩLIN (µ), Y h,M,N (µ) of
dimension J M (µ), as defined in (14). Our hybrid approximation space is then given by X H (µ) ≡ {v ∈ X h (µ) | vLIN ∈
Y h,M,N (µ), vNLIN ∈ X NLIN
h
(µ)}; note that implicitly the degrees of freedom on the interface are dictated by Y h,M,N (µ),
since M ≤ Mh . The superscript H represents the several discretization parameters which characterize X H (µ),
including the finite element resolution in ΩLIN (µ) and ΩNLIN (µ) and also M and N α , 1 ≤ α ≤ n arch (µ). The dimension
of space X H (µ) will be denoted N H (µ); it shall also prove convenient to define N H,+ (µ) = N H (µ) + Mh , which
includes the finite element “non-degrees of freedom” on ΓINT .
We define our hybrid approximation as
h,M,N ;inhom
u H (µ) = u H ;hom (µ) + u LIN (µ), (24)
h,M,N ;inhom
for u LIN (µ) defined in (15) (and extended by zero to the entire domain, Ω (µ)). Given µ ∈ P, u H ;hom (µ) ∈
H
X (µ) satisfies
( )
H ;hom
a(u LIN (µ), vLIN ; µ) − f (vLIN ; µ) − a(u LIN
h,M,N ;inhom
(µ), vLIN ; µ)
(25)
H ;hom
− g(u NLIN (µ), vNLIN ; µ) = 0, ∀v ∈ X H (µ).
We recall that g may include inhomogeneous terms, and in particular loads within ΩNLIN (µ); the superscript “hom”
h,M,N ;inhom
refers only to elimination of the inhomogeneities in ΩLIN (µ) thanks to the bubble contributions in u LIN (µ). We
shall focus attention on a formulation which preserves the decoupled nature of the heterogeneous approximation, in
particular a strict decomposition into “non-invasive” linear and nonlinear contributions.
α=1
To begin, we define liftings of the port functions on the interface ΓINT , ξ j ≡ (χ h;fid
j ◦(τ α=1,i=1 )−1 )◦(Tν(µ,ℓ=1) )−1 , 1 ≤
j ≤ M, over ΩNLIN (µ): for j = 1, . . . , M, LNLIN ξ j is given by
h
NNLIN (µ)
∑
(LNLIN ξ j )(µ) = ξi, j (µ)κi (µ), (26)
i=1
for ξi, j (µ), 1 ≤ i ≤ NNLIN h

(µ), chosen such that LNLIN ξ j |ΓINT = ξ j ; such a construction is perforce possible given
our assumptions on the underlying finite element approximation. (Recall that the κi (µ) are the nodal basis functions
h
associated with X NLIN (µ), enumerated such that the degrees of freedom associated with ΓINT appear last.) We note
that a particularly simple, and also efficient, choice for the lifting coefficients is ξi, j = 0, 1 ≤ i ≤ NNLIN,0
h
(µ); we
N H,+ ×N H
consider that particular case in this paper. We then define a constraint, or prolongation, matrix C ∈ R as
δi, j h
1 ≤ i, j ≤ NNLIN,0
⎧
⎪ (27)-a
⎨ξ
⎪
h 1 ≤ i ≤ NNLIN , NNLIN,0
h h h
+ 1 ≤ j ≤ NNLIN,0 + M (27)-b
Ci, j ≡ i, j−NNLIN,0
, (27)
⎩δi−Mh , j + 1 ≤ i ≤ NH+ , NNLIN,0
h h
⎪
⎪ NNLIN + 1 ≤ j ≤ NH (27)-c
0 otherwise (27)-d
where δi, j is the Kronecker-delta symbol. It is understood that C may depend on µ, even for ξ j , 1 ≤ j ≤ M,
independent of µ, since the approximation spaces and the dimensions of the approximation spaces may depend on µ.
In the remainder of this section we continue to leave the latter implicit.
In order to de-mystify the construction (27) we consider the action of C. In particular, given a basis-coefficient
H M h
vector W H ∈ RN , which we further decompose as W H ≡ (WLIN H
, WNLIN
H H
) for WLIN ∈ R J , WNLIN
H
∈ RNNLIN,0 , we
can express any member w ∈ X H (µ) as
h
NNLIN M
∑ J
∑
w= (C W )m κm (µ) +
H
(C W H )N h Φmh,M,N (µ); (28)
NLIN +m
m=1 m=1
here we interpret κ· (respectively, Φ·h,M,N ) outside Ω NLIN (µ) (respectively, outside Ω LIN (µ)) as zero. Note that w is
continuous over Ω (µ), since the coefficients in the two sums in (28) are coupled – through C – to ensure a common
representation over ΓINT . In words, (27)-a directly “passes through” the coefficients of the interior degrees of freedom
within ΩNLIN (µ); (27)-b impresses the trace (over ΓINT ) of the approximation over ΩLIN (µ) onto the approximation
over ΩNLIN (µ); and (27)-c directly passes through the coefficients of the degrees of freedom within ΩLIN (µ). As
regards the latter, (27)-c constitutes a J M × J M identity block in the lower right corner of C, and thus we may also
write
h
NNLIN M
∑ J
∑
w= (C W H )m κm (µ) + H
WNLIN Φmh,M,N (µ); (29)
m=1 m=1
we require the identity block for the transpose operation, which we presently enlist.
H
We introduce the basis-coefficient vector U H (µ) ∈ RN in terms of which we express u H ;hom (µ) through an
M h
expansion of the form (28); we further decompose U H (µ) as (ULIN H
, UNLIN
H H
) for ULIN ∈ R J , UNLIN
H
∈ RNNLIN,0 .
We similarly introduce the basis-coefficient vector V H in terms of which we express any test function v ∈ X H (µ)
through an expansion of the form (28). We now insert our expansions for u H ;hom (µ) and v into (25) to arrive at a
H
discrete statement of our heterogeneous approximation: find U H (µ) ∈ RN such that
C T R H,+ (U H (µ); µ) = 0, (30)

T H,+ N H,+
where denotes matrix transpose. Here the residual vector R ∈R is given by
h
(U H (µ); µ)
( )
G NLIN
R H,+ (U H (µ); µ) ≡ h,M,N H h,M,N , (31)
SLIN (µ)ULIN (µ) − FLIN (µ)
H h
where, for any W H ∈ RN , G NLIN
h
(W H ; µ) ∈ RNNLIN is given by
h
NNLIN
∑
h
(G NLIN (W H ; µ))i ≡ g( (C W H ) j κ j , κi ; µ), 1 ≤ i ≤ NNLIN
h
, (32)
j=1
h,M,N h,M,N
and SLIN (µ) and FLIN (µ) are defined in (18)–(19). The residuals on ΩLIN (µ) and ΩNLIN (µ) are coupled through
C T , which imposes the necessary weak flux continuity condition through the Neumann operators on the respective
domains.
We make several observations. First, we note that the nonlinear contribution is restricted solely to ΩNLIN (µ).
Second, we note from (31)–(32) that the residual may be calculated independently on each subdomain, ΩLIN (µ)
and ΩNLIN (µ), and then connected very “sparsely” through the constraint matrix C; in particular, the residual over
ΩLIN (µ) is the standard ΓINT -Neumann residual over ΩLIN (µ), with no reference to (the nonlinearity on) ΩNLIN (µ).
The advantages of the linear–nonlinear decomposition are well established in development of more efficient solution
algorithms: static condensation (equivalently, substructuring) may be performed on the linear part of the model only
once, at the outset, rather than at each step of a nonlinear iteration [34]. Our emphasis, and innovation, is the efficient
treatment of parametric variation: we take advantage of the linear–nonlinear decomposition to very easily adapt,
as regards both formulation and also implementation, the SCRBE approach and attendant benefits to parametrized
models with local nonlinearities.
We shall solve our nonlinear system by Newton’s Method, which thus requires both the residual as well as the
H
Jacobian. We briefly describe here the representation of the Jacobian. We denote by W̄ H ∈ RN the basis vector
H H
associated with some member w̄ in X H (µ) (in actual practice, the current iterate). The Jacobian J H (µ) ∈ RN ×RN
can then be expressed as
J H (µ) ≡ C T J H,+ (µ) C (33)

