Data-driven reduced order modeling for time-dependent problems

Mengwu Guo∗, Jan S. Hesthaven

Chair of Computational Mathematics and Simulation Science,
École Polytechnique Fédérale de Lausanne, 1015 Lausanne, Switzerland

Abstract
A data-driven reduced basis (RB) method is proposed for parametrized time-dependent problems. This
method requires the offline preparation of a database comprising the full-order solutions of time history at
parameter locations. Based on the full-order data, a reduced basis is constructed by the proper orthogonal
decomposition (POD), and the maps between the time-parameter values and the projection coefficients onto
the RB are approximated as a regression model. With a natural tensor grid between the inputs of time
and parameters in the database, the singular-value decomposition (SVD) is used to extract the principal
components in the data of projection coefficients, and the regression functions are represented as the linear
combinations of several tensor products of two Gaussian processes, one of time and the other of parameters.
During the online stage, the solutions at new time-parameter locations in the considered domain can be
recovered rapidly via direct outputs from the regression models. Featuring a non-intrusive nature and the
complete decoupling of the offline and online stages, the proposed approach provides a reliable and efficient
tool for solving parametrized time-dependent problems, and its effectiveness is illustrated by non-trivial
numerical examples.
Keywords: Data-driven, non-intrusive reduced order modeling, time-dependent problem, proper
orthogonal decomposition, Gaussian process regression, tensor decomposition, machine learning

1. Introduction

In science and engineering, many time-dependent problems are described as parametrized partial differ-
ential equations [16, 28], in which the parameters characterize material properties, source terms, underlying
geometry, boundary or initial conditions, and so on. In the context of multi-entry or real-time analysis, it is
often required to solve the model for a large number of different parameter values. The great demands on
both CPU time and memory make the high-fidelity simulations too expensive to allow repeated solutions of
the time history for varying parameters. On the other hand, the repeated high-fidelity evaluations are often
not really necessary as intrinsic similarities among the parameter-dependent solutions can be exploited to
recover the solutions for new parameter values.
During the past decades, reduced order modeling (ROM) has been developed to address this issue. The
key idea of the ROM is to replaced the original full-order model with a reduced-order model with significantly
reduced dimensionality and controlled loss of accuracy, to ensure reliable evaluations of the reduced model
at substantially reduced computational cost.
Featuring an offline-online framework, the reduced basis (RB) method [16, 26, 28, 30] is a powerful
technique for the ROM of parametrized problems. With a significantly smaller dimension than the full-
order model, a reduced space is spanned by a set of RB modes that are extracted offline from a collection
of full-order snapshots at several time-parameter locations. Two main approaches for extracting the RB are

∗ Correspondingauthor.
Email addresses: [email protected] (Mengwu Guo), [email protected] (Jan S. Hesthaven)

Preprint submitted to Computer Methods in Applied Mechanics and Engineering May 23, 2018
the Greedy algorithm [26, 30, 34] and the proper orthogonal decomposition (POD) [16, 19, 28]. The former
selects a set of snapshots as the basis, following some error estimator/indicator and optimal criterion, while
the latter utilizes the singular-value decomposition (SVD) to extract the RB. For time-dependent problems,
required to capture the causality with the evolution over time, a combined version is proposed as the POD-
greedy approach [14, 16]. Here the POD is used for the compressions over time, and the greedy procedure
is employed in the parameter domain to enable an efficient treatment of parameter variations. Once the RB
space is constructed, the approximated solution for a new desired parameter value is recovered online in the
reduced space. Conventionally, a Galerkin projection is adopted to determine the (evolving) combination
coefficients associated with the RB, yielding the (time-dependent) reduced-order solutions during the online
stage.
For an RB scheme, the key to a successful reduction in computational cost is a full decoupling between
the offline and online stages, i.e. the reduced space is parameter-independent and the online computations
are independent of the size of the high-fidelity solution. However, such a decoupling is often difficult or
even impossible for a general nonlinear problem with non-affine dependence on parameters. In such a case,
assembly of the reduced model is embodied online, within the time steps, loading increments, nonlinear
iterations and updates of the configuration. The empirical interpolation method (EIM) [4] and its discrete
variants [10, 23], also referred to as the hyper-reductions [23], are developed to deal with this problem by
recovering an affine expansion of the differential operator in a non-affine case. Such strategies are problem-
dependent, or intrusive, and suffer from inflexibility in complex applications. For many time-dependent
problems, some intrinsic structures inside the truth model are lost during order reduction, which can result
in a qualitatively wrong or unstable reduced model. Stability of the reduced model for a time-dependent
problem remains an important open question in many cases [6].
There are plenty of works on the ROM. For time-dependent problems, there are many successful cases of
ROM, e.g. the RB method for linear evolution equations [14], the POD-based reduced-order fluid/structure
modeling [20], the nonlinear ROM based on local reduced-order bases [3], some developments of the RB
method for complex models [18, 27, 31], and some recently proposed structure-perserving RB approaches for
Hamiltonian problems [1, 2]. One can refer to [6] for a comprehensive survey. To the best of our knowledge,
however, no non-intrusive approaches have been proposed for the order reduction of parametrized time-
dependent models.
In this work, we consider the parametrized time-dependent problems in the following form:

L[u̇(x, t; µ); µ] + N [u(x, t; µ); µ] = f (x, t; µ) , (x, t, µ) ∈ Ω × T × P , (1)

with some properly defined initial and boundary conditions. Here Ω ⊂ Rm , T = [0, T ] and P ⊂ Rd represents
the domains of space, time and parameters, respectively, with m = 1, 2 or 3 and d being the number of the
parameters characterizing the model. u : Ω × T × P → Rn denotes the parameterized time-dependent
solution field, n is the dimension of the system, and f : Ω × T × P → Rn is the source term. Moreover,
L[·; µ] is a linear operator and N [·; µ] is a nonlinear differential operator, both associated with the space
coordinates x and characterized by the parameters µ.
To enable good performance, flexibility, robustness and online efficiency, a regression-based approach is
used in this work. A Gaussian-type regression was combined with the RB method in [24, 25] to predict
scalar quantities of interest of the high-fidelity simulation. A regression-based approach has been developed
to evaluate the reduced-order solution field for nonlinear time-independent problems, equipped with the
artificial neural networks [15] in [17] and a Gaussian process regression (GPR) [29, 35] in [13]. The regression-
based approach enjoys a complete decoupling of the offline and online stages. During the offline stage, the
RB modes are extracted from a set of full-order snapshots by the POD, and the maps between the parameters
and the projection coefficients onto the RB are approximated as regression models, which are trained from
high-fidelity data by supervised learning [7, 21]. The evaluations for new parameter values only require
direct outputs from the regression models and linear combinations of the RB, implying a very fast online
computation. One can notice that the high-fidelity solver generates the full-order samples as a ’black-box’,
and these samples are used to construct the RB and regression models in a data-driven way. The whole
offline-online procedure is carried out at the algebraic level, which guarantees a non-intrusive nature of the

