Upload
phungngoc
View
223
Download
0
Embed Size (px)
Citation preview
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Computational Methods for MEMS
N. R. AluruBeckman Institute for Advanced Science and Technology
Thanks to: Xiaozhong JinGang LiRui QiaoVaishali Shrivastava
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Course Objectives
Understand how MEMS are designedUnderstand some of the computational techniques that go into the development of MEMS simulation toolsSpecific examples: electrostatic MEMS, microfluidics
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Outline
Some MEMS ExamplesMixed-Domain Simulation of electrostatic MEMS and microfluidics
Techniques for interior problems (e.g. FEM)Techniques for exterior problems (e.g. BEM)Algorithms
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Pressure Sensor
ApplicationsBiomedical (e.g. blood pressure)
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Accelerometer
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Micro Mirror
ApplicationsHigh performance projection displays
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Outline
Some MEMS ExamplesMixed-Domain Simulation of electrostatic MEMS and microfluidics
Techniques for interior problems (e.g. FEM)Techniques for exterior problems (e.g. BEM)Algorithms
Dynamic Analysis
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Is Electrostatics a good idea?
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Scaling Laws
Useful to understand where macro-theories start requiring corrections with the aim of better understanding the physical consequences of downscalingDevelop an understanding of how systems are likely to behave when they are downsizedExamples
By reducing the size of a device, the structural stiffness generally increases relative to inertially imposed loadsThe mass or weight scales as l3, while the surface tension scales as l as the system size becomes smallerMore difficult to empty liquids from a capillary compared to spilling coffee from a cup because of increased surface tension in a capillaryHeat loss is proportional to l2; Heat generation is proportional to l3; As animals get smaller, a greater percentage of their intake is required to balance heat loss; Insects are cold blooded
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Scaling in Electrostatics
L2Electrostatic forceLSurface TensionL3Gravity
L3Mass
LVelocity
LDistance
L3TorqueL3PowerL2Muscle force
L0Time
L1/4van der Waals
L2Friction
h
w
d
Consider a capacitor
The electrostatic P.E. stored in a capacitor is:
dhwVE br
me 2
20
,εε=
=bV electrical breakdown voltage
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Scaling in Electrostatics
31
2110
, ll
llllE me →=
Assume scales linearly with d (the gap)bV
The maximum energy stored in the capacitor scales as L3
If L decreases by a factor of 10, the stored energy in the capacitor decreases by a factor of 1000
Electrostatic Force
∂∂−= 2
21 CV
xFx
∂∂−= 2
21 CV
yFy
∂∂−= 2
21 CV
zFz
23
lllFx →=
Electrostatic force scales as L2; This is an advantage because mass and inertial forces scale as L3; The electrostatic force gains over inertial forces as the size of the system decreases
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Scaling Laws: Vertical Bracket NotationDifferent possible forces can be written as
=4
3
2
1
llll
Fcase where the force scales as L1