for
dG NLIN (W̄ H ; µ)
( h )
0
J H,+ (µ) ≡ h,M,N , (34)
0 SLIN (µ)
h h
h
where dG NLIN ∈ RNNLIN × RNNLIN is given by
h
NNLIN
∑
h
(dG NLIN (W̄ H ; µ))i, j ≡ dg(κ j , κi ; (C W̄ H )k κk ; µ), 1 ≤ i, j ≤ NNLIN
h
. (35)
k=1
H,+ H,+
We observe that J H,+ (µ) ∈ RN ×N (or the action of J H,+ (µ)) may be evaluated independently on ΩLIN (µ)
and ΩNLIN (µ), and hence the only coupling between the two subdomains is through the very sparse constraint matrix
C. We further note that the Jacobian on ΩLIN (µ) is the standard ΓINT -Neumann Jacobian – stiffness matrix – over
ΩLIN (µ), with no reference to (the nonlinearity) on ΩNLIN (µ). Hence, as for the residual, the SCRBE part of the
formulation and implementation is effectively “ignorant” of the linear–nonlinear hybrid approximation to which it so
courageously contributes.
Finally, we return to the motivation for the linear–nonlinear hybrid approximation: the size of our discrete system,
h
and hence the cost of residual and Jacobian inversion, can be greatly reduced – from N h to NNLIN,0 + J M – if the
extent of the nonlinear domain is relatively small. In this regard, the sparsity associated with decoupled evaluation
plays an important role. We also remark here that the linear–nonlinear coupling strategy described here is applicable
to both the Galerkin and Petrov–Galerkin formulations since Φmh,M,N (µ) and Ψmh,M,N coincide on global ports and
hence on ΓINT .
3.2.2. Coupling considerations

In the development of the hybrid approximation we have not, apart from the identification of the interface with a
global port, in any way adapted the standard SCRBE approach. There are, however, two aspects of the coupling which
do require special attention: development of a flexible framework for construction of a compatible finite element
approximation over ΩNLIN (µ); incorporation of the nonlinear model in the port training of the SCRBE approximation
of the linear model. (We note that the reduced basis training is restricted to bubbles and thus by definition interior to
ΩLIN (µ)—hence affected by the nonlinear model only indirectly, through the port modes.)
We first consider the finite element preparations over ΩNLIN (µ). There are many ways to proceed, however we
advocate a component approach which is directly compatible with the SCRBE formulation. Of particular interest are
dual-purpose components which can serve either within ΩLIN (µ) as part of the SCRBE approximation of the linear
model or within ΩNLIN (µ) as part of the unreduced FE approximation of the nonlinear model. In effect, we form and
train the components for the linear regime. Then, if in any given model and in any particular instantiated component
our linearity criteria (such as no contact with other bodies, maximum stress less than yield stress) are satisfied, we
include the component in ΩLIN (µ) and take advantage of the SCRBE acceleration; these components are denoted
“SCRBE components”. If, however, the linearity criteria are not satisfied, the instantiated component is assigned
to ΩNLIN (µ) and the corresponding (formerly, part of the underlying) finite element approximation is “stitched in”
h
to the unreduced space X NLIN (µ); these components are denoted “FE components”. This process can be effected
dynamically, for example as part of a loading path, since all the necessary data is available in the Online Dataset. (For
another example, in a different context, of online reduced basis adaptivity, we refer to [35].) The assignment from
linear to nonlinear can in some cases be reversed, for example in the context of elastic contact, but in other cases
not, for example in the context of plasticity. In all cases the fiducial ports ensure easy integration and compatible
linear–nonlinear approximations.
We note that dual-purpose components are not always of interest, especially in situations in which the behavior in a
component is known a priori to be nonlinear. In particular, there is a computational overhead associated with dynamic
allocation: we must evaluate our linearity criteria in all instantiated components. Furthermore, since the linearity
criteria are typically evaluated in terms of nonlinear output functionals (of the displacement), such as the maximum
operator, the cost can be considerable: we must appeal, even in ΩLIN (µ), to the underlying FE approximation. In
actual practice, and as we quantify subsequently, verification of linearity criteria is typically not a predominant cost.
Furthermore, it is often possible to facilitate the evaluation of certain linearity criteria through search heuristics: for
contact, we might target instantiated components based on the (linear-functional) average displacement; for plasticity,
we might target instantiated components based on geometry (sharp corners) and boundary conditions (clamped).
Nevertheless, these heuristics are not rigorous, and full field evaluation must often be considered. More efficient
evaluation of linearity criteria remains an outstanding challenge.
We next consider the port training. We first note that Method I – model training – requires no modification within
the linear–nonlinear context; we shall exercise this particular approach in several examples in this paper. On the other
hand, Method II – transfer eigenproblem pairwise training – does rely on linearity. For the application of Method II to
the current linear–nonlinear context there are two possibilities. The first possibility is to simply apply the port modes
generated in the linear regime also on the ports on ΓINT . This optimistic approach is not without some justification:
the nonlinear model presumably relaxes to linear form in the vicinity of ΓINT ; hence from evanescent arguments we
might hope that the linear port modes generated for the library might also be relevant on ΓINT . The second possibility
is to include buffer linear components (components which experience linear behavior) in ΩNLIN (µ) which reside
between the nonlinear components (components which in fact experience nonlinear behavior) and the designated
linear–nonlinear interface, ΓINT . Then, the behavior on ΓINT will perforce be included in the port modes since the
linear components on both sides of ΓINT will be included in the training procedure. This buffer approach is rigorous,
however the method requires the incorporation of additional components in ΩNLIN (µ), and hence more degrees of
h
freedom in X NLIN (µ) and commensurately higher computational cost.
4. Nonlinear PDE models

