1
Parameterized Mode Truncation Technique Parameterized Mode Truncation Technique
in MEMS Designin MEMS Design
V. Kolchuzhin, W. Dötzel and J. Mehner
Berlin, 4 Dec 2010
2
Parameterized Mode Truncation Technique in MEMS Design
• Introduction and Motivation
• Parametric FE technique and pROM
• Numerical Examples
• Conclusion and Outlook
Outline
3
What is the MEMS ?
KD M
V hu
Fel
F
Fu
uCVuKuDuM +
∂
∂=++
)(
2
2
&&&
VuCVuCtd
Qdi && )()( +==
Vibrationssensor TUC/ZfM
200um
2n
d o
rde
r O
DE
4
MEMS Inertial sensors: application
Microsoft Freestyle Pro Gamepad(ADXL202)
HDD Active Protection System, 2003
InvenSense: Angle camera shake(ITG-3200)
Guidance and Navigation
Photo: Claudia Dewald/iStockphoto
Ford: Curve Control
Bosch: ESP
iPhone4STM: L3G4200D
Samsung YP-S1 TicToc
Wii Remote (ADXL330), 2005
SegwayScooter
Nike+iPod
5
MEMS Design
TCAD SPICE, Simulink, VHDL-AMS,Verilog-A
Componentsimulations
System simulations
Process simulations
Lumped elements
Processsequence
RedesignRedesign
Behavioral modelBehavioral modelRequirementsRequirements
FDM, BEM, FEM, FVM
Mask design
L-EDIT, AutoCADMATLAB, MathCAD MATLAB, ANSOFTANSYS
D K
CM
Low-levelsimulations
Sentaurus Topography
Ma
cro
mo
delin
g
To Fab Virtual prototype
REM image
ANSYS, MATLAB
ROM, MOR
S
+
Nodal ports
Voltage ports
Modal ports
Auxillary ports(element load)
S
+
S
+
ROM144Electrostatic-Structural Reduced Order Model
q1
e1
u1+
V1
Pa
e3V2
q2q3
ROM140Fluid-Structural Interactions
Acel Pres
Package model
ROMq4
ASIC
@ master node
THM1
S
+
displ
contact
S
+ V2V1
ΘT1
u1-
r
u1-u1+
Curvature
R1
C1
Epol
The MEMS designing is a very challenging and interdisciplinary task!
Subsystem
6
MEMS Simulations
),(),(2
2
txfx
R
t
Rtxp =
∂
∂+
∂
∂
PDE ODE
>105 DOF x n time steps DOF< 10
mq′′+dq′+kq=f(t)
BEM
FDM
FEMReduced order (2001-…)
SPICE, Simulink, VHDL-AMS,Verilog-A
Physical level System level
FDM, BEM, FEM, FVM
MATLAB, ANSOFTANSYS
ANSYS, MATLAB
ROM, MOR
S
+
Nodal ports
Voltage ports
Modal ports
Auxillary ports(element load)
S
+
S
+
ROM144Electrostatic-Structural Reduced Order Model
q1
e1
u1+
V1
Pa
e3V2
q2q3
ROM140Fluid-Structural Interactions
Acel Pres
Package model
ROMq4
ASIC
@ master node
THM1
S
+
displ
contact
S
+ V2V1
ΘT1
u1-
r
u1-u1+
Curvature
R1
C1
Epol
Subsystem level
Online-coupling
(1995-2000) Matlab/Simulink
VHDL-AMS
Verilog-AMS
gen
era
lized
Kir
ch
off
ian
law
s
PSPICE
-level-physics
-scale-material
multi
mq′′+dq′+kq=f(t)
dq′+kq=f(t)
Lumped elements
Electrical equivalent
Modal decomposition
Krylov based method
Control theory methods
1. Loading design parameters2. Material design parameters3. Geometrical design parameters 4. Boundary conditions
mechanicalelectrofluidicthermal
MFM
FVM
CVM
Subsystem
?