case where the force scales as L2
[ ][ ] 33 −− === FF lllmFa Flll
Fmxt −⋅⋅== 32
3400
11ll
llVt
xFVP
F
F
−=
=
→
=4
3
2
1
llll
F
=−
−
1
0
1
2
llll
a
=
12
1
2/3
ll
lt
=−
−
0.2
5.0
1
5.2
0llll
VP
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Scaling Laws: Remarks
Even in the worst case when F = L4, the time required to perform a task remains constant when the system is scaled downUnder more favorable force scaling (e.g. F = L2), the time required decreases as L (a system 10 times smaller can perform an operation 10 times faster i.e. small things tend to be quick)
For Electrostatics 2ll F =0.1
0
11 −− === lVPltla
When the force scales as F = L2, the power per unit volume scales as L-1 => When the scale decreases by a factor of 10, the power that can be generated per unit volume increases by a factor of 10
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Outline
Some MEMS ExamplesMixed-Domain Simulation of electrostatic MEMS and microfluidics
Techniques for interior problems (e.g. FEM)Techniques for exterior problems (e.g. BEM)Algorithms
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Elastostatics
Finite-difference methodsFinite-element methodsMeshless methods
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Key steps in FEM:
•Construct a weak or a variational form of the problem
•Obtain an approximate solution of the variationalequations through the use of finite element functions
Finite Element Method: Introduction
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
• Trial function q)1(u,Hu|u 1 =∈=δ
•Test function (or weighting functions) 0)1(w,Hw|w 1 =∈=ν
• Strong form
0dx)fu(w1
0xx,∫ =+
• Integrate by parts 0dxfwdxuwwu1
0
1
0x,x,
10x, ∫ ∫ =+−
1x0 0fu xx, ≤≤=+
B.C. Dirichlet q)1(u =
B.C. Neumann h)0(u x, =−
(S)
• Derive weak form
Weak formh)0(wdxfwdxuw
1
0
1
0x,x,∫ ∫ += (w) ( ) ( )WS ⇔
1-D Example: Strong and Weak Forms
0 1
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
•Notations: ∫=1
0x,x, dxuw)u,w(a ∫=
1
0
dxfw)f,w(
h)0(w)f,w()u,w(a +=•The weak form can be rewritten as
∑=
=N
1AAA
h CNw ∑=
=N
1BBB
h dNu
• Galerkin Approximation Method
hhu δ∈
hhw ν∈
h)0(w)f,w()u,w(a hhhh +=Galerkin form
( ) hC0Nf,CNdN,CNaN
1AAA
N
1AAA
N
1BBB
N
1AAA ∑∑∑∑
====
+
=
(G)
Apply the interpolation
1-D Example: Galerkin Form
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( ) 0h0NfdNddNNCN
1BAABx,Bx,A
N
1AA =
−Ω−Ω∑∫ ∫∑=
Ω Ω=
CA’s are arbitrary, so
( )∑ ∫ ∫=
Ω Ω+Ω=Ω
N
1BAABx,Bx,A h0NfdNddNN for A=1,2, …, N
∑=
=N
1BABAB FdK for A=1,2, …, N
where
∫Ω Ω= dNNK x,Bx,AAB ( )h0NdfNF AAA +Ω= ∫Ω
The matrix form FdK = (M)
1-D Example: Matrix Form
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
[ ]
==
NNNN
N
N
AB
KKK
KKKKKK
KK
21
22221
11211 ......
M
==
N
A
F
FF
FFM2
1
==
N
A
d
dd
ddM2
1
Remarks:
• K is symmetric;
• For any given problem, ( ) ( ) ( ) ( )MGWS ⇔≈⇔
1-D Example: Matrix Form
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( ) 211
21 xxx,
hxxxN ≤≤
−= ( ) n1n
1N
1NN xxx,
hxxxN ≤≤
−= −
−
−
1
x1 x2 xA-1 xA xA+1 xn-1 xn
N1 N1Nn
Shape functions
( )
≤≤−
≤≤−
= ++
−−
−
otherwize,0
xxx,h
xx
xxx,h
xx
xN 1AAA
1A
A1A1A
1A
A
Shape Functions
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
[ ]1AA x,x +Domain [ ]21 ,ξξ
global local
Nodes 1AA x,x + 21 ,ξξ
1AA d,d + 21 d,dDegree of freedom
1AA N,N + 21 N,NShape functions
Interpolation function
( ) ( )( ) 1A1A
AAh
dxNdxNxu
++
+= ( ) ( )( ) 22
11h
dNdNu
ξ+ξ=ξ