In this section we introduce the different examples of the nonlinear operator g(·, ·; µ) on ΩNLIN (µ) that we shall
consider in Section 5. Note that to simplify notation below we introduce, for w, v ∈ X NLIN h
(µ), glin (w, v; µ) ≡
a(w, v; µ) − f (v; µ): here a and f are defined as in Section 2, (1)–(2), but now for LIN replaced by NLIN ; glin (·, ·; µ)
thus refers to the linear elasticity forms but over ΩNLIN (µ).
4.1. Frictionless contact
Contact analysis is a mature field and many approaches have been studied extensively in the literature; we refer
to the monograph by Wriggers [36] for details. Here we consider the augmented Lagrangian method [37], since
this provides an advantageous combination of accuracy and efficiency. We emphasize, however, that the SCRBE/FE
framework presented in this work would apply equally well to other contact formulations too.
In this case, we suppose that Γc (µ) denotes a pair of contact surfaces in ΩNLIN (µ), and for w, v ∈ X NLIN
h
(µ) we set
∫
g(w, v; µ) ≡ glin (w, v; µ) + (λk (s) + ϵ N h N (s))n im (s)vi (s)ds, (36)
Γc (µ)
where h N denotes the normal-direction “gap function”, ϵ N is the normal-direction contact penalty, λk is the augmented
Lagrangian parameter at step k and at s ∈ Γc , and n m (s) is the normal direction on the master surface. As in [37], h N
is defined as
h N (s) ≡ (u im (s) − u is (s))n im (s), (37)
where u m (respectively, u s ) is the displacement on the master (respectively, slave) surface. This definition implies
that h N > 0 if there is overlap, and h N < 0 if there is a gap. At each augmented Lagrangian iteration we solve the
nonlinear system by Newton’s method, and we then apply the standard first-order update formula for λk (s) for s ∈ Γc ,
λ (s) + ϵ N h N (s), if λk (s) + ϵ N h N (s) > 0
{ k
λk+1 (s) = (38)
0, otherwise.
In our implementation, we impose the augmented Lagrangian contact terms as nodal forces on Γc , as described
in [36] Section 9.3.1. It is often desirable to plot the contact pressure on Γc : the pressure can be reconstructed from the
Lagrange multiplier-cum-nodal force at each node as p N (s) = λ(s)/Anodal (s), where Anodal (s) is the area associated
to the node at s.
We terminate the augmented Lagrangian updates once the maximum gap or overlap on Γc is below a pre-specified
tolerance. Note that Γc in fact denotes a pair of potential contact surfaces and that, upon convergence, the actual
contact surface is extracted as that part of Γc over which λk is non-zero.
4.2. Frictional contact
Next we consider the addition of frictional terms to g based on the classical Coulomb model. This approach is
well-documented in the literature [36,37] and hence we do not reproduce all the details here. The key point is that we
now include a tangential traction σT on the contact surface Γc . This tangential traction imposes the stick/slip behavior
that is characteristic of friction. We let µfr denote the coefficient of friction. “Stick” is obtained if ∥σT ∥ < µfr |σ N |,
where σ N is the normal stress (contact pressure); “slip” is obtained otherwise. In the slip case, the tangential and
normal stresses are related as ∥σT ∥ = µfr |σ N | for σT oriented in the direction opposite to the slip. Note also that,
following [36,37], we make no distinction between the static and kinetic coefficients of friction; the latter is a common
simplifying assumption in engineering practice.
To resolve numerical issues associated with the inherent discontinuity of the Coulomb model, we follow the well-
established regularization approach based on the analogy between friction and plasticity: the stick regime is modeled
as an elastic response through a stiffness coefficient cT ; the slip regime is modeled as a perfectly plastic response [36].
In principle we wish to consider the limit cT → ∞, however in actual practice stability (of the solution strategy)
demands a finite value; we choose a value cT∗ consistent with stability and sufficiently large that we observe relatively
little change in the solution as cT approaches cT∗ .
We combine the normal contact formulation from (36) and the frictional tangential traction described above to
obtain, for w, v ∈ X NLIN
h
(µ),
∫
g(w, v; µ) ≡ glin (w, v; µ) + (λ (s) + ϵ N h N (s))n im (s) + σT,i (s) vi (s)ds.
[ k ]
(39)
Γc (µ)
We apply the same augmented Lagrangian scheme with inner Newton loop as described in Section 4.1; the key
difference in this case is that we also must update within each inner Newton iteration the tangential traction based
on the regularized stick/slip formulation . The normal stress on the contact surface may then be extracted as in the
frictionless case: the Lagrange multiplier-cum-nodal force divided by the associated nodal area. The tangential traction
may be extracted at each contact surface node from the stiffness relation associated with the regularization of the
Coulomb friction model; this approach ensures satisfaction of the Coulomb inequality constraint ∥σT ∥/µfr ≤ |σ N |, as
we shall confirm empirically in the numerical results of Section 5.2.
4.3. Plasticity
Next we introduce our implementation of plasticity. In this case we introduce, for w, v ∈ X NLIN
h
(µ),
∂vi
∫
g(w, v; µ) ≡ (σEP (w))i j d x − f (v; µ), (40)
ΩNLIN (µ) ∂x j
where the “EP” subscript refers to “elastic–plastic”, and σEP is obtained from the radial return algorithm. Once again,
we refer to the extensive literature on this topic [38] rather than reproduce all the details here. Note that the linear
elasticity model over ΩLIN (µ) and g(·, ·; µ) from (40) over ΩNLIN (µ) are “compatible” in the sense that at the interface
between the plastic and linear domains, ΓINT , we correctly recover (weak) continuity of stress; in fact, ΩNLIN (µ) is
chosen such that g(·, ·; µ) reduces to glin (·, ·; µ) in a neighborhood of ΓINT . In Section 5 we only consider perfect
plasticity, though other models (e.g. isotropic and kinematic hardening) could easily be accommodated via the radial
return method.
We utilize the “consistent elastoplastic tangent” [39] when we apply Newton’s method to this nonlinear system
to ensure second-order convergence. We may also incorporate contact terms from Sections 4.1 and 4.2 into (40), if
desired, in order to model plasticity with contact.
Remark 4.1. Friction and plasticity can be path-dependent phenomena. As a result, we perform load-stepping in
order to implement the formulations in Sections 4.2 and 4.3. Also, we neglect inertia effects in this work: the models
we consider are effectively quasi-static.
5. Numerical results
We now illustrate the coupled SCRBE/FE method on some illustrative problems drawn from solid mechanics and
structural analysis. In each model we include both linear and nonlinear regions via the formulation from Section 3. In
the first three examples we demonstrate application of the SCRBE/FE approach to very simple problems in contact
without friction, contact with friction, and finally plasticity. We illustrate in these examples not the flexibility afforded
by the SCRBE method but rather the different aspects of the SCRBE/FE approximation and associated computational
procedure. In particular, the examples include only a single model parameter, and relatively few components. In the
fourth example we consider a much larger parametrized model which includes (implicitly) many model parameters,
many archetype components, and multiple instantiations of several archetype components; we thereby demonstrate
the flexibility of the approach and also the potential for much larger computational savings.
We use the following software libraries to implement the numerical examples shown here: the libMesh [40] finite
element framework, the rbOOmit [16] implementation of the Reduced Basis method, and PETSc [41,42] for linear
and nonlinear algebraic solvers. We use the PETSc interface to the parallel sparse direct solver MUMPS [43,44]
for problems that are sufficiently small, but the PETSc iterative solvers and preconditioners for larger problems.
Each of these libraries is optimized for large-scale parallel computation, which is highly beneficial in order to treat
industrial-scale problems. All computations are performed on 3.60 GHz Intel Core i7–4790 CPUs.
5.1. Frictionless Hertzian contact
We begin with a frictionless contact problem for which there is a closed-form Hertzian solution. Our contact
solver is based on the formulation in Section 4.1. We consider the situation depicted in Fig. 4: a sphere of radius R
subjected to a force F is pressed into contact with a half-space. The model geometry is shown in Fig. 5 (left). We
impose symmetry boundary conditions on the teal surfaces such that we need only consider one quarter of the full
domain of Fig. 4. We clamp the base of the truncated half-space domain, and we impose the force F as a uniform
pressure on the green surface. We also include a “distributed spring” boundary condition (on the green surface) in
order to constrain the sphere component in the vertical direction—without this term the system would be singular;
this displacement-dependent restoring pressure is incorporated in the weak form as a Robin term,
∫
κ (u j n j )(vi n i )ds, (41)
Γ
where κ is a “distributed spring constant” and n is the normal vector on the surface. The constant κ is set to 104 N/m3 ,
which is sufficiently small such that it does not significantly affect the contact pressure, as we verify below in the close
agreement with analytical results. Finally, we impose a frictionless contact condition according to the formulation in
Section 4.1 to prevent surface penetration.
In this example we consider F to be our only parameter, and hence, in the notation from Section 2.2, µ ≡ F;
we consider the parameter domain P = [0, 20] N. We fix the following material and geometric properties:
E 1 = 432 × 109 Pa, E 2 = 72 × 109 Pa, ν1Poisson = ν2Poisson = 0.32, R = 1 mm, where ν1Poisson , E 1 (respectively,
ν2Poisson , E 2 ) are Poisson’s ratio and Young’s modulus for the sphere (respectively, half-space).
For the physical constants prescribed, and the parameter domain chosen for the applied force F, we satisfy the
condition a ≪ R required for small-strain Hertz theory, and we can thus compare our numerical predictions to these
Fig. 4. The frictionless contact test problem.
Fig. 5. The spatial domain (left), component synthesis (middle), and global FE mesh (right) for the contact test problem. Symmetry boundary
conditions are imposed on the teal surfaces, a pressure and distributed spring condition is imposed on the green surface, and the base of the
half-space domain is clamped. The lower part of Γc , the potential contact surface, is shown in magenta (the upper part of Γc is hidden from
view in this figure). The system is assembled from four components: the two small inner components each contain a contact surface, and these
two components – FE components – form ΩNLIN (µ) within the SCRBE/FE formulation of Section 5.1.2; the two outer components – SCRBE
components – form ΩLIN (µ). We refer to the linear/nonlinear interfaces (the visible edges of which are) shown in brown and yellow as ΓINT,1 and
ΓINT,2 , respectively. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
well-established closed-form asymptotic solutions. We recall from√Hertzian contact theory the following results for
1 3
our test problem [45]: the radius of the contact area is given by a = Rd, where d is determined from F = 43 E ∗ R 2 d 2
1 1−ν12 1−ν22 3F
for E∗
= E1
+ E2
; the maximum contact normal stress (pressure) is given by Pmax = 2πa 2
.
5.1.1. Global FE results

We first create a global finite element approximation for the configuration of Fig. 4. The global finite element
approximation shall provide the foundation for the SCRBE/FE approximation developed in Section 5.1.2: (i) the
h
space X NLIN (µ) associated with the full unreduced FE approximation over ΩNLIN (µ), and (ii) the underlying space
h
X LIN (µ) associated with the SCRBE approximation over ΩLIN (µ). The global finite element approximation also
Table 1
Comparison between global FE predictions and closed-form Hertzian re-
sults for F = 10 N.
Quantity Hertzian theory Global FE value % Difference
a 0.04778 mm 0.0479 mm 0.25%
Pmax 2091.4 MPa 2058.85 MPa 0.98%
provides reference values, both for quantities of interest and computational times, with respect to which we can
assess SCRBE/FE performance.
To verify the global FE approximation, we consider the case F = 10 N. In Table 1 we present the contact radius a
and the maximum normal stress (pressure) Pmax . We calculate the pressure as described in Section 4.1, p N = λ/Anodal ,
where Anodal is the area associated to the node; the maximum pressure, Pmax , is realized at the “corner” contact node
(i.e., the contact node that belongs to both symmetry planes) on the half-space. The global FE and analytical results
in Table 1 agree well, which provides confidence that the FE contribution to the SCRBE/FE error, u − u h , is suitably
small.
The global finite element mesh of trilinear elements, shown in Fig. 5 (right), comprises 124 370 nodes and hence
373 110 FE displacement DOFs; in fact, the global FE mesh is stitched together from the component synthesis of
the model to be described shortly. Note that, in general, stress quantities are calculated directly and consistently from
the finite element displacement field, and are thus discontinuous across element boundaries. We adopt the PETSc
implementation of Newton’s method and, within each Newton step, the MUMPS direct solver for solution of the
Jacobian system. Calculation of the global FE solution requires 422 s of wall-clock time on 8 CPU cores.
5.1.2. SCRBE/FE results