2
regression-based approach. This allows dealing with different problems in the same way.
In this paper, the GPR-based scheme is utilized and further developed for time-dependent problems.
The time coordinate t is considered as another parameter in the system. Over the time and parameter
domains in consideration, a set of full-order solutions are prepared offline, which includes the snapshots for
the POD and the samples for training the regression models. A tensor-decomposition-based approach is
proposed for a more reliable and efficient regression. We use the SVD to extract the principal components
of the data of each projection coefficient and represent the data as the combination of tensor products of
discrete time- and parameter-models. The regression function for the projection coefficient is expressed in a
similar formulation, with the continuous time- and parameter-modes approximated as GPR models. After
obtaining the reduced model, an error surrogate can be recovered from the available data as another GPR
model, which can provide indicators to certify the quality of the ROM.
Following this introduction, the data-driven RB method is presented in Section 2. After a brief review of
the ideas of GPR, the tensor-decomposition-based regression and the error surrogate model are introduced
in Section 3. In Section 4, the proposed method is validated by three numerical examples, the 1-D viscous
Burgers’ equation, incompressible fluid flow around a cylinder, and the large deformation analysis of a
trussed frame. Conclusions are drawn in Section 5.

2. A data-driven reduced basis method

After discretizing (1) in space Ω by some discretization scheme, one has the following system of param-
eterized ODEs:
Lh [u̇h (t; µ); µ] + Nh [uh (t; µ); (µ)] = fh (t; µ) , (t, µ) ∈ T × P , (2)
where uh : T × P → RNh is the discrete solution, Nh is the number of degrees of freedom (DOFs), Lh ,
Nh and fh : T × P → RNh are the discrete counterparts of the operators L, N and the source term f ,
respectively.
Equipped with some approach for time integration, solving the parametrized time-dependent nonlinear
problem (2) requires the assembly and solution of a number of linear systems. The dimension of such linear
systems, i.e. Nh , is determined by both the underlying mesh and the polynomial order of the discretization
scheme. The high-fidelity simulation for a complex real-world problem often requires a large number of
DOFs and a lot of time steps and iterations, suggesting that such a full-order model is not affordable in
many-query or real-time context of parametrized problems.
The RB method is a reliable and efficient tool for the model order reduction of parametrized problems.
Spanned by a set of parameter-independent RB functions, a reduced space is constructed during the offline
stage, and the approximate solution to the parameterized problem is sought online using this reduced space.
From a collection of high-fidelity snapshots at different parameter values, the RB functions are carefully
chosen either by the Greedy algorithm, or by the principal component analysis of the snapshots. The former
requires an error estimator/indicator for the full-order solution, and picks the snapshot that maximizes the
estimator/indicator until a criteria is satisfied. For time-dependent problems, a combination of the POD
over time and the Greedy procedure in parameter domain can be used. In this work, however, suitable
error estimators or indicators are not available for a general nonlinear time-dependent problem over the
parameter domain, so the proper orthogonal decomposition (POD) is used to extract the RB, as detailed in
the following.
To evaluate the reduced-order solution for any desired time-parameter location, a regression-based ap-
proach will be utilized in this work, rather than the conventional Galerkin-projection-based approach.

2.1. The database of the full-order solutions

For a parametrized time-dependent problem, the notion of a solution manifold M can be introduced,
comprising all the solutions of (1) under variation of the time and parameters, i.e. M = {u(t; µ) : (t, µ) ∈
T × P}, and its discrete counterpart Mh = {uh (t; µ) : (t, µ) ∈ T × P} ⊂ RNh .

3
 Due to the data-driven nature of this work, a set of full-order solutions is prepared as a database prior
to the construction of the reduced-order model. Here the database is represented as

D = {{(t, µ), uh (t; µ)} : (t, µ) ∈ Td × Pd } ,

Td = {t1 , t2 , · · · , tNtd } ⊂ T , (3)
Nµd
Pd = {µ1 , µ2 , · · · , µ }⊂P,

in which the time series are obtained by some selected time discretization scheme. Note that the discrete
time-parameter inputs (t, µ), corresponding to the elements in this database, form a tensor grid between a
point-set in the time domain and one in the parameter domain, i.e. Td and Pd . Usually, Td contains a series
of time steps in the full-order simulation, and Pd can be a uniform lattice or generated from the parameter
domain. We assume that the database includes the essential information about the solution manifold Mh .

2.2. The proper orthogonal decomposition and the reduced basis space
To generate the RB, we consider a collection of snapshots

{s1 , s2 , · · · , sN s } = {uh (t; µ) : (t, µ) ∈ Θ} (4)

associated with a discrete point-set Θ ⊂ Td × Pd . A subspace of RNh can be spanned by the snapshots as

MΘ = span{s1 , s2 , · · · , sN s } ⊂ RNh . (5)

If the point-set Θ is fine enough, MΘ acts as a good representation of Mh .

To reduce the model, a low-rank approximation with rank L  min{Nh , Ns } should be found for MΘ .
Towards this end, the POD is employed to extract a set of orthogonal bases V = [v1 |v2 | · · · |vL ] ∈ RNh ×L
from the snapshots, so that the column space of V, denoted by Col(V), served as a reduced-order replacement
for the full-order space RNh . Then we assume that the reduced-order solution for time-parameter location
(t, µ) ∈ T × P, denoted as urb (t; µ), is represented as a linear combination of the bases V, i.e. urb (t; µ) =
PL T L
l=1 ql (t; µ)vl = Vq(t; µ), where q = {q1 , q2 , · · · , qL } ∈ R collects the combination coefficients.
Nh ×Ns
The snapshot matrix S ∈ R , collecting all the snapshots, is defined as

S = [ s1 | s2 | · · · | sNs ] . (6)

Note that Col(S) = MΘ . The POD takes advantage of the singular value decomposition (SVD) of S, given
as
S = UΣZT (7)
with U ∈ RNh ×Nh and Z ∈ RNs ×Ns being orthogonal matrices, i.e. UT U = INh and ZT Z = INs , and
Σ = diag{σ1 , σ2 , · · · , σNs } contains the singular values σ1 ≥ σ2 ≥ · · · ≥ σNs ≥ 0.
At the algebraic level, one seeks to find the ’best’ approximation of Col(S) in some optimal sense, among
all L-dimensional subspaces with L ≤ rank(S). Let V ∈ RNh ×L be taken as the first L columns of U,
and let YL = {W ∈ RNh ×L : WT W = IL } represent the set of all L-dimensional orthogonal bases. The
projection error Pof the snapshots onto the orthogonal bases W ∈ YL , measured in the Euclidean norm, can
Ns
be expressed as i=1 ksi − WWT si k2RNh .
The Schmidt-Eckart-Young theorem [12, 28, 32] states that the basis consisting of the first L left singular
vectors of S minimizes the projection error of the snapshots among all L-dimensional orthogonal bases in
RNh . The error can be evaluated by the (L + 1)th to Ns th singular values, i.e.
Ns Ns Ns
X 2 X 2 X
si − VVT si = min si − WWT si = σi2 . (8)
RNh W∈YL RNh
i=1 i=1 i=L+1

4
Hence a relative error, corresponding to the minimized projection error, is defined as
PNs 2
si − VVT si PNs 2
i=1
RNh i=L+1 σi
PNs 2
= PNs
. (9)
2
i=1 ksi kRNh i=1 σi

Thus Col(S) can be well approximated by Col(V) with a small L if the singular values decay rapidly.