7
Parameterized Mode Truncation Technique in MEMS Design
• Introduction and Motivation
• Parametric FE technique and pROM
• Numerical Examples
• Conclusion and Outlook
Outline
8
Taylor Series and HODM
1715
Brook Taylor
"Taylor's Theorem"
1993
The Intel Pentium supports 8 transcendentalfunctions through direct execution of microcodeusing polynomial approximation techniques.
…i
n
i
ipppf
ipf )()(
!
1)( 0
0
)(
0 −≈∑=
High order sensitivity analysis:
Transient analysis[M]a+[D]v+[K]u=F
Modal analysis[K-LM]X = 0
Static analysis[K]u=F
Harmonic analysis(-Ω2[M]+iΩ[D]+[K])u=F
Taylor expansion
Rational approximation
PDE → DAEMatrix EquationDifferentiation
Mode truncation
method
−= −
=
− ∑ )()(
1
)(1)( )()()()()( inin
i
i
n
nnppCppp uKFKu
−= −
=
− ∑ )()(
1
)(1)( )()()()()( inin
i
i
n
nnppCppp uAFAu
=+
=+−+−++
+++
)()1(
)()1()1()1(
)1(
)1()1()1(n
L
n
i
T
i
n
xi
n
i
n
ii
n
i
n
LnLnn
SMxx
SMxMx Kx
FEM
stiffnesspermittivity
displacementpotential
forcecharge
u = F = K =
9
II. Parametric Method
Parametric FE mesh
2. 3.
4.n
n
p
u
∂
∂
Extraction
Assembling
design parameter
obje
ctive f
unction
1st
2nd
5th TS4th TS
in
i
i
ppi
pfpf )(
!
)()( 0
0
)(
0 −=∑=
n+1
][K
5
5][
p
K
∂
∂
3
3][
p
K
∂
∂ 4
4][
p
K
∂
∂
2
2][
p
K
∂
∂
p
K
∂
∂ ][
3rd TS
2. 3. 4.
FE Technique
2. 3. 4.
I. Ordinary Data Sampling Method
][ 1 FKu −=
FE mesh
Global matrix Solution
for i=1:n
Assembling
1.
CAD model
)( ipf
design parameter
computation computation
time?time?
accuracy?accuracy?
pi
∑=el
kK ][][
end
pi
…
obje
ctive f
unction
...)( 2
210 +++= papaapx
...)( 2
210 +++= pbpbbpy
...)( 2
210 +++= pcpccpz
10
Data Structure
Arrays of polynomials terms (Taylor series coefficients) of 3 variables and 4th order
Pyramid array (Pascal tetrahedron)
∑∑∑∑====
++++====
n
j
j
j
jKR
1
)(µ
size = 35
2D array
+=
p
p
n
nordersize np = 3
order = 4
…
ji
000 R(µµµµ)
100
010
103013
004
0
1
2
32
33
34
µµµµ R(µµµµ)j
…
i
k
y
y2
y3
y4
x
x2
x3
x4
z2yzxz
xy
z3
y2z
yz2
x2z
yz2
xy2
xyz
x2y
x2y2
xy3x3y
y2z2x2z2
y3z
yz3xz3
x3zxyz2
xy2zx2yz
1
z4
z
µ 1
µ 3
µ 2
3D array
11
Geometrically Parameterized Finite Element