We now apply the SCRBE/FE method to this frictionless contact problem. The model in Fig. 5 in fact consists of
four archetype components (each with a single instantiation): ΩNLIN (µ) comprises the two “inner” (FE) components
since these components both contain contact surfaces and hence require the nonlinear model and associated full FE
treatment; ΩLIN (µ) comprises the two “outer” (SCRBE) components in which we consider a linear elasticity model.
The two FE components are shown in greater detail in Fig. 6. Note in this case we anticipate the nonlinear behavior
and hence specify the linear–nonlinear decomposition a priori rather than consider dual-purpose components and
linearity criteria. The interface region ΓINT is composed of two sub-regions, ΓINT,1 and ΓINT,2 , which are shown in
Figs. 5 and 6 as brown and yellow, respectively; each interface sub-region is a single global port shared by a SCRBE
component and a FE component. We note that, as demonstrated in Fig. 6, the SCRBE/FE approach does not require
ΓINT to be a smooth manifold. In this case ΩNLIN (µ) contains 42 769 nodes and 128 307 displacement DOFs, which
is roughly a third of the size of the global FE model from Section 5.1.1.
We prepare the SCRBE data for this model by proceeding with the following steps. We first consider port reduction,
and in particular pursue Method I of Section 2.4.2: we solve the global FE model at 21 different parameter (force)
values in order to extract a set of “empirical port data” on the two fiducial ports which map respectively to ΓINT,1
and ΓINT,2 ; we then apply Proper Orthogonal Decomposition (POD) to compress this data for each fiducial port
into a set of respective port modes. We truncate each POD expansion based on the spectrum and thereby obtain 7
modes on ΓINT,1 , and 11 modes on ΓINT,2 ; the number of global FE DOFs on ΓINT,1 and ΓINT,2 is 5628 (≫ 7) and
6120 (≫ 11), respectively, and thus the port reduction is substantial. We emphasize that the port reduction does not
affect the representation of the solution on the actual contact interface: the contact interface is interior to ΩNLIN (µ)
and hence disjoint from ΓINT,1 and ΓINT,2 . Finally, we perform the offline RB training for each of the two SCRBE
archetype components and the particular port modes identified. In total, the offline stage requires approximately
8500 s to perform the 21 global FE solves and identify the port modes, and approximately 300 s to train each SCRBE
component.
We reconsider the test case F = 10N from Section 5.1.1 but now for the SCRBE/FE approach. Unless otherwise
stated, the number of port modes in the SCRBE/FE approximation is fixed at 7 modes and 11 modes on ΓINT,1 and
ΓINT,2 , respectively. The SCRBE/FE approximation yields Pmax = 2065.61 MPa, which agrees with our global FE
result from Table 1 to within 0.33%. Fig. 7 provides a more detailed comparison of the normal stresses predicted by
the SCRBE/FE and global FE approximations, respectively; we observe good agreement in the spatial distribution
Fig. 6. The two FE components for the frictionless contact model of Fig. 5; these FE components both contain a contact surface. The FE region
(ΩNLIN , which comprises the two FE components indicated) contains 42 769 nodes and 128 307 displacement DOFs.
Fig. 7. The contact normal stress as a function of the radial coordinate, r , for the frictionless contact problem for parameter value F = 10 N. We
observe very good agreement between the global FE and SCRBE/FE solutions. We also include the Hertzian solution as a reference.
of the contact pressure over the contact surface. We recall the normal stress is deduced from the Lagrange multiplier
(normal contact force) and associated nodal area patch on Γc . Finally, in Fig. 8, we present a plot of contact radius as
a function of F, which demonstrates that the SCRBE/FE approximation accurately reproduces the expected behavior
over the full parameter domain. (Note in Fig. 8 we consider 90 parameter values—hence many more than the 21
parameter values associated with the offline port training.) We note that r in Fig. 7 represents the radial coordinate on
the contact surface, whereas a in Fig. 8, an output, represents the deduced radius of the contact region.
We next discuss the online computational efficiency of the SCRBE/FE method. The SCRBE/FE calculation is
effected with the same solvers and the same number of CPU cores as the global FE calculation in Section 5.1.1.
Fig. 8. Contact radius a as a function of parameter F for the SCRBE/FE approximation and the Hertzian closed-form solution. We sample the
SCRBE/FE solution at 90 distinct values of F. The “step” pattern in the SCRBE/FE result arises because the contact radius exhibits discrete
increments each time a new element is incorporated within the contact region; the parameter values at which these increments occur is not related
to the parameter values associated with the offline port training procedure. This plot agrees closely with Figure 6.(b) from [46]. (For interpretation
of the references to color in this figure legend, the reader is referred to the web version of this article.)
However, the SCRBE/FE calculation requires only 70 s of wall-clock time: the SCRBE/FE method thus realizes a
speedup of approximately six relative to the global FE approach. We make several comments. First, the speedup
demonstrated here is superlinear: a reduction to one-third in the number of DOFs yields a speedup of six. This is
to be expected since the computational cost in the SCRBE/FE solver is dominated by the full FE formulation on
ΩNLIN (µ), and the cost of the latter typically scales superlinearly with the number of DOFs. Second, due to the nature
of this Hertzian test problem, the FE region is quite large compared to the overall model, and hence the speedup factor
compared to global FE is relatively modest. In many cases it is realistic to expect much larger DOF reductions, and,
correspondingly, a much larger speedup; we illustrate some examples with larger speedup below. Third, the offline
costs are considerable (preparation of port modes, training of SCRBE/FE components), but easily amortized over
many subsequent queries, for example as illustrated in Fig. 8.
We close with a brief discussion of convergence in M, the number of port modes retained (on each port). In general
we retain at least six port modes – associated with rigid-body modes – on each fiducial port. In our particular example
here, we retain somewhat more than six modes for each fiducial port. To demonstrate the effect of port reduction,
we repeat the normal stress calculations of Fig. 7 for M = 6 modes: we retain, on both ΓINT,1 and ΓINT,2 , only the
respective first six port modes. The results, shown in Fig. 9, indicate that M = 6 is in fact not adequate: the error in
the normal stress is as large as 20%.
5.2. Frictional contact
Next we illustrate that we can incorporate frictional effects into the model from Section 5.1 via the formulation of
Section 4.2. We inherit from the frictionless case the geometry and definition of Γc , the boundary conditions (except
at the contact surface), the value of the distributed spring constant κ, and the material properties (E 1 , E 2 , ν1Poisson ,
ν2Poisson ); we also inherit the four archetype-cum-instantiated components – two FE and two SCRBE – and associated
ports, and hence also the (stitched-together) global FE mesh. In this section we fix F = 10 N and furthermore consider
a fixed load path (in all calculations): the applied force increases from 0 N to 10 N in 10 equal-increment load steps.
The only parameter (given the fixed load path) is the friction coefficient, µ ≡ µfr ; we choose for our parameter domain
P ≡ [0, 0.6].
Fig. 9. The normal stress as a function of the radial coordinate, r , for parameter value F = 10 N, as predicted by the SCRBE/FE approximation
for M = 6, the global FE approximation, and the Hertzian closed-form solution.
Table 2
Relative extent of the stick contact zone, c/a, for F = 10 N and two dif-
ferent values of the friction coefficient: analytical result of Spence [47];
computational result of Goyut et al. [46]; SCRBE/FE result.
µfr c/a
Spence Goyut et al. SCRBE/FE
0.1 0.530 0.690 0.528
0.6 0.99 0.99 0.985
We next develop the SCRBE/FE formulation. We apply the same offline workflow as in the frictionless case,
Section 5.1, to obtain the ΓINT port modes, except now we sample different values of µfr ∈ P (for the fixed final
load F = 10 N); this process provides 18 interface port modes in total, 7 modes on ΓINT,1 and 11 modes on ΓINT,2 .
Note we obtain the same number of port modes for frictionless contact, Section 5.1, and frictional contact, in the
current section, and indeed the respective modes are almost indistinguishable; the latter in fact could be anticipated
from evanescence arguments. Finally, we proceed to perform the offline RB training for the two SCRBE components
and associated port modes. We note that in the port training procedure we retain snapshots only at the final force of
the load path, F = 10 N, and hence we cannot in principle interpret F as a second parameter; however, given that
the frictionless and frictional port modes are very similar, and that F appears linearly in the SCRBE components, we
conjecture that we could a posteriori declare F a second parameter in domain [0, 10] N.
We demonstrate in Fig. 10, for µfr = 0.1, the close agreement between the global FE solution and the SCRBE/FE
solution over the entire domain. Next, and more quantitatively, we compare our SCRBE/FE solutions to the results
provided in Goyut et al. [46] and Spence [47]; note that Goyut et al. uses a coupled boundary element/finite element
formulation, whereas Spence pursues an analytical approach. Table 2 presents a comparison of the results for the ratio
c/a, where c is the radius of the “stick contact zone” and a is the full contact radius; the SCRBE/FE predictions agree
particularly well with the theory of Spence.
Finally, we present in Fig. 11 plots of contact normal stress (σ N ), as well as the ratio of tangential stress to friction
coefficient (σT /µfr ), at µfr = 0.1 and µfr = 0.6. Our results agree reasonably well with Fig. 8 of Goyut et al., and
we further note that the inequality ∥σT ∥/µfr ≤ |σ N | is always satisfied, as expected for the Coulomb friction model.
We cannot provide any supportable explanation for the non-smooth contact tangential stress. However, we emphasize
that (i) all other aspects of the SCRBE/FE solution (inspected), including the extent of the stick zone and the contact
normal stress, are accurate, and (ii) the SCRBE/FE and global FE solutions agree well, and hence the oscillations are
Fig. 10. The zz-component of the stress tensor for the global FE solution (left) and the SCRBE/FE solution (right) for the frictional contact problem
for parameter value µfr = 0.1. Note that we use the same colorbar range in both cases (based on the min/max from the global FE solution) for