2.3. Regression-based approach for reduced-order solutions

The computational efficiency of the RB method relies on the decoupling of the offline and online stages.
As discussed, the RB modes are prepared offline from the high-fidelity snapshots and are parameter-
independent. The reduced-order solution for a new parameter value is recovered in the online stage. In
the conventional framework of the RB method, a standard Galerkin approach is used to determine the
combination coefficients of the RB. Due to the non-affinity in parameter dependence, however, a Galerkin-
projection-based scheme will not significantly save the computational cost for a general nonlinear problem.
Moreover, there are online stability issues for time-dependent problem in many cases, as already discussed
in the introduction.
To overcome these difficulties, a regression-based approach is used to calculate the reduced-order solutions
for new time-parameter values. In this scenario, the projection of a full-order discrete solution uh (µ) onto
Col(V) acts as the corresponding reduced-order solution at the algebraic level,

urb (t; µ) = VVT uh (t; µ) = arg min kuh (t; µ) − wh kRNh , (10)
wh ∈Col(V)

in which VT uh (t; µ) = q(t; µ) collects the coefficients associated with column bases of V.
To obtain the projection coefficients q(t; µ) for any desired time-parameter location (t, µ) ∈ T × P, one
can resort to a nonlinear regression q̂ between d + 1 = dim(P) + 1 inputs and L outputs:

(t, µ) 7→ q(t; µ) = VT uh (t; µ) ≈ q̂(t; µ) . (11)

During the offline stage, this regression model q̂(·; ·) should be constructed from a set of training data

Dtr = {(t, µ), VT uh (t; µ)} : t ∈ Ttr , µ ∈ Ptr ,



Ttr = {ti : i = n1 , n2 , · · · , nNttr } ⊂ Td , (12)

j
Ptr = {µ : j = m1 , m2 , · · · , mNµtr } ⊂ Pd .

with 1 ≤ n1 ≤ n2 ≤ · · · ≤ nNttr ≤ Ntd and 1 ≤ m1 ≤ m2 ≤ · · · ≤ mNµtr ≤ Nµd . The model is used during
the online stage to recover the output q̂(t∗ ; µ∗ ) for any new input (t∗ , µ∗ ) ∈ T × P. Correspondingly, the
reduced-order solution urb,reg (t∗ , µ∗ ) ∈ Col(V) is

urb,reg (t∗ ; µ∗ ) = Vq̂(t∗ ; µ∗ ) . (13)

Once the regression model is recovered, the online stage only requires direct outputs from the regression
models, i.e. the online solutions are obtained at very low cost.

5
Algorithm 1 Regression-based RB method for time-dependent problems
1: Offline stage:
2: Prepare the database D of full-order solutions;
3: Extract some snapshots from the database D and form the snapshot matrix S ∈ RNh ×Ns ;
4: Perform POD for S and get the L orthogonal bases V ∈ RNh ×L ;
5: Prepare the training set Dtr from the database D;
6: Construct the regression model q̂(·; ·) from Dtr .

7: Online stage:
8: Recover output q̂(t∗ ; µ∗ ) for a new parameter value (t∗ , µ∗ );
9: Evaluate the reduced-order solution urb,reg (t∗ ; µ∗ ) = Vq̂(t∗ ; µ∗ ).

To meet the high demand on the quantity of data, all samples in the database are usually used when
training the regression models. We note the complete decoupling of the offline and online stages, and the
non-intrusive nature of the regression-based RB method. In this work, Gaussian process models are utilized
to construct the regression q̂(·), as discussed in the following section.

3. Gaussian process regression models

Regression is concerned with prediction of continuous quantities of interest by the construction of a
model from a set of observation data. Let Dtr = {(xi , yi ) : i = 1, 2, · · · , M } denote the training set of M
observations, where each input xi ∈ P ⊂ Rd consists of d entries and lies in the input domain P, and yi
is the output corresponding to xi . In a Gaussian process regression (GPR) model [29, 35], the observed
input-output pairs are assumed to follow some regression function f : P → R, the prior of which is defined
as a Gaussian process (GP). Given the training data, the model uses the posterior GP to make predictions
for new inputs.

3.1. Gaussian processes for regression

A Gaussian process (GP) is a collection of random variables, any finite number of which obeys a joint
Gaussian distribution. In the case of GPR, let the prior on the regression function be a GP corrupted by
an independent Gaussian noise term, i.e. for (x, x0 ) ∈ P × P,

f (x) ∼ GP(m(x), κ(x, x0 )) , y = f (x) +  ,  ∼ N (0, σy2 ) . (14)

Here m(x) := β T H(x), H(x) = {H1 (x), H2 (x), · · · , HN (x)}T contains N basis functions defined in P, and
β = {β1 , β2 , · · · , βN }T are the corresponding combination coefficients. There are many different options for
the covariance function κ : P × P → R. A frequently used one is the automatic relevance determination
(ARD) squared exponential (SE) kernel :
d
!
0 2 1 X (xm − x0m )2
κ(x, x ) = σf exp − , (15)
2 m=1 `2m

which considers an individual correlated lengthscale `m for each input dimension, and allows for differentiated
relevances of input features to the regression.
Given a finite number of points in the input domain, referred to as the parameter locations of training
data, a prior joint Gaussian is defined for the regression outputs:

y|X ∼ N (m(X), Ky ) , Ky = cov[y|X] = κ(X, X) + σy2 IM , (16)

where y = {y1 , y2 , · · · , yM }T , X = [ x1 | x2 | · · · | xM ] and IM is the M -dimensional unit matrix.

From a regression model, the goal is prediction of the noise-free output f ∗ (s) for a new test input
s ∈ P. Then one can combine the information from training set with the predictions for test samples in
6
a joint density of the observed outputs y and the noise-free test output f ∗ (s). By the standard rules for
conditioning Gaussians, the posterior predictive distribution can be obtained as a new GP:

f ∗ (s)| s, X, y ∼ GP(m∗ (s), c∗ (s, s0 )) ,

(17)
m∗ (s) = m(s) + κ(s, X)K−1
y (y − m(X)) , c∗ (s, s0 ) = κ(s, s0 ) − κ(s, X)K−1 0
y κ(X, s ) ,

The values of the hyperparameters θ make significant difference on the predictive performance, with θ =
{β1 , · · · , βN , `1 , · · · , `d , σf , σy } for the case of ARD SE kernel. In this paper, an empirical Bayesian approach
of maximizing likelihood is adopted to determine a set of optimal values of the parameters. Using a standard
gradient-based optimizer, the optimal hyperparameters θ opt can be estimated via the maximization problem:

θ opt = arg max log p(y|X, θ)

θ

1 1 M
 (18)
T T −1 T
= arg max − (y − β H(X)) Ky (θ)(y − β H(X)) − log |Ky (θ)| − log(2π) ,
θ 2 2 2

where p(y|X, θ) is the conditional density function of y given X under hyperparameters θ, also considered
as the marginal likelihood Z
p(y|X, θ) = p(y|f , X, θ)p(f |X, θ) df .

