from a finite element model to system level simulation by means of
TRANSCRIPT
From a finite element model to
system level simulation by
means of model reduction
MUMPS Users Meeting 2010, Toulouse
Evgenii Rudnyi, CADFEM GmbH
http://ModelReduction.com
Content
System level simulation and
compact modeling
Model order reduction
Case studies with MOR for
ANSYS
Experience with MUMPS
solver
1
Device and System Level Simulation
How to make a model at system level from that at device level?
2
Thermal Domain
Electrical Domain MechanicalDomain
0
R1 R2 R3
MASS_ROT1
DR1
SINE2
SINE1
A
A
A
IN_A
IN_B
IN_C
OUT_A
OUT_B
OUT_C
I_motSimplorer4
A
B
C
N
ROT1
ROT2
Induction_Motor_20kW
Q
Ambient
P1
P2
P3
P4
P5
P6
P7
P8 P
9
P1
0
P1
1
P1
2
P_
RE
F
z_um z_vm z_wm
z_wpz_vpz_up
+
V
VM311 R6 R5 R4
E1
RZM
DR1
RZM
0.00 2.50 5.00 7.50 10.00 12.50 15.00 17.50 20.00Time [ms]
-500.00
-250.00
0.00
250.00
500.00
Y1 [
V]
Curve Info rms
R1.VTR 230.9405
R4.VTR 313.3812
0.00 2.50 5.00 7.50 10.00 12.50 15.00 17.50 20.00Time [ms]
1475.00
1485.00
1495.00
1500.00
MA
SS
_R
OT
1.O
ME
GA
[rp
m]
0.00
50.00
100.00
150.00
180.00
MA
SS
_R
OT
1.P
HI
[deg]
Curve Info
MASS_ROT1.OMEGATR
MASS_ROT1.PHITR
Compact Modeling: Transistor Compact Model
Figure from J. Lienemann, E. B. Rudnyi and J. G. Korvink. MST MEMS model
order reduction: Requirements and Benchmarks. Linear Algebra and its
Applications, v. 415, N 2-3, p. 469-498, 2006.
Compact Thermal Models
Looks understandable – but how to do it in practice?
Figure from the book “Fast Simulation of Electro-Thermal MEMS: Efficient Dynamic
Compact Models.” Springer, 2006.4
Comparison: Electrical vs. Thermal
Thermal phenomena are much more distributed, it is hard to lump
them.
Figure from the book “Fast Simulation of Electro-Thermal MEMS: Efficient Dynamic
Compact Models.” Springer, 2006.5
Electrical flow
cond / ins 108
Heat flow
cond / ins 102
Conductor
Insulator
Content
System level simulation and
compact modeling
Model order reduction
Case studies with MOR for
ANSYS
Experience with MUMPS solver
6
7
Model Order Reduction
Relatively new technology:
It is not mode superposition;
It is not Guyan reduction;
It is not CMS.
Solid mathematical background:
Approximation of large scale
dynamic systems
Dynamic simulation:
Harmonic or transient simulation
Industry application level:
Linear dynamic systems
8
Linear Model Order Reduction
Linear Dynamic
System, ODEs
Control Theory
BTA, HNA, SPA
O(N3)
Moment Matching
via Krylov subspaces.
Iterative, based on
matrix vector product.
Low-rank
Grammian,
SVD-Krylov
one-side
Arnoldi process
double-side
Lancsoz algorithm
MOR for ANSYS
9
Model Reduction as Projection
Projection onto low-
dimensional subspace
How to find subspace?
Mode superposition is not the best way to do it.
zx V
x V
z
=
uxx BKE
uzz BVKVVEVV TTT
E
Er
Hankel Singular Values
Dynamic system in the state-
space form:
Lyapunov equations to
determine controllability and
observability Grammians:
Hankel singular values (HSV):
square root from eingenvalues
for product of Grammians.
Cxy
BuAxxH(s) C(sI A) 1B
AP PAT BBT 0
AT Q QA CT C 0
i i (PQ)
10
Global Error Estimate
Infinity norm
Global error for a reduced model
of dimension k
Model reduction success
depends on the decay of HSV.
Log10[HSV(i)] vs. its number.
From Antoulas review.