purposes of easier comparison. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this
article.)
Table 3
Comparison of solution times for frictional contact. We use 10 load steps
for both the global FE and SCRBE/FE approximation. Note that for the
SCRBE/FE approximation we report only the online time.
µfr Global FE time (s) SCRBE/FE online time (s) Speedup
0.1 4138 610 6.8
0.6 3280 524 6.3
not an artifact of port or bubble reduction. Inasmuch, we anticipate that the non-smoothness is a numerical artifact of
the penalty-based regularized friction formulation; we could consider alternative frictional contact implementations
to mitigate his effect, but we do not pursue these here. Finally, we note that the maximum stresses in Fig. 11 would
certainly exceed the yield stress of common materials (e.g., steel), and hence we would need to consider plasticity—
which is the subject of the next section.
To close, we discuss the computational performance. The FE and SCRBE/FE solution times for this model, for
two different values of µfr , are presented in Table 3; the speedup observed in the frictional case is comparable to the
speedup observed in the frictionless case of Section 5.1.2.
5.3. Frictionless contact with plasticity
As a third test case, we apply the SCRBE/FE method to a test problem which incorporates both frictionless contact
and plasticity, and thus exercises the FE formulations of both Sections 4.1 and 4.3. The test problem is provided in
Vu-Quoc et al. [48]: elastoplastic analysis of two spheres (of the same material) of radius R pressed into contact; the
problem is depicted in Fig. 12 (left). We adapt the boundary conditions of Section 5.1 for elastic contact to the current
elastoplastic context: on the top (green) surface we impose force F and distributed spring with constant κ = 50 N/m3 ;
on the bottom surface we impose zero displacement (clamped). We consider as parameter a load path {Fℓ }ℓ=0,...,L
(F0 = 0) which takes on values for the force F· in the domain [0, 1500] N; note that our parameter is, in principle, an
L-tuple, in which we may vary L and both the load path shape and magnitudes. As we discuss further below, we shall
train the SCRBE approximation on a particular “training load path”: a linear increase in the force – loading – from
F0 = 0 N to F40 = 1500 N, followed by a linear decrease – unloading – from F40 = 1500 N to F80 = 0 N; the
training load path thus corresponds to L = 80 load steps in total. Finally, in accordance with [48], we set Young’s
modulus to 70 GPa, Poisson’s ratio to 0.3, and sphere radius R = 0.1; we consider perfect plasticity with a yield stress
of 100 MPa.
Fig. 11. SCRBE/FE solution for the contact normal stress (σ N ) and contact tangential stress divided by friction coefficient (σT /µfr ) as a function
of radial coordinate r for µfr = 0.1 (top) and µfr = 0.6 (bottom).
The domain, shown in Fig. 12 (left), comprises four archetype-cum-instantiated components, shown in Fig. 12
(middle): two “inner” FE components which constitute ΩNLIN (µ) and accommodate the (nonlinear) contact and plastic
behavior; two “outer” SCRBE components (identical in geometry) in which we consider a linear elasticity model. We
stitch together the component meshes to create the global FE mesh shown in Fig. 12 (right); the associated global
FE approximation space is of dimension 436 386. We next develop the SCRBE/FE approximation following the same
procedures as in Sections 5.1 and 5.2. We first obtain the respective port modes for the ports associated with each
component: we find the global FE solution for our single (L = 80-step) training load path and perform a POD on
each set of port data to identify 7 port modes for the upper component port and 9 port modes for the lower component
port; note that although the geometry is symmetric in z, the boundary conditions on the top and bottom surfaces
are different. We then perform the offline RB training for each of the two SCRBE archetype components and the
corresponding port modes. In the end we obtain an SCRBE/FE model with 247686 FE dofs (over ΩNLIN (µ)) and our
16 SCRBE port modes over ΩLIN .
Fig. 12. The domain (left), component synthesis (middle), and global FE mesh (right) for the two-spheres contact test problem. We use the
same color scheme as in Fig. 5 to indicate symmetry boundary conditions, loads, contact surface(s) Γc , and linear–nonlinear interface ΓINT . (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
Fig. 13. Comparison of σzz over ΩNLIN (µ) for the global FE (left) and SCRBE/FE (right) approximations for the test problem of [48]. Note
that we use the same colorbar range in both cases (based on the min/max from the global FE solution) for purposes of easier comparison. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
We now consider as our first test case the particular training load path with which we train our port modes and
subsequently develop our SCRBE/FE approximation. In fact, we consider just the loading phase of the load path; we
will address unloading in a second test case. In Fig. 13 we demonstrate the good agreement between the global FE
and SCRBE/FE solutions for this test problem. In Figs. 14 and 15, we further compare our global FE and SCRBE/FE
predictions to the reference results from Figures 6 and 8 of [48]: in Fig. 14 we consider the normal stress, calculated
from the contact nodal forces, as a function of radial coordinate; in Fig. 15 we present the area of the contact radius
as a function of load step force. In both cases we observe good agreement with the reference results.
We next demonstrate the capability to consider load paths other than the training load path. In particular we present
in Fig. 16 the normal displacement – as defined in [48] – as a function of force for the load path ADE of Figure 14
of [48]. Note that now we consider both loading and then subsequent unloading, and furthermore for a maximum
Fig. 14. Contact normal stress as a function of radial coordinate for the two-spheres test problem with frictionless contact and plasticity. We
compare our global FE and SCRBE/FE solutions to the numerical results of Figure 8 from [48]. We also superimpose the elastic Hertzian solution
as a reference.
Fig. 15. Contact area as a function of F (within our single load path from 0 N to 1500 N) for the two-spheres test model with frictionless contact
and plasticity. We observe close agreement with the results of Figure 6 from [48]. We also superimpose the elastic Hertzian solution as a reference.
force F = 1000 N which is different from the maximum force F = 1500 N of the training load path. We observe,
during both the loading and unloading phases, good agreement between our global FE and SCRBE/FE predictions
and also the FE results of [48].
Finally, we discuss the computational performance. For both the global FE and SCRBE/FE calculations we adopt
the PETSc conjugate gradient solver on 8 cores. The solution times for the first test case – load path F0 = 0 N to
F40 = 1500 N – are 6362 s and 1567 s for the global FE and SCRBE/FE calculations, respectively. We observe the
superlinear speedup with respect to DOF reduction, as already noted in Section 5.1.2; the latter ensures a substantial
speedup – here a factor of four – even for small problems.
Fig. 16. Normal displacement for a loading/unloading path for the two-spheres test model with frictionless contact and plasticity. We observe close
agreement between our SCRBE/FE predictions and the results from Figure 14 of [48].
Fig. 17. Equivalent plastic strain over ΩNLIN (µ) for the SCRBE/FE solution of the two-spheres test problem. (Note the plot domain is a subset
of the two FE components.) The plastic strain decays to zero well before the boundary of ΩNLIN (µ), and hence the nonlinear behavior is indeed
contained within ΩNLIN (µ), as required.
We close with an important comment regarding the size of the full FE region, ΩNLIN (µ), in the SCRBE/FE
approximation: ΩNLIN (µ) must be large enough to fully encapsulate the region of plastic behavior. We demonstrate in
Fig. 17 that the equivalent plastic strain [38] in ΩNLIN (µ) goes to zero well within the interior of the full FE region, as
required. If the latter condition were not satisfied then we would need to enlarge the full FE region, ΩNLIN (µ), which
of course would have a deleterious effect on the speedup provided by the SCRBE/FE formulation. In fact, for the
current example, we could perhaps have considered smaller FE components, which would in turn have decreased the
size of ΩNLIN (µ) and improved computational performance accordingly.
5.4. Local plasticity in a large model
The examples in Sections 5.1, 5.2, and 5.3 provide a detailed illustration of the accuracy of the SCRBE/FE
approximation and our ability to treat a range of nonlinear physical phenomena. However, these examples do not
illustrate major speedup compared to conventional (global) FE treatment because the number of DOFs in ΩNLIN (µ) is
Fig. 18. The shiploader structure of Section 5.4: the full shiploader is shown on the left, and the “shuttle” subregion, which supports a conveyor
belt, is shown on the right.
Fig. 19. Several instantiations of two distinct archetype components. Both archetype components have parametrized length, density, and surface
loads, and moreover the left archetype component has parametrized curvature, and the right archetype component has parametrized defect location.
The demarcations in the component on the right indicate regions that are used to define the component’s piecewise mapping function.
a large fraction of the total number of DOFs in the global FE mesh. In this section we present a model of industrial-
scale infrastructure which demonstrates the speedup that may be achieved with the SCRBE/FE approach in larger-
scale problems.
The model we consider is a shiploader structure, illustrated in Fig. 18. A shiploader is a machine used at ports to
load cargo (e.g., ore) via a conveyor belt – supported by a “shuttle” sub-assembly – onto ships. We impose only a
“self-weight” load; we also clamp the structures at several points of rigid connection. The component description of
the full shiploader model comprises n inst = 182 instantiated components and n glob = 317 global ports synthesized
from a library with n arch = 57 archetype components and n fid = 33 fiducial ports. The component meshes use
tetrahedral elements with second-order basis functions and the resulting global FE approximation, which we shall
never construct, has 18903687 DOFs in total.
The shiploader model has hundreds of model parameters which can be modified, including geometry (location of
defects, angle of the shuttle), material properties (density, Young’s modulus, Poisson’s ratio), and loads. In Fig. 19 we
present some examples of the types of parameters that are incorporated in the shiploader archetype components. In
particular, we show several instantiations of two archetype components: each archetype component has parametrized