We only use one constant basis function in this paper, i.e. H(x) = H(x) = 1, so there exists only N = 1
combination coefficient β. In the numerical implementations for time-dependent problems, moreover, the
predictive mean function m∗ (·) is of the most interest.

3.2. Tensor-decomposition-based regression

In this work, the regression is equipped with tensor decomposition to ensure reliable and efficient fitting
results. For the lth entry ql = vlT uh of the projection coefficients q = {q1 , q2 , · · · , qL }T , l = 1, 2, · · · , L, the
training data can be written in a matrix as

Pl = [ql (tni ; µmj )]ij , 1 ≤ i ≤ Nttr , 1 ≤ j ≤ Nµtr , (19)

as a natural result of the tensor grid between the time and parameter locations in the training data.
To decompose the data of a projection coefficient into several time- and parameter-modes, the SVD is
employed again as
XQl
Pl ≈ P̃l = λlk ψ lk (φlk )T . (20)
k=1

Here ψ lk and φlk are the kth discrete modes of time and parameters for the lth projection coefficient,
respectively, λlk is the kth singular value for the same coefficient, and Ql is the corresponding rank of
truncation.
From the data of discrete modes, Gaussian processes are trained to approximate the corresponding
continuous modes, as follows

t 7→ ψ̂kl (t) trained from {(tni , (ψ lk )i ) : i = 1, 2, · · · , Nttr } ,

(21)
µ 7→ φ̂lk (µ) trained from {(µmj , (φlk )j ) : j = 1, 2, · · · , Nµtr } ,

where ψ̂kl (t) and φ̂lk (µ) are the kth continuous time- and parameter-modes for the lth projection coefficient,
respectively. Hence we have
Ql
X
mj
(Pl )ij = ql (tni ; µ )≈ λlk ψ̂kl (tni )φ̂lk (µmj ) , 1 ≤ i ≤ Nttr , 1 ≤ j ≤ Nµtr .
k=1

7
A continuous regression function q̂l for the lth projection coefficient ql with respect to time-parameter values
can be recovered as
Ql
X
ql (t; µ) ≈ q̂l (t; µ) = λlk ψ̂kl (t)φ̂lk (µ) , (t, µ) ∈ T × P , (22)
k=1

which shares a similar formulation with (20).

We would like to comment on the advantages of the tensor-decomposition-based regression approach
in the context of time-dependent problems. Usually, the projection coefficients are evolving substantially
with respect to time, leading to difficult (global) GPRs. After the tensor decomposition, the 1D regressions
with respect to time, combined with the multi-dimensional regressions with respect to the parameters, are
both easier. For the lth projection coefficient, there are N + d + 3 hyperparameters in the global regression
model, while Ql (2N + d + 5) in the tensor-decomposition-based one. Much more hyperparameters in the
tensor-decomposition-based approach imply a more flexible scheme for function fitting.
As can be noticed, the major computational effort for obtaining the posterior GP is to calculate K−1 y ,
the complexity of which is O((Nttr Nµtr )3 ) for the global regression and Ql [O((Nttr )3 ) + O((Nµtr )3 )] for the
alternative. Thus the tensor decomposition significantly reduces the computational cost of the regression of
each projection coefficient, even though extra effort of the SVD with complexity O(Nttr Nµtr max{Nttr , Nµtr })
is required. However, this is usually not expensive.

3.3. Error control for the truncations

In this subsection, we suppose the database D is completely used, i.e. all samples have been used both
as snapshots for constructing the RB and for training samples of regressions, i.e.

{uh (t; µ) : (t, µ) ∈ Td × Pd } = {uh (t; µ) : (t, µ) ∈ S} = {uh (t; µ) : (t, µ) ∈ Ttr × Ptr } := F ,

where F collects all the full-order solution vectors in the database.

Following the truncations in the SVDs for the projection coefficients, a relative error for the recovery
can be given as:
PNtd PNµd PL
i=1 kuh (ti ; µj ) − l=l (P̃l )ij vl k2RNh
j=1
ē2rcv (F) = P 2
u∈F kuk2
 d 
Ntd Nµ L  2
1 X
T
X XX
=P 2
 ku − VV uk2RNh + (Pl )ij − (P̃l )ij 
u∈F kukRNh u∈F i=1 j=1 l=1 (23)
PL 2
PL
l=1 kPl − P̃l kF kPl k2F δl2
= ε2L + P 2 = ε2L + Pl=1 2
u∈F kukRNh u∈F kukRNh
L L
X kPl k2F X
= ε2L + P 2 δ 2
l = ε 2
L + ηl2 δl2 ,
l=1 u∈F kukRNh l=1

where P̂l represents the coefficients after the SVD truncation, as defined in (20). Here δl2 stands for the
truncation error of the SVD for the lth projection coefficient, i.e. δl2 = kPl − P̃l k2F /kPl k2F , ηl2 is defined as
PNs 2 PL
σl2 / l=1 σl , σl is the lth singular value in the POD, used to construct the RB, and ε2L = 1 − l=1 ηl2 gives
the truncation error by the POD. In the data-driven context, (23) can be used to evaluate the truncation
errors due to both the POD for constructing the RB and the SVDs for the training data of projection
coefficients. This provides an estimator for the corresponding error control.

3.4. An error surrogate model

Regression have also been used to approximate the errors introduced by reduced-order models. As
presented in [11], the errors in quantities of interest are modelled statistically, and the GPR is employed

8
to map from a small number of inexpensive error indicators to a distribution over the true error. Recently
in [33], some regression functions are directly constructed between the input parameters and the errors in
both the state and quantities of interest. Based on a similar idea, we give a regression model as an error
surrogate to assess or certify the quality of the proposed reduced model.
For each time-parameter location (tni , µmj ) in the training set, i = 1, 2, · · · , Nttr , j = 1, 2, · · · , Nµtr , the
relative error of the reduced-order solution is given as
1
er (tni , µmj ) = kuh (tni ; µmj )) − VE[q̂(tni ; µmj )]kRNh , (24)
kuh (tni ; µmj ))kRNh

which can be evaluated once all the GPR models are constructed. Another regression model can be trained
as an error surrogate, denoted by êr (·, ·) : T × P → R, i.e.

(t, µ) 7→ êr (t, µ) trained from {(tni , µmj , er (tni , µmj )) : i = 1, 2, · · · , Nttr , j = 1, 2, · · · , Nµtr } . (25)

In this work, the error surrogate is constructed as a Gaussian process, and the outputs are given as distribu-
tions, such that the predictive error indicator for any time-parameter location is represented by a mean value
bounded by confidence bounds. Usually, the relative error of a reduced-order solution with respect to time
and parameters is a much more complex function than the projection coefficients onto the RB. As noticed,
the training for error surrogate shares the same sampling locations as those for the projection coefficients,
thus one should not expect too much about the accuracy of its predictive mean function. Even so, one can
resort to the confidence level for non-rigorous bounds of the relative error.

