Professional Documents
Culture Documents
DOI 10.1007/s00158-011-0757-1
INDUSTRIAL APPLICATION
Received: 26 August 2011 / Revised: 2 December 2011 / Accepted: 26 December 2011 / Published online: 13 January 2012
c Springer-Verlag 2012
1 Introduction
Using high-fidelity simulation models to predict the
response of structures for design optimization and uncertainty quantification often leads to an unacceptable computational cost, thus motivating the research of techniques
to extract features from complex physical fields using a
reduced number of full-order numerical experiments. The
Proper Orthogonal Decomposition (Berkooz et al. 1993)
is of particular interest in the optimization of engineering
problems, where a set of scalar objective/constraint functions that depends on the values of design variables, is
evaluated such as a surface/volume integration of a velocity/stress/. . . field obtained from a finite element/volume
simulation. Literature reviewed thus far shows a large volume of work recently published on improving the precision
of POD-like ROMs and on interpolation between these
ROMs.
One of the research directions focuses on minimizing
the number of full-scale analyses by including information about the gradients. Weickum et al. (2009) enriched a
POD for the transient response of linear elastic structures
by using the gradients of the POD modes with respect to
the design/random parameters for robust shape optimization. In the same spirit (Carlberg and Farhat 2011; Hay et al.
2010) extended the concept of POD snapshots to include
derivatives of the state variables with respect to system input
parameters.
In order to avoid generating additional sampling points,
an effort has been made to develop interpolating strategies
between the ROMs. Missoum (2008) used Lagrange interpolation to control the relative contributions of individual
modes in order to perform a random-field-based probabilistic optimization of a tube impacting a rigid wall. Mathelin
and Le Maitre (2010) proposed polynomial transformation
130
M. Xiao et al.
(1)
k=m+1
M
(3)
i=1
(k)
= Argmin
(4)
v v(k) 2
k=1
L ()
In Xiao et al. (2010), (4) was exploited in order to introduce an additional constraint requiring that some functions
(objective or optimization constraints) have the same value
whether evaluated using an exact or an approximate field,
thus
(5)
J v = J (v)
at least for the snapshot sampling points. Our approach then
was to express the approximate snapshot using a standard
POD basis (4)
v (k) = v + (k)
(6)
(k) = Argmin
v (k) v(k) 2
=const
k=1
L ()
Enhanced POD projection basis with application to shape optimization of car engine intake port
(7)
(8)
=
CT ,m
0
c(k) CT v
(9)
with the matrix W representing both the interpolation
operator over the reference grid and the numerical integration of the quantity of interest J (v), M a standard
finite element mass matrix and the Lagrange multipliers.
This approach was successfully applied to several problems using a bi-level approximation, coupling POD with
kriging/RBF/Diffuse Approximation and we showed that
satisfying the interpolation property (5) was critical in
a multi-objective problem. There remain however some
difficulties, since the existence of a solution to (9) cannot
be guaranteed in a general case.
(10)
(k)
(k)
= Argmin
(11)
v v 2
T =I
k=1
L ()
and with the coefficients obtained by projecting the snapshots on the modified basis in the same way as in standard
POD (2)
(k) = v(k) , , k = 1 . . . M.
(12)
The main idea behind this algorithm is based on the observation that the necessary and sufficient condition (NSC) for
the existence of a solution to (11) is that for any velocity
field v in the high fidelity model space, there should exist a
field w in the reduced order space such that
C v=C w
T
(13)
(15)
(16)
where
is the matrix of first m p eigenvectors of the
M
T
covariance matrix
v(k) v Q2 Q2T v(k) v of the
k=1
131
the constrained POD has analogous convergence properties to the standard POD,
132
M. Xiao et al.
T
=
T
=
(y)
(x x0 ) vz d
(18)
v
d
S
Q
=
Q
=
(y)
(19)
Enhanced POD projection basis with application to shape optimization of car engine intake port
(a)
(b)
133
(c)
Fig. 3 POD and CPOD2 error for the velocity field (a) and for objective functions T (b) and Q (c)
The first p constrained modes do not depend on the velocity field as they are obtained by a QR decomposition of the
constraint matrix C without considering the snapshots. The
matrix C size is r n, where r = 2 is the number of constraints and n is the number of nodes on the reference grid
in the zone of interest. In our case, the QR decomposition
gives p = 2 and the corresponding velocity fields are illustrated in Figs. 4 and 5. We observe readily, that the first
mode (Fig. 4) corresponds to the pure rotation of the fluid
with no mass flow, while the second mode (Fig. 5) depicts a
uniform flow with no vorticity.
Figure 6 compares the first POD modes (left) with their
constrained counterparts p + 1. . . m (right).
As is usual with the standard POD, the physical interpretation of these modes is not straightforward. However,
we may make two observations here. First, there is an
overall similarity between the corresponding vector fields.