length, density, and surface loads; the archetype component shown on the left (respectively, right) in the figure
additionally has parametrized curvature (respectively, defect location). These components can be instantiated within
the shuttle with different parameter values in order to realize a system associated to any particular desired model
parameter value. We show in Fig. 20 two alternative configurations of the full shiploader model in which the shuttle
angle – a model parameter – is varied; the instantiated components which have different parameters relative to the
model in Fig. 18 are highlighted.
In the interest of brevity, we consider here only one model parameter value – corresponding to the configurations
indicated in Fig. 18 for shiploader and shuttle – in the remainder of this section. We begin here with a validation
Fig. 20. Two variations in the model parameter of the full shiploader model of Fig. 18 in which the shuttle is at an angle of (left) −10◦ and (right)
10◦ from horizontal. The components that are parametrically modified compared to Fig. 18 are highlighted in yellow. (For interpretation of the
references to color in this figure legend, the reader is referred to the web version of this article.)
Fig. 21. Von Mises stress for the SCRBE/FE solution for the shiploader shuttle; we consider linear elasticity over the entire domain Ω (µ). In the
right figure we focus on small regions (shown in pink) near the clamped support at which the stress exceeds the yield stress. (For interpretation of
the references to color in this figure legend, the reader is referred to the web version of this article.)
study in which we compare the global FE and SCRBE/FE predictions. For this comparison we must restrict attention
to a subregion of the shiploader, since the global FE solution for the entire shiploader model would constitute a very
large-scale computation not practical within the scope of the present work. In particular, we shall consider the “shuttle”
subregion shown in Fig. 18 (right), which is a sufficiently small model such that we may indeed obtain a global FE
solution. The shuttle model is constructed from the Shiploader library introduced above: the system consists of
n inst = 47 instantiated components and n glob = 70 global ports synthesized from n arch = 21 archetype components
and n fid = 16 fiducial ports.
We first consider the shuttle problem for a linear elasticity model, and associated SCRBE/FE approximation, over
the entire domain, Ω (µ) ≡ ΩLIN (µ). We obtain the Von Mises stress of Fig. 21, in which we observe small regions
near the shuttle clamped support points at which the stress is higher than 300 MPa, a typical yield stress of steel. This
motivates us to re-consider this problem but now introducing a FE plasticity zone, a region ΩNLIN (µ), in the vicinity
of the clamped supports shown in Fig. 21. Towards that end, we next consider two SCRBE/FE approximations which
incorporate plasticity, in particular perfect plasticity with yield stress of 300 MPa, in the region identified in Fig. 21.
The two options are shown in Fig. 22. The first configuration, Case I, contains a minimal choice of ΩNLIN (µ), and the
second configuration, Case II, includes a “buffer” linear zone within ΩNLIN (µ) as described in Section 3.2.2.
We compare four displacement fields in Fig. 23. The first considers linear elasticity over the entire domain,
SCRBE/FE approximation; the remaining three consider elastoplasticity, SCRBE/FE approximations for Case I and
Case II, and global FE approximation. We observe very good agreement between the three elastoplastic solutions.
We note that the maximum displacement (at the tip of the shuttle) predicted by the linear elasticity model over
the entire domain, the (left, top) panel of Fig. 23, differs by 0.23% from the maximum displacement predicted by
the elastoplasticity models; furthermore, and as expected, the latter is more negative than the former. On the other
hand, the differences between the SCRBE/FE predictions and the global FE prediction for the maximum elastoplastic
displacement differ by only 0.025%. We could thus reasonably conclude that our SCRBE/FE approximation accurately
captures the elastoplastic effect on the (or an) engineering quantity of interest. In this particular case the effect is not
Fig. 22. Two SCRBE/FE configurations with different choices of ΩNLIN (µ), highlighted in yellow: Case I (left) with a minimal choice of ΩNLIN (µ),
and Case II (right) in which ΩNLIN (µ) is chosen to include a “buffer” linear zone. (For interpretation of the references to color in this figure legend,
the reader is referred to the web version of this article.)
Table 4
Number of DOFs and solution times for the three shuttle approximations. Solution times are obtained for 8 CPU cores and the MUMPS direct
solver; note that for SCRBE/FE we report only the online time.
Model FE DOFs SCRBE DOFs Solution time (s) Speedup vs. Global FE
Case I 24 882 4102 25.3 93.2
Case II 258 777 4038 242.3 9.73
Global FE 2 203 230 0 2358.4 1
too large – on the order of 2 mm – and perhaps not overly significant given other uncertainties in the model and
operating environment. Nevertheless, the example provides proof of concept.
Fig. 24 demonstrates that the equivalent plastic strain clearly decays to zero within ΩNLIN (µ), which in turn justifies
our linear–nonlinear SCRBE/FE spatial decomposition. In this sense, our hybrid approximation is self-consistent
and in particular there is no need to adapt ΩLIN (µ) and ΩNLIN (µ). We note, however, that the shiploader exhibits
many re-entrant corners at which we might expect an unbounded stress, and hence a stress above yield, as the finite
element mesh is increasingly refined. Arguably the plastic regions identified first, on coarser meshes, will ultimately
correspond to the largest plastic regions, with largest effect on engineering outputs. This argument is important to
computational cost since typically we will not have the resources, even within the SCRBE/FE context, to refine the
mesh locally – and introduce plastic regions – at all corners in a large-scale structure.
We now discuss the computational aspects of the global FE approach and SCRBE/FE method for this shuttle
test case. As regards the former, the global FE approximation comprises 2203230 DOFs. As regards the latter, we
effect in this example port reduction through Method II [13] described in Section 2.4. For port reduction Method II,
unlike port reduction Method I applied in the earlier examples, we do not require any global FE solutions to generate
the empirical data; furthermore, as described in Section 3.2.2, if we include a buffer region in ΩNLIN (µ) then the
port space is system- and model-independent even in the presence of a plastic region. The procedure yields, for our
shuttle library of archetype components and associated fiducial ports, 4102 SCRBE port modes. The number of FE
DOFs and the solution times for the three elastoplastic models (SCRBE/FE Cases I and II, and global FE) are given in
Table 4. We note from Table 4 that Case II – with buffer components – is (unsurprisingly) slower than Case I—without
buffer components. However, Case II offers the advantage that the port modes are more rigorously justified. In actual
practice, an engineer might design based on Case I but then for final confirmations resort to Case II; the latter is
still much less expensive than global FE (which is furthermore infeasible for the full shiploader). For our particular
situation it appears that Case I is indeed sufficient. We also note that the time required (on 8 CPU cores) to identify
the plastic zone as in Fig. 21 is roughly 15 s, and hence smaller than the solution time of 25.3 s. More generally, the
computation of the maximum stress is linear in the number of global FE DOFs, and furthermore easily parallelized;
in many cases, search heuristics will not be needed. (We recall that search heuristics identify a subset of instantiated
components for evaluation of linearity criteria; for example, we may evaluate maximum stress relative to yield stress
only in components with re-entrant corners and clamped boundary conditions.)
Finally, we consider the full shiploader model. In this case we report results only for the SCRBE/FE method; the
global FE solution – 18903687 DOFs in total – is prohibitively expensive. Note that although we cannot obtain a
Fig. 23. Comparison of the displacement fields for the shiploader shuttle model: (left top) elasticity, SCRBE/FE approximation, over the entire
domain; (right top) elastoplasticity, SCRBE/FE approximation Case I; (left lower) elastoplasticity, SCRBE/FE approximation Case II; and (right
bottom) elastoplasticity, global FE approximation. Note left top assumes elasticity over the entire shuttle domain, whereas the other three models
incorporate local plasticity in the region identified in Fig. 21. The undeformed structure is indicated transparently in the figures, and we apply a
displacement scaling factor of 5 for better visualization.
Fig. 24. Equivalent plastic strain for the SCRBE/FE Case I solution of the shiploader shuttle model. The plastic equivalent strain is plotted in
ΩNLIN (µ), and the rest of the domain is shown transparently.
global FE solution, and we never even construct the entire global FE mesh, the global FE space nevertheless exists:
the global FE space provides the full FE approximation over ΩNLIN (µ), and the underlying approximation space over
ΩLIN (µ). Port modes are again obtained via port reduction Method II, which in this case leads to 13793 modes in the
SCRBE region. We consider Case I – the smaller region ΩNLIN (µ) – as shown in Fig. 25, and hence again we obtain
24882 DOFs in ΩNLIN (µ). We use 8 CPU cores and the MUMPS direct solver to obtain the solution in 57.3 s, which
constitutes only a modest increase in solution time compared to the SCRBE/FE solution time for the much smaller
shuttle model. In contrast, the global FE model, already slow for the shuttle model (over 2000 s), is now infeasible for
the full shiploader, at least for the computational resources available in this work.
We note that in engineering practice it is common to develop a localized model of a high-stress region (such as
the plasticity region considered here), and to impose boundary conditions on the localized region to represent the
effect of the larger structure. This workflow recognizes that conventional FE high fidelity models are typically very
computationally intensive and can thus be invoked only sparingly. This local modeling approach does have significant
drawbacks: decoupling a local region from the rest of the model introduces additional errors into the analysis, and in
particular it can be difficult to ensure that the boundary conditions imposed on the local model are physically realistic.
We propose that the SCRBE/FE approach, as illustrated by the shiploader examples in this section, offers an efficient