H(s) ˆ H (s)
maxs abs(H(s) ˆ H (s))
H(s) ˆ H (s)
2( k 1 n )
11
12
Implicit Moment Matching
Padé approximation
Matching first moments for the
transfer function
Implicit Moment Matching:
via Krylov Subspace
0
0 )( i
i ssmH
mi mi,red, i 0, ,r
BKsEsH1
)(
0
0, )( i
redired ssmH
s0 0
V span{ (K 1E,K 1b)}
uxx BKE
13
Second Order Systems
Cxy
BuKxxExM
Ignore the damping
matrix during MOR:
Proportional Damping
Transform to 1st
order system.
Use Second Order
ARnoldi.
default-1 -2
14
Error Indicator
Key question :
What is a suitable order of the reduced
system for a desired accuracy?
“Rule of thumb”: r = 30-50
Proposed engineering approach:
comparison of reduced systems of order
r and r + 1
System of
n ODEs
Reduced
System of
r << n ODEsMOR
r
User
r = ?
error estimate:
a way to automate MOR
15
Convergence of Relative Error
)(
)()()(
sH
sHsHsE r
r
)(
)()()(ˆ 1
sH
sHsHsE
r
rr
r
True error:
Error indicator:
E r(s)ˆ E r (s)
1,5mm
Main result:
16
R R(T ) R0 (1 T T 2 )
)(
)(),(
2
TR
tUFTtQFTKTC
Input function does not
participate in MOR
Q(t,T ) Q(TN (t)) Q(V * z)
Nonlinear input
Control temperature is a
single node temperature TN
Uheater=14V
),( * zVtQFVzKVVzCVV TTT
Nonlinear Input
Content
System level simulation and
compact modeling
Model order reduction
Case studies with MOR for
ANSYS
Experience with MUMPS solver
18
19
MOR for ANSYS: http://ModelReduction.com
FULL files
Linear Dynamic
System, ODEsCxy
BuKxxExM MOR Algorithm
Small dimensional
matrices
Current version 2.5
ANSYS Model
Simulink,
Simplorer, VerilogA,
…
Implicit Moment Matching
Take matrix and vector
corresponding to a given
expansion point.
Compute the orthogonal basis by
the Arnoldi process.
Project the original system on this
basis.
One can prove that this way the
reduced model matches k
moments.
Multiple inputs:
Block Krylov Subspace;
Block Arnoldi;
Superposition Arnoldi.
s0 0 P A 1E r A 1b
V span{ (A 1E,A 1b)}
V span{r,Pr,P2r, Pk 1r}
20
Treating Matrix Inverse
The Arnoldi process requires
matrix vector product.
Yet, we have a matrix inverse.
Instead solve a system of linear
equations.
This is the biggest computational
cost:
Number of vectors x time for
linear solve.
Direct solver has strong
advantage.
vi 1 A 1Evi
ui 1 A 1ui
Aui 1 ui
The only right hand side is
different.
Can be used to speed it up.
21
MOR for ANSYS Timing: MOR as Fast Solver
Reduced model of dimension 30
22
Dimension,
DoFs
nnz MOR Time
/ANSYS
static
4 267 20 861 1.4
11 445 93 781 1.8
20 360 265 113 1.7
79 171 2 215 638 1.5
152 943 5 887 290 2.2
180 597 7 004 750 1.9
375 801 15 039 875 1.7
Chip and its Model in Workbench
Two power MOSFET transistorsThe example is from T. Hauck, I. Schmadlak, L. Voss, E. B. Rudnyi. Electro-Thermal Simulation of Multi-channel Power
Devices: From Workbench to Simplorer by means of Model Reduction. NAFEMS, Seminar: Multi-Disciplinary
Simulations - The Future of Virtual Product Development, 2009
23
Thermal Impedance and Comparison with ANSYS
ANSYS: about 300 000 DoFs, Reduced model: 30 DoFs
The difference is less than 1%
Timing: 60 timesteps is about 30 min in ANSYS
25
Transient Turn-on of an Automotive Light-Bulb
28
0.00 10.00 20.00 30.00 40.00 50.00Time [ms]
0.00
5.00
10.00
15.00
20.00
25.00
30.00
35.00
E2
.I [A
]
Curve Info
SL1='1' SL2='0' SL3='0'
SL1='1' SL2='1' SL3='0'
SL1='1' SL2='1' SL3='1'
Ansoft LLC FET_Light_Bulb_P21W_2Lamp Currents at turn-on
1 lamp
2 lamps
3 lamps
0.00 10.00 20.00 30.00 40.00 50.00Time [ms]
250.00
300.00
350.00
400.00
450.00
500.00
550.00
TH
M1
.T [ke
l]
Curve Info
THM1.TSL1='1' SL2='0' SL3='0'
THM1.TSL1='1' SL2='1' SL3='0'
THM1.TSL1='1' SL2='1' SL3='1'
Ansoft LLC FET_Light_Bulb_P21W_2Junction Temperature at turn-on
3 lamps
2 lamps
1 lamp
MOR for ANSYS in turbine dynamics
Felix Lippold / Dr. Björn Hübner
Voith Hydro Holding
ANSYS Conference & 27. CADFEM Users Meeting, 18 - 20
November 2009, Congress Center Leipzig.