∑=ngp
i
iii
T
i wJBDBk][
ngp – number of Gauss points
n
i
n
n
i
n
n
i
n
p
z
p
y
p
x
∂
∂
∂
∂
∂
∂,,
Nodal coordinates and its derivatives at p0
AD:
[f]=[u][v]
AD:
[f]=[u].v
AD:
[f]=det[u]
[f]=inv[u]
n
n
p
k
∂
∂ ][
Solid
element
),,(][ ZYXfJ =
])[(][ JinvfB ====
)(
)(
)(
pfZ
pfY
pfX
=
=
=|J|, (tuple)
[J-1], (3,3,tuple)
[B] = [J-1] [dN]
(3,8,tuple) = (3,3, tuple) . (3,8,tuple)
[BD] = [B]T[D]
(8,3,tuple) = (8,3, tuple) . (3,3)
ngp = 8 wi, ξi, ηi, ζi
Ni = (1 ± ξi)(1 ± ηi)(1 ± ζi) (8 x ngp)
dNi=[Ni /dξ; Ni /dη; Ni /dζ;] (24 x ngp)
[J] = [dN] . [XYZ]
(3,3,tuple) = (3,8, ngp) . (8,3,tuple)
[BDB] = [BD].[B]
(8,8,tuple) = (8,3, tuple) . (3,8,tuple)
[BDBJ] = [BDB]. |J|
(8,8,tuple) = (8,8, tuple) . (tuple)
[ke] = [ke] + [BDBJ].wi
(8,8,tuple) = (8,8, tuple) . (ngp)
Nodal Coordinates:
n
∑=ngp
i
iii
T
i wJBDBk][∫∫∫=V
TdxdydzBDBk][
=
)()()(
)()()(
)()()(
)()()(
)()()(
)()()(
)()()(
)()()(
][
PPP
OOO
NNN
MMM
KKK
JJJ
III
pzpypx
pzpypx
pzpypx
pzpypx
pzpypx
pzpypx
pzpypx
pzpypx
XYZLLL
1 2
34
5 6
78
I
K
J
ON
P
M
L
ζ
0
η
ξ
M(p0)
N(p0)
O(p0)P(p0)
I(p0)
J(p0)
K(p0)
K(p)
O(p) p
y
z
x
i:=i+1
Isoparametrictransformation
Cell
12
Generation of MEMS Macromodels Using HODM
Test loads:couple domain run
),,( zyxu∑≈ ),,(),,( zyxwzyxu iiφ
0.0343.382
………
0.09420.067
18.2758.553
81.5834.031
Contribution
factor [%]
Frequency,
kHz
Mode #
Computation of basis functions:
iw
),,( zyxiφ
i
Modal analysis
Mode 1 Mode 3 Mode 7
),,( zyxiφ
E1 conductor 2
conductor 1
conductor 3
∑=
=3
1
),,()(i
iiii zyxqwqu φ
Static structural domain Electrostatic domain
Mode i
),( qkii Ω),( qcii Ω
Para
mete
rs e
xtra
ctio
n
Fluid domain
12C 13C 23C1 3
2
Strain energy Capacitance Damping coefficient
n
i
n
q
qqqW
∂
∂ ),,( 731
ROM database
iω
ji
ijqq
Wk
∂∂
∂=
2
n
i
n
q
P
∂
∂n
i
n
q∂
∂ ϕn
i
n
q
u
∂
∂
1ω3ω 7ω
0=iqiωω =0
Static parametric analysis at Static parametric analysis at Harmonic parametric analysis at0=iq
1. 2.
4.1. 4.2. 4.3.
( )∑ −∂
∂+=
∂
∂++
r
sk
i
ksq
i
SENEiiiiii tVtV
q
pqCptf
q
pqWtqpmppqtqpm )()(
),(
2
1),(
),()()()(),(2)()( &&& ωξ
Para
metr
ic r
uns
3.