alternative for analysis of large-scale structures, and in particular overcomes the main drawback – unidirectional
coupling – of the conventional localized modeling workflow. (Note our emphasis here is on structures in operation, as
opposed to forming processes in which the nonlinear behavior is often more global.)
Fig. 25. SCRBE/FE approximation of the entire shiploader using the Case I ΩNLIN (µ) region (depicted in yellow) from above. (For interpretation
of the references to color in this figure legend, the reader is referred to the web version of this article.)
6. Conclusions
In this work we propose a novel approach to combine component-based model reduction via the SCRBE
formulation with conventional FE. This leads to a hybrid formulation, the SCRBE/FE method, which incorporates
attractive features of both approximation methods: SCRBE/FE brings the ability to develop fast parametrized models
of large systems, and FE brings the ability to incorporate the full range of nonlinear physics. The formulation is fully
conforming on the interface of the SCRBE/FE and FE regions, which leads to a numerically robust method. Moreover,
the formulation leads to a convenient implementation in which standard SCRBE and FE matrix and vector assembly
routines can be fully reused.
We apply the method to several illustrative problems in Section 5 which demonstrate that the method can accurately
reproduce full nonlinear FE results at a fraction of the overall cost—under the condition that the nonlinearities are
localized. The computational speedup that the method provides is of course strongly dependent on the size of ΩNLIN (µ)
relative to the entire domain. We demonstrate test cases (e.g. the Hertzian contact models) in which the FE region is
a relatively large fraction of the overall model, and for which we obtain modest speedup. We also demonstrate cases
in Section 5.4 with a much greater DOF reduction and a correspondingly larger speedup. We believe the latter case is
highly relevant in engineering practice as it is frequently necessary to incorporate localized nonlinear behavior within
a large model in particular for localized contact or failure analysis.
Overall, we believe that the SCRBE/FE method presents an attractive middle ground between the two extremes of
global FE and global model reduction, and it provides a set of capabilities that fit well with practical workflows that
are required for industrial-scale engineering analysis.
Acknowledgments
Jonas Ballani was supported by the Swiss Confederations Innovation Promotion Agency (CTI) under Grant No.
17802.1 PFIW-IW. Anthony Patera was supported by an MIT Ford Professorship. We would like to thank Sylvain
Vallaghé, Brian Sabbey, and Alejandro Echeverria of Akselos S.A., Prof. Marco Picasso of EPFL, and Prof. Kathrin
Smetana of University of Twente, for many fruitful discussions.
References
[1] A. Noor, J. Peters, Reduced basis technique for nonlinear analysis of structures, AIAA J. 18 (4) (1980) 455–462.
[2] B. Almroth, P. Stern, F. Brogan, Automatic choice of global shape functions in structural analysis, AIAA J. 16 (5) (1978) 525–528.
[3] T. Porsching, Estimation of the error in the reduced basis method solution of nonlinear equations, Math. Comp. 45 (1985) 487–496.
[4] G. Rozza, D. Huynh, A. Patera, Reduced basis approximation and a posteriori error estimation for affinely parametrized elliptic coercive
partial differential equations, ARCME 15 (3) (2008) 229–275.
[5] H. Ly, H. Tran, Modeling and control of physical processes using proper orthgonal decomposition, Math. Comput. Modelling 33 (2001)
223–236.
[6] P. Chinesta, P. Ladeveze, E. Cueto, A short review on model order reduction based on proper generalized decomposition, Arch. Comput.
Methods Eng. 18 (2011) 395–404.
[7] P. Benner, S. Gugercin, K. Willcox, A survey of projection-based model reduction methods for parametric dynamical systems, SIAM Rev.
57 (4) (2015) 483–531.
[8] K. Veroy, C. Prud’homme, D. Rovas, A. Patera, A posteriori error bounds for reduced-basis approximation of parametrized noncoercive and
nonlinear elliptic partial differential equations, AIAA Paper No. 2003-3847, 2003, pp. 1–18.
[9] P. Binev, A. Cohen, R. Dahmen, G. Petrova, P. Wojtaszczyk, Convergence rates for greedy algorithms in reduced basis methods, SIAM J.
Math. Anal. 43 (3) (2011) 1457–1472.
[10] D. Huynh, D. Knezevic, A. Patera, A static condensation reduced basis element method: Approximation and a posteriori error estimation,
M2AN Math. Model. Numer. Anal. 47 (1) (2013) 213–251.
[11] J.L. Eftang, D.B.P. Huynh, D.J. Knezevic, E.M. Rønquist, A.T. Patera, Adaptive port reduction in static condensation, in: 7th Vienna
International Conference on Mathematical Modelling, MATHMOD, 2012.
[12] J. Eftang, A. Patera, Port reduction in component-based static condensation for parametrized problems: Approximation and a posteriori error
estimation, IJNME 96 (5) (2013) 269–302.
[13] K. Smetana, A. Patera, Optimal local approximation spaces for component-based static condensation procedures, SIAM J. Sci. Comput. 38 (5)
(2016) A3318–A3356.
[14] E. Wilson, The static condensation algorithm, IJNME 8 (1) (1974) 198–203.
[15] M. Barrault, Y. Maday, N. Nguyen, A. Patera, An empirical interpolation method: Application to efficient reduced-basis discretization of
partial differential equations, C.R. Acad. Sci. Paris, Sér. I 339 (2004) 667–672.
[16] D.J. Knezevic, J.W. Peterson, A high-performance implementation of the certified reduced basis method, Comput. Methods Appl. Mech.
Engrg. 200 (13–16) (2011) 1455–1466.
[17] W.C. Hurty, On the dynamic analysis of structural systems using component modes, in: First AIAA Annual Meeting, Washington, DC, AIAA
Paper no. 64-487, June 29–July 2, 1964.
[18] R. Craig, M. Bampton, Coupling of substructures for dynamic analyses, AIAA J. 3 (4) (1968) 678–685.
[19] U. Hetmaniuk, R. Lehoucq, A special finite element method based on component mode synthesis, M2AN Math. Model. Numer. Anal. 44 (3)
(2010) 401–421.
[20] Y. Maday, E. Rønquist, The reduced basis element method: Application to a thermal fin problem, SISC 26 (1) (2004) 240–258.
[21] L. Iapichino, A. Quarteroni, G. Rozza, A reduced basis hybrid method for the coupling of parametrized domains represented by fluidic
networks, Comput. Methods Appl. Mech. Engrg. 221–222 (2012) 63–82.
[22] L. Iapichino, A. Quarteroni, G. Rozza, Reduced basis method and domain decomposition for elliptic problems in networks and complex
parametrized geometries, Comput. Math. Appl. 71 (1) (2016) 408–430.
[23] N. Nguyen, A multiscale reduced-basis method for parametrized elliptic partial differential equations with multiple scales, J. Comput. Phys.
227 (2007) 9807–9822.
[24] H. Antil, M. Heinkenschloss, R.H.W. Hoppe, D.C. Sorensen, Domain decomposition and model reduction for the numerical solution of PDE
constrained optimization problems with localized optimization variables, Comput. Vis. Sci. 13 (6) (2010) 249–264.
[25] H. Antil, M. Heinkenschloss, R.H.W. Hoppe, Domain decomposition and balanced truncation model reduction for shape optimization of the
Stokes system, Optim. Methods Softw. (2010).
[26] K. Veroy, A.T. Patera, Certified real-time solution of the parametrized steady incompressible Navier-Stokes equations: Rigorous reduced-basis
a posteriori error bounds, Internat. J. Numer. Methods Fluids 47 (2005) 773–788.
[27] M.A. Grepl, Y. Maday, N.C. Nguyen, A.T. Patera, Efficient reduced-basis treatment of nonaffine and nonlinear partial differential equations,
M2AN Math. Model. Numer. Anal. 41 (3) (2007) 575–605.
[28] B. Haasdonk, J. Salomon, B. Wohlmuth, Reduced basis method for parametrized variational inequalities, SIAM J. Numer. Anal. 50 (2012)
2656–2676.
[29] Z. Zhang, E. Bader, K. Veroy, A slack approach to reduced-basis approximation and error estimation for variational inequalities, C. R. Math.
354 (2016) 283–289.
[30] J. Munjal, Acoustics of Ducts and Mufflers, second ed., Wiley, 2014.
[31] A. Pinkus, N-widths in Approximation Theory, Springer Science and Business Media, 1985.
[32] I. Babuska, R. Lipton, Optimal local approximation spaces for generalized finite element methods with application to multiscale problems,
Multiscale Model. Simul. 9 (2011) 373–406.
[33] A. Quarteroni, F. Pasquarelli, A. Valli, Heterogeneous domain decomposition principles, algorithms, applications, in: D.E. Keyes, T.F. Chan,
G.A. Meurant, J.S. Scroggs, R.G. Voigt (Eds.), Fifth International Symposium on Domain Decomposition Methods for Partial Differential
Equations, SIAM, 1992, pp. 129–150.
[34] Ansys Manual, Advanced Analysis Techniques, Chapter 5, Substructuring.
[35] M. Ohlberger, F. Schindler, Error control for the localized reduced basis multi-scale method with adpative on-line enrichment, SIAM J. Sci.
Comput. 37 (6) (2015) A2865–A2895.
[36] P. Wriggers, Computational Contact Mechanics, second ed., Springer, 2006.
[37] J.C. Simo, T.A. Laursen, An augmented Lagrangian treatment of contact problems involving friction, Comput. Struct. 42 (1) (1992) 97–116.
[38] J.C. Simo, T.J.R. Hughes, Computational Contact Mechanics, Springer, 1998.
[39] J.C. Simo, R.L. Taylor, Consistent tangent operators for rate independent elastoplasticity, Comput. Methods Appl. Mech. Engrg. (1985)
101–118.
[40] B.S. Kirk, J.W. Peterson, R.M. Stogner, G.F. Carey, libMesh: A C++ library for parallel adaptive mesh refinement/coarsening simulations,
Eng. Comput. 23 (3–4) (2006) 237–254.
[41] S. Balay, S. Abhyankar, M.F. Adams, J. Brown, P. Brune, K. Buschelman, L. Dalcin, V. Eijkhout, W.D. Gropp, D. Kaushik, M.G. Knepley,
L.C. McInnes, K. Rupp, B.F. Smith, S. Zampini, H. Zhang, H. Zhang, PETSc users manual, Tech. Rep. ANL-95/11 - Revision 3.7, Argonne
National Laboratory, 2016.
[42] S. Balay, W.D. Gropp, L.C. McInnes, B.F. Smith, Efficient management of parallelism in object oriented numerical software libraries,
in: E. Arge, A.M. Bruaset, H.P. Langtangen (Eds.), Modern Software Tools in Scientific Computing, Birkhäuser Press, 1997, pp. 163–202.
[43] P.R. Amestoy, I.S. Duff, J.-Y. L’Excellent, J. Koster, A fully asynchronous multifrontal solver using distributed dynamic scheduling, SIAM
J. Matrix Anal. Appl. 23 (1) (2001) 15–41.
[44] P.R. Amestoy, A. Guermouche, J.-Y. L’Excellent, S. Pralet, Hybrid scheduling for the parallel solution of linear systems, Parallel Comput.
32 (2) (2006) 136–156.
[45] K.L. Johnson, Contact Mechanics, Cambridge University Press, 1985.
[46] N. Guyot, F. Kosior, G. Maurice, Coupling of finite elements and boundary elements methods for study of the frictional contact problem,
Comput. Methods Appl. Mech. Engrg. 181 (2000) 147–159.
[47] D.A. Spence, The Hertz contact problem with finite friction, J. Elasticity 5 (1975) 297–319.
[48] L. Vu-Quoc, X. Zhang, L. Lesburg, A normal force-displacement model for contacting spheres accounting for plastic deformation: Force-
Driven formulation, J. Appl. Mech. 67 (2) (1999) 363–371.