When comparing the CPOD2 modes with the POD ones,
the difference is mainly due to the subtraction of the pure
mass flow and the pure vorticity components. The second
observation is quite obvious when looking at the underlying
math: for the CPOD2, the mass flow and vorticity are null
134
M. Xiao et al.
POD mode 1
constrained mode 3
T = 0.806, Q = -0.138
T= -2.7 *10-14, Q = 0
POD mode 2
constrained mode 4
T = -0.406, Q = 5.0*10-3
T = 2.4*10-14, Q = 0
POD mode 3
constrained mode 5
T= -0.068, Q = 1.87*10-3
T = 1.948 *10-14, Q = 0
Enhanced POD projection basis with application to shape optimization of car engine intake port
135
(a)
Solution
(b)
T/T_max
Q/Q_max
Y1
y2
y3
y4
y5
y6
1.0406
0.8448
1.0
0.2699
0.5721
0.7421
0.8981
0.1108
0.4635
0.2541
1.0
1.0
1.0
0.4682
0.9482
0.8717
1.0
0.0226
0.5343
0.0172
1.0
0.2029
1.0016
1.0
1.0
1.0
1.0
1.0
1.0
number
(a) point 1
(b) point 2
(c) point 3
(d) point 4
Fig. 8 Optimal velocity fields for four points of the Pareto set. a point 1. b point 2. c point 3. d point 4
136
a simple interpolation of coefficients for the optimal design parameters values followed by a linear combination
of the already available modes, we may use a higher
value for m. The velocity fields reconstructed with the full
set of m = 146 modes for the four solutions are shown in
Fig. 8ad.
We see that for point 1 corresponding to the lowest values of the valve and short turn angles, the minimal bowl
height and a maximal port diameter allow for maximizing
T by adjusting the runner length and the downdraft angle
giving a vorticity dominated flow (Fig. 8a), resulting in an
environmentally friendly design. The power of the engine
is however too low due to an insufficient mass flow. Conversely, for the optimal point 4, corresponding to maximal
values of all the parameters, a high value of Q is obtained
and the field (Fig. 8d) is dominated by the mass flow. This
solution (which belongs to the Pareto set of the initial DOE)
has to be discarded due to the excessive CO2 emission rate.
Points 2 and point 3 (Fig. 8b, c), provide a compromise
between the two objectives, both with a maximal downdraft
angle. Once again, the increase in port diameter and runner
length allows us to adjust the vorticity while a maximum
value of the bowl height permits the tuning of the valve
angle, short turn radius, the port diameter and the runner
length for higher engine power.
5 Conclusions
The numerical results show that with the standard POD, the
reconstruction error may lead to unreliable results, especially in multi-objective optimization problems. This is
because the basis truncation produces an error in the calculation of the functionals of the velocity field. The approach
proposed in this paper overcomes this limitation by modifying the orthogonal modes, by imposing the conservation of
integral quantities.
Two conclusions may be drawn. First, realistic Pareto
sets are obtained more easily than with the original POD
approach. The second advantage is that the number of
modes needed for optimization is reduced to the number
of objective functions: two in the case presented, with no
impact on the quality of estimation of the quantities of
interest. Moreover, the final solutions may be reconstructed
using a high number of modes (much higher than normally
used in POD), providing a finer physical insight into the
optimized solutions.
M. Xiao et al.
In terms of future prospects, even though the use of a constrained approach within local POD approximations seems
straightforward, further research is needed to investigate
dedicated manifold interpolation techniques. An extension
of the approach to problems with energy-type quadratic
constraints is also currently under investigation.
Acknowledgments This work was supported by the Agence Nationale
de la Recherche in the scope of OMD2 project ANR-08-COSI-007.
References
Amsallem D, Cortial J, Carlberg K, Farhat C (2009) A method for
interpolating on manifolds structural dynamics reduced-order
models. Int J Numer Methods Eng 80:12411258. doi:10.1002/
nme.2681
Berkooz G, Holmes P, Lumley JL (1993) The proper orthogonal
decomposition in the analysis of turbulent flows. Ann Rev Fluid
Mech 25:539575
Carlberg K, Farhat C (2011) A low-cost, goal-oriented compact
proper orthogonal decomposition basis for model reduction of
static systems. Int J Numer Methods Eng. doi:10.1002/nme.3074
DAKOTA Reference Manual. http://dakota.sandia.gov/licensing/votd/
html-ref/index.html, last accessed on 26 August 2011
Degroote J, Vierendeels J, Willcox K (2010) Interpolation among
reduced-order matrices to obtain parameterized models for
design, optimization and probabilistic analysis. Int J Numer
Methods Fluids. doi:10.1002/fld.2089
Eddy J, Lewis K (2001) Effective generation of Pareto sets using
genetic programming. In: Proc. of DETC01 ASME 2001 design
engineering technical conferences and computers and information
in engineering conference, Pittsburgh, PA, 912 September
Filomeno Coelho R, Breitkopf P, Knopf-Lenoir C (2007) Model reduction for multidisciplinary optimizationapplication to a 2D wing.
Struct Multidiscipl Optim. doi:10.1007/s00158-007-0212-5
Filomeno Coelho R, Breitkopf P, Knopf-Lenoir C, Villon P (2008) Bilevel model reduction for coupled problemsapplication to a 3D
wing. Struct Multidiscipl Optim. doi:10.1007/s00158-008-0335-3
Hay A, Borggaard J, Akhtar I, Pelletier D (2010) Reduced-order models for parameter dependent geometries based on shape sensitivity
analysis. J Comput Phys. doi:10.1016/j.jcp.2009.10.033
Mathelin L, Le Maitre O (2010) Equation-free model reduction
for complex dynamical systems. Int J Numer Methods Fluids.
doi:10.1002/fld.2219
Missoum S (2008) Probabilistic optimal design in the presence of
random fields. Struct Multidiscipl Optim. doi:10.1007/s00158007-0126-2
Weickum G, Eldred M, Maute K (2009) A multi-point reduced-order
modeling approach of transient structural dynamics with application to robust design optimization. Struct Multidiscipl Optim.
doi:10.1007/s00158-008-0309-5
Xiao M, Breitkopf P, Filomeno Coelho R, Knopf-Lenoir C,
Sidorkiewicz M, Villon P (2010) Model reduction by CPOD and
Krigingapplication to the shape optimization of an intake port.
Struct Multidiscipl Optim 41(4):555574