n
ij
nc
ω∂
∂
n
i
r
n
q
qqqC
∂
∂ ),,( 731
13
Parameterized Mode Truncation Technique in MEMS Design
• Introduction and Motivation
• Parametric FE technique and pROM
• Numerical Examples
• Conclusion and Outlook
Outline
14
Application: Parametric Linear Static Analysis of the FFB
L
B
H
y
z
x
FE stiffness matrix
FE-mesh
Load: Pressure=12.5 kPa
120x30x2mkm
8415×8415
Elements = 2000Nodes = 2805
Size(Kfull) = 540.25Mb
Size(Knz) = 26.37Mb
nz = 539081
E = 169 GPa, Poisson’s ratio ν = 0.066, and mass density ρ = 2.329 g/cm3
Design parameters: L, B, H
15
Analytical Models of FFB
L-2Resonant frequency
L5Bending energy
L2Maximum stress
(under uniform pressure P)
L4Center deflection
(under uniform pressure P)
L-33rd order stiffness coefficient
L-3Stiffness
(point excitation at the center)
LMass
L-dependenceAnalytical modelProperty
ρLBHm =
ρπ
E
L
Hnf i
i 2
2
3=
3
3
425.10L
EBHK =
KB
2
767.0=α
PEH
L3
4
32=δ
PH
L2
2
2=σ
3
24
96L
EIW
δπ=
12
3BH
I =
16
Application: Results of Parametric Linear Static Analysis
4th order TS
100 200 300 400
0.0
0.5
1.0
50 100 150 200
Evaluationpoint
0.0
0.5
1.0
100
FE solutions
50 100 150 200
0
50
2 4 6 8
0
50
100
0
Padé app.
4th order TS
1st order TS
50 100 150 200
0
20
40
60
Width, µm
4th order TS
0 2 4 6 8
0
20
40
60
Thickness, µm
3rd order
Padé app.
0 2 4 6 8
0.0
0.5
1.0
2nd order TS
Padé app.
100 200 300 400
0
50
100
Evaluationpoint
2nd order TS
100 200 300 400
0
20
40
604th order TS
Length, µm
Evaluationpoint
Str
ain
energ
y, p
J
Str
ain
energ
y, p
J
Str
ain
energ
y, p
J
Max. eqv. str
ess, M
Pa
Max. d
ispla
cem
ent, µ
m
Max. d
ispla
cem
ent, µ
m
Max. d
ispla
cem
ent, µ
mM
ax. eqv. str
ess, M
Pa
Max. eqv. str
ess, M
Pa
∝∝∝∝ H-3
∝∝∝∝ H-2
∝∝∝∝ H-3
∝∝∝∝ B0
∝∝∝∝ B0
∝∝∝∝ B1
∝∝∝∝ L4
∝∝∝∝ L2
∝∝∝∝ L5
17
Application: Parametric Linear Static Analysis
eff
icie
ncy
of
HO
DM
1 2 3 4 5 6 7 8 9 100
100
200
300
400
Number of sampling points / Number of derivatives
Com
puta
tion
tim
e,
sec
Parametric FE technique (HODM)
FE runs for data sampling
∝ DOFsone FE run
one FE run ×5
Benefits of novel approach compared to ordinary data sampling procedures become obvious:
• for multi-parameter problems (np > 3)• for large problems (dof > 104).
The grid-like data sampling process is expected to grow as nk, where n is the number of parameters and k is the number of states for each parameter.