4. Numerical examples

In this section, numerical results of three time-dependent or pseudo-time-dependent examples are pre-
sented. The first one is a simple problem of 1-D Burgers’ equation with parametrized viscosity. Both
the constructed RB and the reduced-order solutions are illustrated, and the effectiveness of the proposed
framework is validated. The second example is a benchmark problem of incompressible fluid flow around a
cylinder. Regression results in the offline stage are plotted, and some online tests are performed to compare
the accuracy of reduced-order solutions with that of full-order ones. Error surrogates are constructed as
GPR models to provide error indicators for the online evaluations. Moreover, by setting a series of carefully
defined criteria for the SVDs of the training data, offline regression models are constructed with further
reduced computational cost and guaranteed online accuracy. In the last example, a structural problem with
large deformation is treated as pseudo-time-dependent to build a reduced model which can recover both
the solution fields and a class of equilibrium paths with respect to several parameters. The GPR models in
these examples are constructed by the MATLAB function fitrgp.

4.1. Example 1: 1-D viscous Burgers’ equation

As a simple test, a one-dimensional viscous Burgers’ equation with parameterized diffusion coefficient is
taken into consideration, given as
µ
ut + uux − uxx = 0 , (x, t, µ) ∈ (−1, 1) × (0, 1] × [1, 7.5] ,
50π
u(−1, t; µ) = u(1, t; µ) = 0 , u(x, 0; µ) = − sin(πx) .

In the offline stage, full-order solutions are calculated by the finite difference method with Nh = 201
nodes at Nµd = 27 uniformly distributed parameter locations of µ, i.e. Pd = {1.00, 1.25, 1.50, · · · , 7.25, 7.5}.
Among the 1000 time steps, Ntd = 250 uniformly distributed ones are included in the database, i.e. Td =
{0.004, 0.008, 0.012, · · · , 0.996, 1.000}. Then all the full-order solution vectors are used both as snapshots
for constructing the RB and training data for the GPR models. By the POD, a set of L = 7 orthogonal
modes are extracted as the RB with a tolerance εL,tol = 0.1%, as shown in Figure 1. As proposed previously,

9
regression models are then constructed as the combinations of time- and parameter-modes, which are trained
as GPs after the SVD truncations with tolerances δl,tol = 0.1%, 1 ≤ l ≤ L = 7.
To validate the accuracy of the reduced model, the reduced-order solutions are recovered at 5 time points
with respect to 4 different values of µ, and are compared with the corresponding full-order solutions in Figure
2. It is seen that the full-order and reduced-order solutions are matching well with each other, even when
the viscous parameter µ is quite small, confirming the effectiveness of the proposed method.

4.2. Example 2: incompressible fluid flow around a cylinder

In this example, we consider a classic benchmark in CFD – the fluid flow around a cylinder. As shown
in Figure 3, the problem is defined in a domain Ω = [0, 2.2] × [0, 0.41] \ Br (0.2, 0.2) with r = 0.05. The
governing equations are the non-stationary Navier-Stokes equations for an incompressible fluid:

ρut − ν∆u + ρu∇u + ∇p = 0, ∇ · u = 0.

Here we take the fluid density as ρ = 1.0 and the viscosity as ν = 0.001. No-slip conditions are set along the
lower and upper walls and ∂Br (0.2, 0.2), and Neumann conditions are set along the right (outflow) edge.
Moreover, the following parabolic velocity profile is prescribed on the left (inflow) edge:
 T
4U y(0.41 − y)
u(0, y) = ,0 ,
0.412

in which the value of parameter U is subject to change in an interval [1.125, 1.5]. Correspondingly, the
Reynolds number Re lies in an interval P = [75, 100], under the definition of Re = Um L/ν, where the mean
velocity is given as Um = 2U/3 and the characteristic length of the flow configuration is L = 0.1. Reduced
models will be constructed for the solution history in time domain T = [2.5, 5.0].
More details about this benchmark problem can be found at http://www.featflow.de/, including some
visualizations of both the velocity solution u and the pressure solution p. In this example, we use the finite
element [5] solver in the MATLAB library redbKIT v2.2 [22, 28] to calculate the full-order solutions. At
51 different values Pd = {75.0, 75.5, 76.0, · · · , 99.5, 100.0} of the parameter Re, we use a small time step to
get the high-fidelity time series of the solutions, for the velocity and the pressure, and add the full-order
solutions at 126 time points t ∈ Td = {2.50, 2.52, 2.54, · · · , 4.98, 5.00} into the database D. All these samples
will be used both as snapshots and training data.
After the POD with a truncation tolerance εL,tol = 0.5%, L = 36 reduced bases are extracted for the
velocity solution and L = 33 for the pressure solution. For comparison we have the full orders Nh = 29318
for velocity and Nh = 3899 for pressure.
Case 1:
In this case, the truncation criteria of the SVDs for all projection coefficients are set as Q1 = 10 and
δl,tol = εL,tol = 0.5% for l = 2, 3, · · · , L. A more strict criterion is defined for the 1st coefficient due to its
dominant position in the accuracy of reduced-order solutions.
After the tensor decompositions for the data of all projection coefficients, the time- and parameter-modes
are obtained as GPR models, some of which are shown in Figures 4 and 5. The mean regression functions
of all the coefficients are recovered in a similar decomposition form, and some of these regression results are
plotted in Figures 6 and 7. It is worth pointing out that all the regressions for the problem can be finished
within minutes.
After the offline training, online tests are performed on the solutions for 4 new Re values that are
not included in the offline database, chosen as Re ∈ {80.25, 85.25, 90.25, 95.25}. The reduced-order
time-dependent for these 4 parameter values are calculated as combinations of the RB modes, with their
coefficients obtained as direct outputs from the regression models. Compared with the corresponding full-
order solutions, the relative errors, measured in 2-norm, are calculated and plotted in Figure 8. As can be
noticed, the relative error are quite close to their lower bounds – the projection errors onto the RB spaces
constructed by the POD – implying a good quality of the regressions. Moreover, the coefficients for lift and
drag, denoted by CL and CD , are also extracted from the reduced-order solutions for the 4 test cases. As

10
shown in Figure 9, they are matching well with the high-fidelity results from the corresponding full-order
solutions.
As discussed in subsection 3.4, error surrogates are constructed as GPR models for the solutions of both
the velocity field and the pressure field. For the online test samples, the predictive errors with their ±σ
bands are shown in Figure 10. It can be seen that the true errors between the reduced-order and full-order
solutions are essentially bounded by the ±σ bands, and they can be strictly bounded by the ±2σ bands or
95% confidence levels. As surrogates for relative errors, the GPs provide non-rigorous error indicators for
the reduced-order solutions.
Case 2:
According to (23), the recovery error caused by all the offline truncations, both the POD for constructing
the RB and the SVDs for the data of projection coefficient, can be controlled using
L
X
ē2rcv,tol (F) = ε2L,tol + ηl2 δl,tol
2
. (26)
l=1

