Available online at www.sciencedirect.

com

Procedia Engineering 00 (2017) 000000

www.elsevier.com/locate/procedia

X International Conference on Structural Dynamics, EURODYN 2017

Model order reduction in design of parameterized structures under

different load configurations
Raul Rodrguez Sancheza,, Martin Buchschmida , Gerhand Mullera
a Chair of Structural Mechanics, Technical University of Munich, Arcisstrae, Munich, 80333, Germany

Abstract
In structural design of mechanical systems dynamic analysis is carried out in time domain or in frequency domain which implies
solving the equation of motion several times. Usually the systems depend on a set of parameters which influence their responses.
Thus the design process includes numerical simulations using a full-scale finite element (FE) model for each set of parameters
which is computationally demanding and time consuming. In this contribution the response in the frequency domain due to different
load configurations is investigated by using a mixed approach of two related methods for parametric model order reduction (MOR)
based on interpolation in matrix manifolds of the reduced order models (ROMs) and by using a global basis over the parametric
space. Furthermore an approach based on interpolation of the reduced solution is presented. These approaches of MOR permit
the computational efficient evaluation of different load configurations and avoid the generation of a new FE model for each case.
A numerical example illustrates the capability of those methods. The respective results using parametric MOR approaches are
compared with the solution obtained by using the corresponding full-scale FE model and the direct application of the Krylov
subspace method (KSM).
c 2017 The Authors. Published by Elsevier Ltd.
Peer-review under responsibility of the organizing committee of EURODYN 2017.

Keywords:
Parametric model order reduction, Interpolation in matrix manifolds, Frequency response, Krylov subspace, Interpolation

1. Introduction

The equation of motion resulting from a finite element discretization of a mechanical system is given by
Mu(t) + Du(t) + Ku(t) = f (t) (1)
  
Applying a Fourier transformation to Eq.1, with u() = F u(t) and f() = F f (t) leads to
( j 2 M + i j D + K)u( j ) = f( j ) j = 1, 2, 3...n (2)
which is the equation of motion in the frequency domain, where M, D, K RNN are the mass, damping and stiffness
matrix respectively, f RN1 is the load vector (force vector) and u RN1 is the displacement vector. A reduced-
order model (ROM) which would lead to a lower computational time is achieved by using a suitable MOR technique.

Corresponding author. Tel.: +49-89-289-28329 ; fax: +49-89-289-28665.

E-mail address: raul.rodriguez@tum.de

1877-7058 c 2017 The Authors. Published by Elsevier Ltd.

Peer-review under responsibility of the organizing committee of EURODYN 2017.
2 Raul Rodrguez et al. / Procedia Engineering 00 (2017) 000000

The main idea of MOR techniques is to find a vector space spanned by the columns of V CNnr , with nr  N, which
maps a reduced set of degrees of freedom (dofs) ur Cnr 1 into the original set of degrees of freedom u, such that

u u = Vur (3)

The reduced order model is given by

( j 2 Mr + i j Dr + Kr )ur ( j ) = fr ( j ) (4)

where Mr , Dr , Kr Cnr nr , fr Cnr 1 are defined by

Mr = VH MV Dr = VH DV Kr = VH KV
(5)
fr = VH f

For the construction of that matrix V different approaches have been proposed in the field of structural mechanics,
i.e. Real Modal Analysis, Guyan-Irons Reduction, Improved Reduction System, Dynamic Reduction, Craig-Bampton
Method, Krylov subspace method and Derivative-based Galerkin Projection. A review of those MOR techniques is
presented in [1].
When Eq. 2 depends on a set of parameters i Rk it is possible to have n ROMs corresponding to n sets of
parameters, but the question arises how to use those ROMs in order to determinate the solution in a new parameter set
n +1 . Interpolation on Matrix Manifolds is a promising approach because it works only at the level of ROMs. In this
contribution those methods are applied to a system with different load configurations. The results are compared with
results from a ROM using a truncated global basis Vg and the one from interpolation of the reduced solutions. In the
next section the theoretical background for each approach is summarized. In section 3 a numerical example illustrates
the capabilities of the pMOR mentioned above and a conclusion is given in section 4.