18
Application: Parametric Electrostatic Analysis
2.580
2.581
2.582
2.583
2.584
2nd order TS
2.579
p4, degree
-5 0 5
1st order TS
Padé app. [2/2]
2.5
3.0
3.5
4.0
4.5
-2 0 2p2, µm
Evaluationpoint
Evaluationpoint
p5, degree
4th order TS
0 2-2-4 42.40
2.45
2.50
2.55
2.60
2.65
2.70
Padé app. [2/2]
p6, degree
0 2-2
2.5
3.0
3.5
4.0
4.5
4th order TS
2nd order TS
2.40
2.45
2.50
2.55
2.60
2.65
2.70
Capacitance, fF
p1, µm-10 -5 0 5
4th order TS
2nd order TS
2 40-2 6p3, µm
3rd order TS
2.40
2.45
2.50
2.55
2.60
2.65
2.70
Levitation
p1 – motion in operating directionp2 – horizontal shift in y-directionp3 – vertical shift in z-directionp4 – x-rotationp5 – y-rotationp6 – z-rotation
Conductor 3SubstrateConductor 1
Moving finger
Conductor 2 Fixed fingers
Vc
Capacitance stroke function of the movable finger with regard to six degrees of freedom C=f(p1,…, p6)
Linear electrostatic analysis
FE model
Solid model
C(p0)=2.57982e-15 F
=
0
c
e
c
ee
QV
KK
KK
ec
eccc
ϕ
p2
p3
p1
19
Application: Parametric Electrostatic Analysis
Capacitance, fF
p1 – motion in operating direction
Conductor 3SubstrateConductor 1
Moving finger
Conductor 2 Fixed fingers
Vc
Comparison between mesh-morphing and remesh approaches
y
x
y
x
y
x
a) Initial FE-mesh p1= 0 µm
b) Mesh-morphing p1= −18 µm
c) Remesh p1= −18 µm
Solid model
p1
-20 0
2.0
2.5
3.0
p1, µm
4th order TS
FE-remesh
FE-morphing
20
Generally, HOD FE technique is limited to
moderate parameters variations ±±±±15% due to the mesh-morphing procedures!
• Multi-Point Taylor Expansions• BEM
20
Application: Discrete Analysis of Perforated Beam
p2: element component 2
p3: element component 3
p1: element component 1
p2: element component 2
p3: element component 3
FE-model
0 111
12
13
14
15
16
Discrete parameter p
Str
ain
energ
y, p
J
3rd order TS
1st order TS
Evaluation point
0 150
55
60
65
Discrete parameter p
4th order TS
p
3rd order TS
0 1Discrete parameter p
-0.50
-0.45
-0.40
Evaluationpoint
Estimatedpoint
5th order TS
2nd order TS
4th order TS
y
z
x
Ma
x. e
qv.
str
ess,
MP
a
Ma
x. d
ispla
cem
ent,
µm
pEpE 0)( =
Linear Static Analysis
Perforated beam with circular holes having a radius of 25 µm
Young modulus:
Discrete parameter p corresponds to sets of elements in the model that can be turned 1 or 0.
Design parameter p ≡≡≡≡ number of holes
0,0,0 1,0,0 0,1,0 1,1,0 0,0,1 1,0,1 1,0,0 1,1,1
21
∂
∂=
∆
∂
∂−
∆∇
g
z
tPa
P
tPa
PPag 22
12η
pPa
P→
∆zv
g
z
t→
∂
∂
zv
FC
Re
=
21 ippp +=
gapvz
Linearized Reynolds squeeze film equation
zv
FK
ωIm
=
is the pressure
Harmonic Problem
is the normal velocity componentzv
Damping Coefficient, Ns/mSqueeze Stiffness Coefficient, N/m
Real component of the pressureImaginary component of the pressure
Application: Parametric Squeeze Film Analysis
0.0
0.5
1.0
1.5
),,( 02 fyxP
×101
2nd order TS
1st order TS
air gap:3.8µm4.0µm4.2µm
-0.5 4.5
5.0
5.5
6.0
×10-6
),,( 01 fyxP
Operating frequency, MHz
0.0 0.5 1.0 1.5f0
gap = 4.2µm
gap = 3.8µm
gap = 4.0µm
6.5
Operating frequency, MHz
=
−
02
1Q
P
P
KC
CK
ω
ω
∫=A amb
TdA
P
gapNNc
η1210
01 3gap
=D
∫=A
TdABDBk
2nd order TS
0.0 0.5 1.0 1.5f0
∫= dAPF 2
Im
∫= dAPF 1
Re
22
pROM Example: Fixed-fixed beam
FE Model of the coupled domains
Mode1 Mode2 Mode3 Mode4
MODAL CONTRIBUTION FACTOR84.45 % 2.74%
3. right conductor
1. leftconductor
bulk
anchor
2. beam
insulator
Vs
squeezeelementsmaster
node
L
y
zx
airelements
structuralelements
Solid Model
200×20×2 µm
Air-gap 4 µm
1033 elements 1767 nodes
23
pROM Example: Parametric Modal Analysis and Mesh-Morphing
Mode 1 Mode 3
The relative error is less than 0.5 %.