Note that the weight ηl2 is decaying as l increases, meaning that the lower-order projection coefficients
contribute more to the recovery accuracy. Moreover, the regressions usually get more difficult when l gets
larger. A natural thought is to set a more strict SVD-truncation criterion, i.e. a smaller δl , for a lower-order
projection coefficient. In this way, the regressions are easier and more efficient, provided the accuracy is
under control. √
In this case, we set δl,tol = αεL,tol /(ηl L), so that ηl2 δl,tol
2
= (αεL,tol )2 for each 1 ≤ l ≤ L, i.e., each
term of the summation in (26) contributes equally. To ensure similar values of ē2rcv,tol as those in Case 1,
we take α = 0.2 for the velocity and α = 0.27 for the pressure, in which case one has ē2rcv,tol = 1.040ε2L,tol
for the velocity and ē2rcv,tol = 1.073ε2L,tol for the pressure, accordingly.
Following the same procedure of offline training and online tests as in Case 1, the reduced-order solutions
for the 4 test values of Re are obtained, and their relative error are shown in Figure 11, compared with those
of Case 1. It can be seen that the solutions in both cases have a similar accuracy, as a natural results of
sharing similar ē2rcv,tol ’s. On the other side, the rank of SVD-truncations, Ql , for both velocity and pressure
in the two cases are PLshown in Figure 12. It is evident that the total number of regressions for time- or
parameter-modes, l=1 Ql , is much reduced in Case 2, which significantly enhances the offline efficiency.
One advantage of the proposed regression-based framework is that the regression models for the projection
coefficients are trained separately, and one can set an individual criterion for the construction of each model.
In Case 2, we take advantage of this to further reduce the efficiency. The way of setting criteria is problem-
independent and can be adopted in other numerical applications.

4.3. Example 3: equilibrium paths of a trussed frame

A frame made of a beam and a column is considered in this example. As shown in Figure 13, the frame
is trussed by 596 one-dimensional hyperelastic elements and loaded by a concentrated load on the beam.
The number of DOFs of the full-order model is Nh = 476, and the FLagSHyP MATLAB program [8, 9] is
used as a high-fidelity solver for this problem. The quantities in Figure 13 are given as: Young’s modulus
E = 210 GPa, unit force F0 = 1 N and unit displacement ∆0 = 1 mm. With a uniform Young’s modulus E
in the whole structure, the equilibrium paths, referred to as the load-displacement curve, is shown in Figure
14 together with some configurations of the frame at different loading stages.
There are two parameters µ1 and µ2 in this example: µ1 ∈ [0.5, 1.5] is the scaling factor of Young’s
modulus of a 10 mm × 10 mm inverted L-shaped zone at the beam-column joint, and µ2 ∈ [0.8, 1.2] is that
of the lower half of column. We seek to build a reduced model for a class of equilibrium paths with respect
to the two varying material parameters.
The loading increments for an equilibrium path are obtained by the arc-length method. It is natural
to consider the incremental procedure as a time series, and consider the deformation of this structure
along an equilibrium path as a pseudo-time-dependent problem. For the full-order database, we calculate
the 100 loading increments with fixed arc-length at randomly generated Nµd = 100 parameter locations in

11
P = [0.5, 1.5]×[0.8, 1.2]. We define T = [0, 1], and the Ntd = 101 pseudo-time steps, representing the loading
increments, are ti = (i − 1)/(Ntd − 1), i = 1, 2, · · · , Ntd . All full-order samples will be used both as snapshots
and training data.
The proposed approach is employed to construct the RB and the regression models. As in Case 2 of
Example 2, we adopt different criteria for the SVDs of projection coefficient data, with εL,tol = 0.1% and
α = 1.0. To plot the paths, one extracts the load F from the training samples and build a GPR model
F̂ (t, µ1 , µ2 ) for this load.
Online tests are performed at 4 new parameter locations, not included in the database. One can extract
the vertical displacement at the loading node, denoted as ∆(t, ˆ µ1 , µ2 ), from the combination of RB modes
with their approximate coefficients, and plot the paths as curves of F̂ (t, µ1 , µ2 ) versus ∆(t, ˆ µ1 , µ2 ) with t
eliminated. The curves for the 4 test locations of (µ1 , µ2 ) are shown in Figure 15, matching very well with
those obtained by full-order simulations.
In the large deformation analysis in structural mechanics, it is important, but usually challenging, to
obtain the equilibrium path of a structure under some loading pattern. As shown in this example, such a
problem can be dealt with by constructing a reduced order model for fast recovery of both the displacement
solution fields and the equilibrium paths. It is worth noting that any state in the loading procedure and
any path within the parameter domain can be recovered by the regression models.

5. Conclusions

A regression-based reduced basis method is proposed for time-dependent problems. A database of full-
order solution samples is prepared offline, and provide data for the constructions of both the RB and the
regression models for projection coefficients. Equipped with the SVDs to extract the principal components of
the data of projection coefficients, each regression function is expressed in the form of a tensor decomposition
via a series of time- and parameter-modes of GPs. The control of truncation errors is discussed, and a GPR
model of an error surrogate is introduced to certify the quality of the reduced-order models. An example of
1-D viscous Burgers’ equation and one of incompressible fluid flow around a cylinder are tested to validate
the effectiveness of the proposed method. Moreover, a large deformation analysis of a trussed frame is
considered as a pseudo-time-dependent problem with loading increment defined as the pseudo-time, and a
reduced model is derived for a class of equilibrium paths with respect to parameter variation.
Due to its non-intrusive nature, the proposed method trains reduced-order models at the algebraic level,
and provides fast and reliable online calculations for parametrized time-dependent problems. This technique
provides a promising tool for the solution of multi-entry or real-time large-scale complex problems in science
and engineering.

References
[1] B. M. Afkham, A. Bhatt, B. Haasdonk, and J. S. Hesthaven. Symplectic model-reduction with a weighted inner product.
arXiv, No. 1803.07799, 2018.
[2] B. M. Afkham and J. S. Hesthaven. Structure preserving model reduction of parametric hamiltonian systems. SIAM
Journal on Scientific Computing, 39(6):A2616–A2644, 2017.
[3] D. Amsallem, M. J. Zahr, and C. Farhat. Nonlinear model order reduction based on local reduced-order bases. International
Journal for Numerical Methods in Engineering, 92(10):891–916, 2012.
[4] M. Barrault, Y. Maday, N. C. Nguyen, and A. T. Patera. An empirical interpolation method: application to efficient
reduced-basis discretization of partial differential equations. Comptes Rendus Mathematique, 339(9):667–672, 2004.
[5] K.-J. Bathe. Finite Element Procedures. Klaus-Jurgen Bathe, 2006.
[6] P. Benner, S. Gugercin, and K. Willcox. A survey of projection-based model reduction methods for parametric dynamical
systems. SIAM review, 57(4):483–531, 2015.
[7] C. M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
[8] J. Bonet, A. J. Gil, and R. D. Wood. FLagSHyP software. http://www.flagshyp.com/, 2016.
[9] J. Bonet, A. J. Gil, and R. D. Wood. Nonlinear Solid Mechanics for Finite Element Analysis: Statics. Cambridge
University Press, 2016.
[10] S. Chaturantabut and D. C. Sorensen. Nonlinear model reduction via discrete empirical interpolation. SIAM Journal on
Scientific Computing, 32(5):2737–2764, 2010.