2. parametric Model Order Reduction

2.1. pMOR based on interpolation in matrix manifolds

The two main approaches for the pMOR based on interpolation in Matrix manifolds are reviewed in this subsec-
tion, i.e. the method developed by Lohmann [2] (Method A) and the method developed by Farhat [4] (Method B).
The general procedure to use method A and B was introduced by Geuss et. al [3]. Here the general procedure is
summarized for the particular problem which is addressed in this work.

Sample the parametric space. A coarse grid is defined in the parametric space of interest, i.e. n parametric
configurations are chosen.
Reduction of the local system. n ROMs are computed using any MOR technique and a data base is build with
Mri , Dri , Kri , Fri , Vi , for i = 1...n . In this contribution the KSM is used to reduced the systems at the samples
points.
Adjustment of the Matrix V. The ROMs should be expressed in a common subspace V x , i.e. all the local
systems share the same generalize coordinates x. It is achieved using a transformation matrix Qi :

V x x = Vi Q i x (6)

The definition of matrix V x and Qi according to Method A and B are given in table 1.
Choice of the interpolation manifold. Method A uses the original matrix manifold for the interpolation while
method B uses a tangent manifold as interpolation space. The projection to a tangent space to the matrix
manifold is given by the logarithm mapping and to return back to the original manifold the exponential mapping
is used. The logarithm mapping of a matrix Mi with respect to M x is given by [5]

Mlog,i (M x , Mi ) = UVT (7)

 Raul Rodrguez et al. / Procedia Engineering 00 (2017) 000000 3

Table 1. Selection of reference subspace and transformation matrix

Method A Method B
V x = U(:, 1 : nr ) V x = Vi0
where   where
UWT = svd [V1 ...Vn ], econ i0 {1, 2, ... n }

Qi = (VTx Vi )1 Qi = Ui VTi
where  
UWT = svd VTi V x , econ

where = tan1 and UVT = svd((I M x M x )(MTx Mi )1 ). The exponential mapping is given by

Mi+1 = M x V cos + U sin (8)

with UVT = svd(Mlog,i+1 , econ).

It is important to mention that few dofs, nout  N, have to be evaluated, otherwise the methods are not compu-
tationally efficient, i.e. nout rows of the Basis are going to be interpolated.
Choice of the interpolation method. The Interpolation is done elementwise, so any interpolation method like
spline interpolation or polynomial interpolation could be used.

2.2. pMOR based on interpolation of the reduced solutions

To apply this approach the second step of the procedure mentioned above needs to be extended, i.e. the solution ui
of ROMi for i = 1..n is computed and saved in a data base along with Vi . Then the interpolation is carried out just in
the few rows, nout  N, of Vi and in the set of reduced solutions.

2.3. pMOR using Global Basis

For this approach step two of subsection 2.1 is stopped after computing Vi , then the union or direct sum of the
basis [6] is computed,
n
[
Vbig = Vi (9)
i=1

A QR decomposition is carried out using Vbig in order to get an orthonormal space

QR = qr(Vbig ) (10)

Vg is obtained by truncating Q:

Vg = Q(:, 1 : n nr ) (11)

where < 1 and its optimal value depends on the problem. Then a standard MOR procedure is applied to the new
model using Vg .

3. Numerical example

The capabilities of the method described above are tested for a solar panel modeled with ANSYS R , see Fig. 1. It
is a 4 m 12 m structure, built out of beam and shell elements. The beam188 element was used for the frame, with
a youngs modulus of 2 1011 mN2 , a Poissons ratio of 0.3 and mass density of 7850 mkg3 . For the panel the shell181
element was used, with an elastic modulus of 0.7 1011 mN2 , a Poissons ratio of 0.3 and a mass density of 2500 mkg3 .
4 Raul Rodrguez et al. / Procedia Engineering 00 (2017) 000000

Fig. 1. ANSYS
R
model for solar panel.

Frequency response
10 0
DM
KSM
iROMs
10 -2 iUr
KSM-Vg
Amplitude [m]

10 -4

10 -6

10 -8
0 20 40 60 80 100 120 140 160 180 200
Frequency [Hz]

Fig. 2. Frequency Response at z = 0.4m for 20 frequency points.