160 180 200 220 2400.2
0.3
0.4
0.5
0.6
0.7
Evalution point
4th order TS
160 180 200 220 2401.0
1.5
2.0
2.5
3.0
3.5
4.0
Fre
quency,
MH
z
Evalution point
4th order TS
Length of the beam in µmLength of the beam in µm
Fre
quency,
MH
z
Perturbations of the nodes with respect to modal amplitude qj and design parameters L are given by:
∑ ⋅+=j
jzjiji wLqLzLqzi
)()(),( φz
x
0
1
Contour plot of design velocity v(x,y,z) for mesh perturbationsq1 q3
Structural mesh perturbation with q3 = 1.0µm
Air mesh perturbation with q3 = 1.0µmAir mesh perturbation with q1 = 2.0µm
Structural mesh perturbation with q1 = 2.0µm
)()( 00 LLaxLx iii −+=
24
pROM Example: SIMULINK Parametric Model of the FFB
cap12
dcap 12
cap23
dcap 23
cap13
dcap 13
current 13
6
current 23
5
current 12
4
displacement at master node
3
modal displacements
2
modal velocities
1
v to u
1
s
modal restoring forces
MATLAB
Function
modal damping forces
MATLAB
Function
modal accelerations
MATLAB
Function
electrostatic
force 23
electrostatic
force 13
electrostatic
force 12
charge Q 23
charge Q 13
charge Q 12
cap 23 +dcap 23
MATLAB
Function
cap 13 +dcap 13
MATLAB
Function
cap 12 +dcap 12
MATLAB
Function
a to v
1
s
Q23 to I23
du /dt
Q13 to I13
du /dt
Q12 to I12
du /dt
Eigenvector
K*u
voltage 13
5
voltage 23
4
voltage 12
3
parameter L
2
modal external forces
1
5
10
15
20
25
increasing L
Cap
acit
ance
C12, fF
-3 -2 -1 0 1 2 3Modal amplitude q1, µm
-5
-4
-3
-2
-1
-3 -2 -1 0 1 2 3Modal amplitude q1, µm
∂C12/∂
q1, fF
/µm
increasing L
-0.5
0.5
-3 -2 -1 0 1 2 3Modal amplitude q1, µm
∂W
/∂q
1, nJ/
µm
0.0
increasing L
160 180 200 220 2405
6
7
8
9
Length L, µm
Modal m
ass
m1, kg
×10-12
∑−
∂
∂=
∂
∂++
r
sk
i
ks
i
iiiiii
tVtV
q
LqC
q
LqWtqLmLLtqLm
2
)()(),(),()()()()(2)()( &&& ωξ
25
pROM Example: Results of Simulation
The transient responses for double step voltage excitation
0
50
100
150
200
Ele
ctr
icalexi
tation
, V
Time, µs
5 10 15 20 25 30 35 40 45 50
L = 240µm
-0.6
-0.4
-0.2
0.0
0.2
Dis
pla
cem
ent
, µ
m
Time, µs
5 10 15 20 25 30 35 40 45 50
L = 160µm
T1T2
Non-linear dynamics:
T1 > T2
),( Lqf ii =ξ
(FSEI)
26
Conclusions and Outlook
A short conclusions of the work are given in the following:
• A parametric FE-solver based on the derivation of discretized FE equations and the computation of a Taylor polynomial of the solution from the high order derivatives has been prototyped
• Automatic differentiations technique for extraction high order derivatives of the FE matrixes, its inverse and determinant has been utilized
• A library of parameterized finite elements has been developed
• A simple mesh-morphing algorithm has been successfully implemented
• A large number of numerical tests have been performed in order to evaluate the accuracy and efficiency.