12
[11] M. Drohmann and K. Carlberg. The ROMES method for statistical modeling of reduced-order-model error. SIAM/ASA
Journal on Uncertainty Quantification, 3(1):116–145, 2015.
[12] C. Eckart and G. Young. The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211–218, 1936.
[13] M. Guo and J. S. Hesthaven. Reduced order modeling for nonlinear structural analysis using gaussian process regression.
Technical report, EPFL, 2017.
[14] B. Haasdonk and M. Ohlberger. Reduced basis method for finite volume approximations of parametrized linear evolution
equations. ESAIM: Mathematical Modelling and Numerical Analysis, 42(2):277–302, 2008.
[15] S. Haykin. Neural Networks: A Comprehensive Foundation. Prentice Hall, 1999.
[16] J. S. Hesthaven, G. Rozza, and B. Stamm. Certified reduced basis methods for parametrized partial differential equations.
Springer, 2016.
[17] J. S. Hesthaven and S. Ubbiali. Non-intrusive reduced order modeling of nonlinear problems using neural networks. Journal
of Computational Physics, 363:55–78, 2018.
[18] T. Lassila, A. Manzoni, A. Quarteroni, and G. Rozza. A reduced computational and geometrical framework for inverse
problems in hemodynamics. International journal for numerical methods in biomedical engineering, 29(7):741–776, 2013.
[19] Y. Liang, H. Lee, S. Lim, W. Lin, K. Lee, and C. Wu. Proper orthogonal decomposition and its applicationspart i: Theory.
Journal of Sound and vibration, 252(3):527–544, 2002.
[20] T. Lieu, C. Farhat, and M. Lesoinne. Reduced-order fluid/structure modeling of a complete aircraft configuration. Com-
puter methods in applied mechanics and engineering, 195(41-43):5730–5742, 2006.
[21] K. P. Murphy. Machine Learning: A Probabilistic Perspective. MIT press, 2012.
[22] F. Negri. redbKIT Version 2.2. http://redbkit.github.io/redbKIT/, 2016.
[23] F. Negri, A. Manzoni, and D. Amsallem. Efficient model reduction of parametrized systems by matrix discrete empirical
interpolation. Journal of Computational Physics, 303:431–454, 2015.
[24] N. C. Nguyen and J. Peraire. Gaussian functional regression for linear partial differential equations. Computer Methods
in Applied Mechanics and Engineering, 287:69–89, 2015.
[25] N. C. Nguyen and J. Peraire. Gaussian functional regression for output prediction: Model assimilation and experimental
design. Journal of Computational Physics, 309:52–68, 2016.
[26] A. T. Patera and G. Rozza. Reduced Basis Approximation and A Posteriori Error Estimation for Parametrized Par-
tial Differential Equations. Copyright MIT 2007, MIT Pappalardo Graduate Monographs in Mechanical Engineering,
http://www.augustine.mit.edu, 2007.
[27] G. Pitton and G. Rozza. On the application of reduced basis methods to bifurcation problems in incompressible fluid
dynamics. Journal of Scientific Computing, 73(1):157–177, 2017.
[28] A. Quarteroni, A. Manzoni, and F. Negri. Reduced basis methods for partial differential equations: an introduction,
volume 92. Springer, 2015.
[29] C. E. Rasmussen and C. K. Williams. Gaussian Processes for Machine Learning. MIT press Cambridge, 2006.
[30] G. Rozza, D. B. P. Huynh, and A. T. Patera. Reduced basis approximation and a posteriori error estimation for affinely
parametrized elliptic coercive partial differential equations. Archives of Computational Methods in Engineering, 15(3):1,
2007.
[31] A. Sartori, D. Baroli, A. Cammi, L. Luzzi, and G. Rozza. A reduced order model for multi-group time-dependent
parametrized reactor spatial kinetics. In 2014 22nd International Conference on Nuclear Engineering, pages
V005T17A048–V005T17A048. American Society of Mechanical Engineers, 2014.
[32] E. Schmidt. Zur theorie der linearen und nichtlinearen integralgleichungen. i. teil: Entwicklung willkrlicher funktionen
nach systemen vorgeschriebener. Mathematische Annalen, 63:433–476, 1907.
[33] S. Trehan, K. Carlberg, and L. J. Durlofsky. Error modeling for surrogates of dynamical systems using machine learning.
International Journal for Numerical Methods in Engineering, 2017.
[34] C. Wang, J. S. Hesthaven, J. Bai, Y. Zhang, and T. Yang. A greedy non-intrusive reduced order model for fluid dynamics.
Technical report, submitted to American Institute of Aeronautics and Astronautics, 2017.
[35] C. K. Williams and C. E. Rasmussen. Gaussian processes for regression. In Advances in neural information processing
systems, pages 514–520, 1996.

13
 The 1st mode The 2nd mode
0.1 0.2

0 0

-0.1 -0.2
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x
The 3rd mode The 4th mode
0.5 0.5

0 0

-0.5 -0.5
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x
The 5th mode The 6th mode
0.5 0.5

0 0

-0.5 -0.5
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x
The 7th mode
0.5

-0.5
-1 -0.5 0 0.5 1
x

Figure 1: The RB modes for the parametrized viscous Burgers’ equation

14
 µ = 1.125 µ = 3.125
1 1

0.5 0.5

0 0
u

u
-0.5 -0.5

-1 -1
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x
µ = 5.125 µ = 7.125
1 1

0.5 0.5

0 0
u

-0.5 -0.5

-1 -1
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x x

Figure 2: Comparison between the full-order solutions and the corresponding reduced-order solutions: solutions at different
t ∈ {0.25, 0.50, 0.75, 0.90, 1.00} with different values of µ: full-order solutions – dashed, reduced-order solution – solid.

0.2

0.21 0.05

0.2

2.2

Figure 3: Underlying geometry for the problem of fluid flow around a cylinder

15
 The 5th coefficent -- time modes The 5th coefficent -- parameter modes
0.5 0.5

0 0

-0.5 -0.5
2.5 3 3.5 4 4.5 5 75 80 85 90 95 100
t Re
The 15th coefficent -- time modes The 15th coefficent -- parameter modes
0.5 0.5

0 0

-0.5 -0.5
2.5 3 3.5 4 4.5 5 75 80 85 90 95 100
t Re
The 25th coefficent -- time modes The 25th coefficent -- parameter modes
0.5 0.5

0 0

-0.5 -0.5
2.5 3 3.5 4 4.5 5 75 80 85 90 95 100
t Re
The 35th coefficent -- time modes The 35th coefficent -- parameter modes
0.5 0.5

0 0

-0.5 -0.5
2.5 3 3.5 4 4.5 5 75 80 85 90 95 100
t Re

Figure 4: Time- and parameter-modes for the 5th, 15th, 25th and 35th projection coefficients for the velocity field: the 1st
modes – black, the 5th modes – blue, the 10th modes – red, the 15th modes – blue dashed, the 20th modes – black dashed