This structure would be loaded with a single load of 100 N in vertical direction (direction of the global coordinate
system) with different positions along z direction (from z = 0 to z = 12m), see Fig. 1. In total, the solar panel contains
9582 dofs, 3 nodes are fixed, i.e. 18 dofs, and it has 33 coupling dofs, thus the equation system to be solved contains
9531 unknowns (displacements and rotations).
For a first investigation n = 21, (0, 0.6 ,1.2,...,12), and the frequency band is defined from 0Hz to 200Hz.
The reduction of the configurations is achieved using KSM. The coarse mesh for the interpolation contains 20 fre-
quency points uniformly distributed between 5Hz and 200Hz. The method based on interpolation in matrix mani-
folds (iMORs) uses Method B for the computation of V x and Qi but interpolates the matrix entries in the original
manifold according to Method A. The method based on interpolation of the reduced solutions (iUr) uses Method
B for the computation of V x and Qi . The method based on the global basis (KSM-Vg) uses KSM and Vg with
= 0.4, n = 21 and nr = 60 columns. The amplitude of the response for the three approaches used in this contribu-
tion, are depicted in Fig. 2 along with the direct method (DM) and the direct reduction using KSM. The relative error
is depicted in Fig. 3.
It is clear from the plot of the relative error that the KSM-Vg performs better than the other two approaches and
also better that the direct KSM. The computation time are given in Tab. 2.

Table 2. Computational time.

Method Time (s)
DM 18.05
KSM-direct 5.30
interp-ROMs 2.58
interp-Ur 4.29
ROM-Vglobal 2.78
 Raul Rodrguez et al. / Procedia Engineering 00 (2017) 000000 5

Relative Error between Original Model and pMORs

10 5

10 0
Relative error

10 -5

10 -10 KSM
iROMs
iUr
KSM-Vg

10 -15
0 20 40 60 80 100 120 140 160 180 200
Frequency [Hz]

Fig. 3. Relative error at z = 0.4m for 20 frequency points.

Frequency Response
10 -2
DM
KSM
iROMs
10 -4 iUr
KSM-Vg
Amplitude [m]

10 -6

10 -8

10 -10
0 20 40 60 80 100 120 140 160 180 200
Frequency [Hz]

Fig. 4. Frequency Response at z = 5m for 40 frequency points.

Relative Error between Original Model and pMORs

10 5

10 0
Relative Error

10 -5

10 -10 KSM
iROMs
iUr
KSM-Vg

10 -15
0 20 40 60 80 100 120 140 160 180 200
Frequency [Hz]

Fig. 5. Relative error at z = 5m for 40 frequency points.

A second investigation is carried out with the same parametric data as the first but with 40 frequency points in the
coarse mesh for the application of the KSM in the generation of the ROMs. = 0.2 and nr = 120 is used for Vg . The
new response is depicted in Fig. 4 and the relative error in Fig. 5.
6 Raul Rodrguez et al. / Procedia Engineering 00 (2017) 000000

4. Conclusion

For analysis of structures with different load configurations the reduction using the global basis, KSM-Vg, seems
to be a promising approach. The accuracy of the results delivered by that method depends on the value of Eq.
11, but its optimal value remains and open question. For the test case in this contribution it is the best approach
considering computational time (a factor of two digits in computational time could be saved depending on the number
of simulations to be carried out) and accuracy of the results. The another advantage of the approach is that the solution
of all dofs is obtained. The approach based on interpolation of the reduced solutions, iUr, is an alternative to tackle
this kind of problems. In this contribution it delivers a similar accuracy to iROMs, but both methods are useful only
when the response of just a few dofs needs to be computed.

References
[1] R. Rodrguez Sanchez, M. Buchschmid, G. Muller, Model Order Reduction in Structural Dynamics, ECCOMAS 2015.
[2] H. Panzer, J. Rudy Eid, B. Lohmann, Parametric Model Order Reduction by Matrix Interpolation, Automatisierungstechnik, 58(2010) 475484.
[3] M. Guess, H. Panzer, B. Lohmann, On Parametric Model Order Reduction by Matrix Interpolation, 2013 European Control Conference,
58(2013) 34333438.
[4] D. Amsallem, C. Farhat, An online method for interpolating linear parametric reduced-order models, SIAM,33 (2011), 21692198.
[5] E. Begelfor, M. Werman, Affine invariance revised, Journal of Mathematical Sciences 153.6 (2008): 727-762.
[6] P. Benner, S. Gugercin, K. Willcox, A survey of Projection-Based Model Reduction Methods for Parametric Dynamical Systems, SIAM, 57
(2015), 483-531.