For a linear finite element
Local/Element Point of View
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
uh
NA NA+11 1
dA dA+1
eA xx 1= e
A xx 21 =+
N1 N21 1
d1e d2
e
12 +=ξ11 −=ξ
eξ
ex
Mapping
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( )21
2N
aa
,a−
=ξ
=ξ
2xx
2hx
e1
e2
ee
,−
==ξ ( ) e
1e,
ex, h
2x ==ξ−
ξ
( ) ( ) ( ) ( ) e22
e11
2
1a
eaa
e xNxNxNx ξ+ξ=ξ=ξ ∑=
( )ξ−= 121N1 ( )ξ+= 1
21N 2
( ) ( )ξξ+=ξ aa 121N
( )A
1AA
hxxx2x −−−
=ξ
• Local Shape function
where
• Derivative of the Shape function
Mapping
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
∑=
=nel
1e
eKK
∑=
=nel
1e
eFF
[ ]ABee KK =
[ ]Aee FF =
( ) ∫Ω==e
dxNNN,NaK x,Bx,Ae
BAABe
( ) ( )h0Nf,NF Ae
AAe +=
[ ]2e
1ee x,x=Ωwhere is the domain of the e-th element
( ) ( )
( )( ) ( )( ) ( ) ξ=ξξξ=
=
−ξ
+
− ξξξ
+
−
Ω
∫∫∫
dxNNdxxNxN
dxxNxNK
1,
1
1 ,b,a,
1
1 x,bx,a
x,bx,aabe
e
−
−=
1111
h1K e
e
where
• Global stiffness
• Force vector
local stiffness
Matrix Assembly
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( )( )
−
−=
1111
h1K 1
1 ( )( )
−
−=
1111
h1K 2
2
(1) (2)
10 11 12
−−+−
−=
MMM
MMM
LLLL
LLLL
LLLL
MMM
MMM
YYYYXX
XXK
10 11 12
1110
12
local stiffness
Global stiffness
example
Assembly Process
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Divergence theorem Γ=Ω ∫∫ ΓΩdnfdf
ii,
∫∫∫ ΩΓΩΩ−Γ=Ω dfgdngfgdf i,ii,Integration by parts
Heat Conduction
Strong form: given ℜ→Γℜ→Γℜ→Ω hg :h,:g,:f
find ℜ→Ω:u such that
hii
g
i,i
onhnq
ongu infq
Γ=−
Γ=
Ω=
where j,iji uKq −= Kij are the conductivities
FEM: Multidimensional Problems
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( ) 0dfqw i,i =Ω−∫Ω
0dfwdqwdnwq ii,ii =Ω−Ω−Γ ∫∫∫ ΩΩΓ
∫∫∫ ΓΩΩΓ+Ω=Ω−
h
whddfwdqw ii,
g on 0w Γ=
Notation: ( ) ∫Ω Ω= duKwu,wa j,iji,
( ) ∫Ω Ω= wfdf,w ( ) ∫ΓΓ Γ=h
dhwh,w
( ) ( ) ( )h
h,wf,wu,wa Γ+=
( ) ( ) ( )h
h,wf,wu,wa hhhhΓ+=
• Weak form
• Galerkin form
Apply the property of the weighting function
Weak Form: Heat Conduction Problem
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
∑=
=N
1AAA
h CNw ∑=
=N
1BBB
h dNu
Γ====
+
=
∑∑∑∑N
1AAA
N
1AAA
N
1BBB
N
1AAA h,CNf,CNdN,CNa
( ) ( ) ( )Γ=
+=∑ h,Nf,NdN,Na AAB
N
1BBA
FdK =• Matrix form
( )BAAB N,NaK =
( ) ( )Γ+= h,Nf,NF AAA
Substitute the interpolation
where
Heat Conduction Problem: Matrix Form
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Strong form: given fi, gi and hi, find ui such that
hijij
giii
ij,ij
onhn
ongu
in0f
Γ=σ
Γ=
Ω=+σ
The stress is related to the strain by Hooke’s law
klijklij c ε=σ
( )j,iu2
uu i,jj,iij =
+=ε
( ) 0dfwdnwdwdfw iijijiijj,iij,iji =Ω+Γσ+Ωσ−=Ω+σ ∫∫∫∫ ΩΓΩΩ
( ) ∑ ∫∫∫=
ΓΩΩ
Γ+Ω=Ωσ
nsd
1iiiiiij dhwdfwdj,iw
Strain tensor
• Weak form
FEM for Elastostatics
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( ) ijijj,i j,iww σ=σ
( ) [ ]jiwjiww ji ,,, +=
( )
[ ] symmetric) (skew 2
,
)(symmetric 2
,
,,
,,
ijji
ijji
wwjiw
wwjiw
−=
+=
So
( ) [ ]( )( ) [ ] ijij
ijijji
jiwjiw
jiwjiww
σσ
σσ
,,
,,,
+=
+=
[ ] [ ] [ ] [ ] ijjiijij jiwijwijwjiw σσσσ ,,,, −=−=−=
[ ] 0, =ijjiw σ
( ) ijijj,i j,iww σ=σWhy
where
rewrite
since
and
FEM for Elastostatics
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Notation: ( ) ( ) ( )∫Ω Ω= dl,kucj,iwu,wa ijkl
( ) ∫Ω Ω= dfwf,w ii ( ) ∑ ∫=
ΓΓ
Γ=
nsd
1iii
hi
dhwh,w
( ) ( ) ( ) vw all for h,wf,wu,wa ∈+= Γ
( ) ( ) ( )Γ+= h,wf,wu,wa hhhh
• Weak form
• Galerkin form
s)dof' 3( 1,2,3i dNun
1AiAA
hi == ∑
=
FdK =• Matrix form
FEM for Elastostatics
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