A Component-Based Hybrid Reduced Basis/finite Element Method For Solid Mechanics With Local Nonlinearities

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

A Component-Based Hybrid Reduced Basis/finite Element Method For Solid Mechanics With Local Nonlinearities

Hochgeladen von

Copyright:

Verfügbare Formate

Available online at www.sciencedirect.

Comput. Methods Appl. Mech. Engrg. 329 (2018) 498–531

A component-based hybrid reduced basis/finite element method for

We refer to the above as single-domain RB shortcomings.

2. The SCRBE approximation

2.1. Expositional problem statement

a(u LIN (µ), vLIN ; µ) − f (vLIN ; µ) = 0, ∀vLIN ∈ X LIN (µ). (2)

2.2. Components and systems

We shall elaborate further on these ingredients in the development below.

2.3. Finite element approximation and static condensation

2.4. Reduced order approximation

2.4.2. Port reduction

2.4.3. Computational procedure

3. Two-domain formulation of linear–nonlinear model

3.2. Hybrid method: SCRBE/FE

X (µ) | vLIN ∈ X LIN

for ξi, j (µ), 1 ≤ i ≤ NNLIN h

C T R H,+ (U H (µ); µ) = 0, (30)

J H (µ) ≡ C T J H,+ (µ) C (33)

3.2.2. Coupling considerations

4. Nonlinear PDE models

4.1. Frictionless contact

4.2. Frictional contact

5.1. Frictionless Hertzian contact

Fig. 4. The frictionless contact test problem.

5.1.1. Global FE results

5.1.2. SCRBE/FE results

5.2. Frictional contact

5.3. Frictionless contact with plasticity

5.4. Local plasticity in a large model

Das könnte Ihnen auch gefallen