Further work will be focused on:
• Adaptation of the HODM to really industry problems (over 105 DOF)
• Fully automation of the algorithm
27
The End
Спасибо за внимание!
Vielen Dank für ihre Aufmerksamkeit!
Thank You for Your Attention!
28
Error Estimation
1.FEM Solution at the point p0:• Discretization errors • Computation errors• Formulation errors
2.Compute higher order Derivatives at the point p0:• The smoothness of the mesh perturbation• Method of differentiation (AD)
3.Build an approximation:Convergence Domain of the Taylor expansion
δ
δ = F(mesh perturbation , order derivatives, range)
-Exact FEM Solution
-Finite element method error
-Parametric FEM Solution
- Derivatives computation and approximation errors
possible to control
p
G(p)
p0
Validity range: p0 ± ∆ p
δ
Parametric solution
Evaluation point
δ0
Reference point
δ1 δ2
δ0 ≠ δ1 ≠ δ2
1
0
)1(
)()!1(
)()( +
+
−+
= mm
m ppm
fpR
ξ , where p0 < ξ < p.
The Lagrange form of the remainder is given by
29
Parameterization of Linear FE Problem
1.Solution at the point p0:
u(p0) = [K(p
0)]-1.F(p
0)
[K] is stiffness matrixu is solution vector (unknown)F is load vectorp
0is initial parameter
Linear FEM Problem:
[K(p0)].u(p0) = F(p0)
2.Compute Higher Order Derivatives at the point p0:
u(p0)′ = [K(p
0)]-1′.F(p
0) + [K(p
0)]-1.F(p
0)′
u(p0)′′ = ([K(p
0)]-1′.F(p
0) + [K(p
0)]-1.F(p
0)′) ′
u(p0)′′′ = …
u(p0)n =
− −
=
− ∑ )(
0
)(
0
1
)(
0
1
0 )()()()( innn
i
i
n
npupKCpFpK
3.Build an approximation:• Taylor expansion: u(p - p0) = Σ u(p0)
(n). (p - p0)n / n!
• Pade approximation: u(p-p0) =Σ r m(p0)(∆p)m /Σ s k(p0)(∆p)k
30
Parameterization of NL FE Problem
1.Solution at the point p0:
u(p0)i = u(p0)i-1 − [J(p
0)]-1.R(p
0)
2.Compute Higher Order Derivatives at the point p0:
u(p0)′ = −[J(p
0)]-1.R(p
0)′ (implicit function)
u(p0)′′ = [J(p0)]
-1.(R(p0)′′ − [J(p
0) ]′.R(p
0)′)
u(p0)′′′ = …
u(p0)n =
3.Build an approximation:• Taylor expansion: u(p - p0) = Σ u(p0)
(n). (p - p0)n / n!
• Pade approximation: u(p-p0) =Σ r m(p0)(∆p)m /Σ s k(p0)(∆p)k
[K] is stiffness matrixu is solution vector (unknown)F is load vectorp
0is initial parameter
NL FEM Problem:
R(u(p0)) = 0j
iij
u
RJ
∂
∂=
p
pu
RJ
∂
∂=][
− −
−
=−
−− ∑ )(
0
)(
0
1
1
1
)1(
0
1
0 )()()()( inin
i
i
n
npupJCpRpJ
31
Differentiation Techniques
[ ] ∑∂
∂=
∂
∂ ngp
i
iii
T
i wJBDBp
kp
]][[][
3.1. AD relies on the fact that every function is a sequence of arithmetic operations: additions, multiplications and elementary functions.3.2. Evaluate numerically values at the point as (2) up to machine precision!3.3. There are two ways to implement:
source transformation;operator overloading.
Analytical Derivatives are usually more complicated than the function itself ! ! !