MOR for ANSYS in turbine dynamics | CADFEM2009 | Felix Lippold | 2009-11-19 | 30
Finite-Element model used in harmonic analysis
Linear-elastic SOLID elements for runner structure.
Acoustic FLUID elements for surrounding water.
Complex load vectors for rotating pressure fields.
ModellingAcoustic-structure coupling
MOR for ANSYS in turbine dynamics | CADFEM2009 | Felix Lippold | 2009-11-19 | 31
ModellingCoupled acoustic-structure equations
Equations of motion
)(A
S
A
SAS
A
S
ASA
Stq
b
b
p
u
K
KK
p
u
E
E
p
u
MM
M
Formally equivalent to
)(tqbxKxExM
But:
non-symmetric matrices
non-proportional damping matrix
fluid damping EA only for impedance boundaries
MOR for ANSYS in turbine dynamics | CADFEM2009 | Felix Lippold | 2009-11-19 | 32
Axial displacement@centre trailing edge – no damping
Amplitude spectrum (Dim=100) Deviation related to ANSYS results
ResultsNo damping
MOR for ANSYS in turbine dynamics | CADFEM2009 | Felix Lippold | 2009-11-19 | 33
Pressure distribution on runner for f = 40 Hz
Ansys (90 000 DOFs) Reduced (Dim=100)
REAL part of pressure
solution
Deviation < 0.2%
IMAG part of pressure
solution
Deviation < 0.2%
ResultsVerification of reduced order method
Hard Disk Drive Actuator/Suspension System
Michael R Hatch, Vibration Simulation Using MATLAB and ANSYS
34
Model Reduction
3352 elements
7344 nodes
21227 equation
400 frequencies takes about 12
min
MOR takes only 3 s
Comparison for head
ANSYS
MOR 30
MOR 80
35
Content
CADFEM
System level simulation and
compact modeling
Model order reduction
Case studies with MOR for
ANSYS
Experience with MUMPS solver
39
www.msnbc.msn.com/id/12409082/
History of solvers in MOR for ANSYS
MUMPS
Fortran 90 – Activation barrier
TAUCS – symmetric matrices
UMFPACK – unsymmetric
matrices
2006 – HPC-EUROPA
SGI Origin 3800 (teras)
Micro-FEM bone models
Bert van Rietbergen
2007 – MUMPS in MOR for
ANSYS
40
Voith Hydro Timing
2 x quad-core Intel X5560 (Nehalem) with 48 GB under Windows
Shared memory
41
nnz dim
K 91627731 1292663
M 31982656 1292663
27.7 Gb memory
Factorization is 287 s
250 vectors is 1500 s
6 s pro vector
nnz dim
K 99934728 1495341
M 35774706 1495341
20.1 Gb memory
Factorization is 243 s
200 vectors is 1300 s
6.5 s pro vector
How MUMPS is used in MOR for ANSYS
Shared memory (serial)
No parallel
Linux and Windows versions (Visual C + Intel Fortran on Windows)
Used to have an Altix version
Intel MKL BLAS
Used to work with ATLAS
Two versions:
32-bit integers
64-bit integers
Mostly in-core, sometimes out-of-core
Symmetric positive definite, symmetric indefinite, unsymmetric
matrices42
Sharing Experience
Oberwolfach Model Reduction Benchmark Collection
http://portal.uni-freiburg.de/imteksimulation/downloads/benchmark
Trabecular Bone Micro-Finite Element Models
http://MatrixProgramming.com
Compiling MUMPS under Microsoft Visual Studio and Intel Fortran with
GNU Make
Sample code in C++ to run MUMPS
Working with Microsoft Visual Studio
Distribution the code with manifests
-02 does not turn safe iterators off (_SECURE_SCL=0)
32-bit code has some problems with memory fragmentation
Using a tree to assemble a matrix
43