16
 The 5th coefficient - time modes The 5th coefficient - parameter modes
0.5 0.5

0 0

-0.5 -0.5
2.5 3 3.5 4 4.5 5 75 80 85 90 95 100
t Re
The 10th coefficient - time modes The 10th coefficient - parameter modes
0.5 0.5

0 0

-0.5 -0.5
2.5 3 3.5 4 4.5 5 75 80 85 90 95 100
t Re
The 20th coefficient - time modes The 20th coefficient - parameter modes
0.5 0.5

0 0

-0.5 -0.5
2.5 3 3.5 4 4.5 5 75 80 85 90 95 100
t Re
The 30th coefficient - time modes The 30th coefficient - parameter modes
0.5 0.5

0 0

-0.5 -0.5
2.5 3 3.5 4 4.5 5 75 80 85 90 95 100
t Re

Figure 5: Time- and parameter-modes for the 5th, 10th, 20th and 30th projection coefficients for the pressure field: the 1st
modes – black, the 5th modes – blue, the 10th modes – red, the 15th modes – blue dashed, the 20th modes – black dashed

17
Figure 6: Regression results for the 5th, 15th, 25th and 35th projection coefficients for the velocity field

18
Figure 7: Regression results for the 5th, 10th, 20th and 30th projection coefficients for the pressure field

19
 ×10 -3 Re = 80.25 ×10 -3 Re = 85.25
12 14
POD-GPR POD-GPR
10 12
POD POD
Relative error

Relative error
10
8
8
6
6
4 4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

×10 -3 Re = 90.25 ×10 -3 Re = 95.25

12 10
POD-GPR POD-GPR
10 POD POD
8
Relative error

8 Relative error
6
6

4
4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

(a)

×10 -3 Re = 80.25 ×10 -3 Re = 85.25

10 10
POD-GPR POD-GPR
POD POD
8 8
Relative error

Relative error

6 6

4 4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

×10 -3 Re = 90.25 ×10 -3 Re = 95.25

12 10
POD-GPR POD-GPR
10 POD POD
8
Relative error

Relative error

8
6
6

4
4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

(b)

Figure 8: Relative errors of the solutions at the time steps with different Re values: (a) the velocity field; (b) the pressure field.

20
 Re = 80.25 Re = 80.25
1 3.35
Full-order Full-order
0.5 Reduced-order 3.3 Reduced-order

CD
CL

0 3.25

-0.5 3.2

-1 3.15
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t
Re = 85.25 Re = 85.25
1 3.3
Full-order Full-order
0.5 Reduced-order Reduced-order
3.25
CD
CL

0
3.2
-0.5

-1 3.15
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t
Re = 90.25 Re = 90.25
1 3.3
Full-order Full-order
0.5 Reduced-order 3.25 Reduced-order
CD
CL

0 3.2

-0.5 3.15

-1 3.1
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t
Re = 95.25 Re = 95.25
1 3.25
Full-order Full-order
0.5 Reduced-order Reduced-order
3.2
CD
CL

0
3.15
-0.5

-1 3.1
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

Figure 9: Comparisons of the drag and lift coefficients extracted from the full-order and reduced-order solutions under different
Re values

21
 ×10 -3 Re = 80.25 ×10 -3 Re = 85.25
14 14
±σ band ±σ band
12 12
True True
Relative error

Relative error
Predictive mean Predictive mean
10 10

8 8

6 6

4 4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

×10 -3 Re = 90.25 ×10 -3 Re = 95.25

12 12
±σ band ±σ band
10 True 10 True
Relative error

Relative error
Predictive mean Predictive mean
8 8

6 6

4 4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

(a)

×10 -3 Re = 80.25 ×10 -3 Re = 85.25

12 12
±σ band ±σ band
10 True 10 True
Relative error

Relative error

Predictive mean Predictive mean

8 8

6 6

4 4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

×10 -3 Re = 90.25 ×10 -3 Re = 95.25

14 12
±σ band ±σ band
12 10
True True
Relative error

Relative error

Predictive mean Predictive mean

10 8
8
6
6
4
4
2
2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

(b)

Figure 10: GP Surrogate of the relative errors of the solutions with different Re values: (a) the velocity field; (b) the pressure
field.

22
 ×10 -3 Re = 80.25 ×10 -3 Re = 85.25
12 14
POD-GPR 1 POD-GPR 1
10 12
POD-GPR 2 POD-GPR 2
Relative error

Relative error
POD POD
10
8
8
6
6
4 4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

×10 -3 Re = 90.25 ×10 -3 Re = 95.25

12 10
POD-GPR 1 POD-GPR 1
10 POD-GPR 2 POD-GPR 2
8
Relative error

Relative error
POD POD
8
6
6

4
4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

(a)

×10 -3 Re = 80.25 ×10 -3 Re = 85.25

10 10
POD-GPR 1 POD-GPR 1
POD-GPR 2 POD-GPR 2
8 8
Relative error

Relative error

POD POD

6 6

4 4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

×10 -3 Re = 90.25 ×10 -3 Re = 95.25

14 12
POD-GPR 1 POD-GPR 1
12 10
POD-GPR 2 POD-GPR 2
Relative error

Relative error

POD POD
10
8
8
6
6

4 4

2 2
2.5 3 3.5 4 4.5 5 2.5 3 3.5 4 4.5 5
time t time t

(b)

Figure 11: Relative errors of the solutions at the time steps with different Re values: (a) the velocity field; (b) the pressure
field.

23
 40

velocity in Case 1
Number of the decomposition modes Q velocity in Case 2
35
pressure in Case 1
pressure in Case 2

30

25

20

15

10
5 10 15 20 25 30 35
Series number of the projection coefficient

Figure 12: Comparison of the number of the decomposition modes Q for the projection coefficients

24
F
10
10
D
m1E

120

m2E
60 Y / mm

X / mm

120

Figure 13: A trussed frame

24
 Figure 14: Configurations at different loading stages

(µ 1,µ 2) = (0.75,1.1) (µ 1,µ 2) = (1.25,1.1)

600 1200
Reduced-order Reduced-order
Vertical concentrated load F / N

Vertical concentrated load F / N

1000
Full-order Full-order
400
800

200 600

400
0 200

0
-200
-200

-400 -400
0 20 40 60 80 100 0 20 40 60 80 100
Vertical displacement ∆ at the loading node / mm Vertical displacement ∆ at the loading node / mm

(µ 1,µ 2) = (0.75,0.9) (µ 1,µ 2) = (1.25,0.9)

600 600
Reduced-order Reduced-order
Vertical concentrated load F / N

Vertical concentrated load F / N

Full-order Full-order
400 400

200 200

0 0

-200 -200

-400 -400
0 20 40 60 80 100 0 20 40 60 80 100
Vertical displacement ∆ at the loading node / mm Vertical displacement ∆ at the loading node / mm

Figure 15: Comparison between the equilibrium paths extracted from the reduced-order and the full-order solutions

25