1. Analytical / Symbol
3. Automatic / algorithmic
Chain rules> 100 pages
308.017
37174)0(
2−=
⋅−⋅−=′f
4)0( −=′u 3)0( =′v
17)0( =v7)0( =u
)(
)()(
pv
pupf =
22
22
)5317(
)103)(1.047()5317)(2.04()(
pp
pppppppf
++
++−−+++−=′
21.047)( pppu +−=25317)( pppv ++=
p
ppfppf
dp
∆
∆∆
2
)()()( −−−−−−−−++++≈≈≈≈
2.1. Truncation and cancellation errors2.2. Partial and high order derivatives
n-point Finite Difference Formulas
2. Numerical
2)(
)()()()()(
pv
pvpupvpupf
′−′=′
Matrix rules:
MathCAD
4. Example: differentiation of the f(p):
4.1.
4.2.
4.3.f = uT.v
f = [u].v[ f ] = [u].v
[ f ] = [u].[v][ f ] = det[u][ f ] = inv[u]
at p = 0
32
Overview of AD software
ADOGEN
ADiMat
FFADLib
ADIC
ADMIT-1
ADOL-C
ADIFOR
Tools
C/C++
MATLAB
C/C++
C/C++
MATLAB
C/C++
Fortran77
Language
source transformation
source transformation
operator overloading
operator overloading
source transformation
operator overloading
operator overloading
source transformation
Technique
up to n
up to 2nd
up to n
up to 2nd
up to 2nd
up to n
1st
Order
patented, proprietary tools
for derivate matrices
2002 Jacobian and
Hessian
2000
1997
1996 / 98
1992
1991
Comment
33
Automatic Mesh-Morphing Algorithm
Ω
Γ
x
Morphing process
x
y
L0
L = L0+∆L
Design velocity Perturbated mesh
Laplacian smoother: 2 it. Final mesh
Original mesh
Vs = 1V
∑
∑
=
==n
i
i
n
i
c
ii
k
A
xA
x
1
1
Γ′
Ω′
pxvxpOpp
xxpx )()(
)0,()0,(),( 2 +≈+
∂
∂+=
TTT
v(x)px′
parameter ≡≡≡≡ L
02
2
=∂
∂
ix
v
1. Laplacian smoother 2. Electrical analogy
How obtain the perturbation of the interior nodes?
T(x,p):Translation
Rotation
Scaling
v(x) ≡≡≡≡ Design velocity
p ≡≡≡≡ Time
34
Polynomial FE Mesh
L=4*nKB=2*nKD=0.5*nK
nL=8 nB=4 nD=2
/prep7
ET,1,SOLID45
K,1,0,0,0 $K,2,L,0,0 $K,3,L,B,0 $K,4,0,B,0K,5,0,0,D $K,6,L,0,D $K,7,L,B,D $K,8,0,B,D
V,1,2,3,4,5,6,7,8
L,1,2,nLL,1,4,nBL,1,4,nB
VMESH,ALL
Parametric Design Language (APDL)
…………
0.00.00.12
0.00.00.01
zyx#
……………
8…211
n8…n2n1#
Elements Table
Nodal Table
EWRITE, Fname, ext ……………………
…
2
2
2
1
1
1
#
…
z
y
x
z
y
x
dof
…
…
…
…
…
…
…
…
…
3.7
1.0
0.5
1.3
0.2
0.0
an-1
0.6…1.00.1
5.4…0.10.0
2.0…0.50.0
6.1…1.30.0
…………
0.1…0.30.0
0.0…0.00.0
ana2a1a0
Polynomial Nodes Table
NWRITE, Fname, ext
n = tuple_size
Parametric FE model vs. Parametric polynomial FE model
L
D
Re-m
esh
n
n papapaapx ...)( 2
210 +++=n
n pbpbpbbpy ...)( 2
210 +++=n
n pcpcpccpz ...)( 2
210 +++=
Z
YX
0
?