4
1 2
3
ξ
η
12
3
4
( )1,1 −−
( )ee yx 22 ,
( )ee yx 33 ,( )ee yx 44 ,
( )1,1 −
( )1,1( )1,1−
ηξ
yx
eΩ
( ) ( )
( ) ( ) ea
aa
ea
aa
yNy
xNx
∑
∑
=
=
=
=
4
1
4
1
,,
,,
ηξηξ
ηξηξ
Bilinear Quadrilateral Element
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( ) ( )( )
( ) ( )( )
( ) ( )( )
( ) ( )( )ηξηξ
ηξηξ
ηξηξ
ηξηξ
+−=
++=
−+=
−−=
1141,
1141,
1141,
1141,
4
3
2
1
N
N
N
N
( ) ( )( ) ( )( ) ( )( ) ( )( )
1
114111
4111
4111
41,
1
=
+−++++−++−−=∑=
ηξηξηξηξηξnen
aaN
• Shape functions
Property of the shape function
Shape Functions
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( ) ( ) ( ) l
n
lll
n
ll wgRwgdg ∑∑∫ ≅+=
+
−
intint1
1ξξξξ
• Trapezoidal rule
1
1
2
2
1
int
+=
−=
=
ξ
ξ
n2,11 == lwl
( )ξξξ,32 gR −=
• Simpson’s rule
1
0
1
3
3
2
1
int
=
=
−=
=
ξ
ξ
ξ
n
34
31
2
31
=
==
w
ww
( )( )ξ4
901 gR −=
Numerical integration
Numerical Integration
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( )3
20
,
11
ξ
ξ
ξξgR
w
=
==
( )( )135
13
13
1
4
21
21
ξ
ξξ
gR
ww
=
==
=−=
( )( )15750
98
95
530
53
6
231
321
ξ
ξξξ
gR
www
=
===
==−=
1int =n
2int =n
3int =n
Gaussian Quadrature Rules
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( ) ( )( )
( )( )
( )
( )
( )( )
( )( )
( )( )
( )
( )
( )( )2l
n
1l
1l
2
l
1
l
1
1
n
1l
1l
1
l
1
1
1
1
2
1int
1121
1int
111
ww,g
dw,gdd,g
∑ ∑
∫ ∑∫ ∫
=
+
−=
+
−
+
−
ηξ≅
η
ηξ≅ηξηξ
ξ
η
ξ
η
ξ
η
( ) ( )0,0g4dd,g1
1
1
1=ηξηξ∫ ∫
+
−
+
−
( )
−+
+
−+
−−=ηξηξ∫ ∫
+
−
+
− 31,
31g
31,
31g
31,
31g
31,
31gdd,g
1
1
1
1
examples
Gaussian Quadrature Rules in 2-D
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( )Xx ϕ= XFx dd =
( )
∂∂
∂∂
∂∂
∂∂
∂∂
∂∂
∂∂
∂∂
∂∂
==
3
3
2
3
1
3
3
2
2
2
1
2
3
1
2
1
1
1
XXX
XXX
XXX
D
ϕϕϕ
ϕϕϕ
ϕϕϕ
XF ϕ
( ) [ ]Fdet=ϕJ
Transformation function
Deformation gradient Volume transformation JdVdv =
1E
Original configuration
Deformed configuration
xX
2E
uXx +=3E
01 ρρJ
=
T
JFP1
=σ
Area transformation
t∂∂
=ϕV
( ) ( )( ) NFSns 1−= TdJd
Density transformation
Cauchy stress
Velocity
Finite Deformation Elastodynamics
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
[ ] BP 02
2
0 ρρ +=∂∂ Div
tϕ
• Boundary conditions:
EcS = FFc T= [ ]IcE −=21
SFP =where
Strong form:
[ ][ ]TΓhnP
TΓ
ihiAiA
g
,0on
,0
at
aton
g
=
=ϕ
• Initial conditions:
( ) Ω=
Ω=
=
=
V
in
in0
0
00
Vt
t ϕϕ
Elastodynamics: Governing Equations
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( ) [ ] ( )( )[ ] 0:, 0 =Γ⋅−⋅−= ∫∫∫ ΓΩΩ h
ddVdVDGradG hBES ηηϕϕηηϕ ρ
( ) 0int =− extFF d
• Weak form:
• Galerkin form:
Where
( ) dVT∫Ω= Sd)
BF int ∫∫ ΓΩΓ⋅+⋅=
h
ddVext NhNB0ρF
Nonlinear equations:
( ) [ ] [ ] 0:, 0 =Γ⋅−⋅−= ∫∫∫ ΓΩΩ h
ddVdVDGradG hhhhhhh hBS ηηϕηηϕ ρ
FEM for Elastodynamics
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
The scalar problem: r(d) is a scalar nonlinear function of d. Find such thatd
( ) 0=dr
Possibilities:
d
r
d
r
d
r
d
one unique solution several solutions no solutions
1d 2d
Solution of Nonlinear Systems: Newton Methods
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( ) ( ) ( ) ( )
( ) ( ) ( )
( ) ( )( )( )ddddKdrd
ddKdr
ddrdddrdr
ddrddddr
dddrdr
∆+=
−=∆
=∆+=
∆++≅
+∆++∆++=
=
==
01
0
0
00
000
002
2
000
0
!21
ε
εε
εε
εε
εε
KK
• Solution strategy:Guess a “good” d0 close to the solution;
d0d
1d
2d
r
• Geometric interpretation
Newton’s Method
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( )
( )( )
for endif end
else
if
econvergenc untilfor
'
,2,1
0
1
0
ddddrdrd
dd
toldri
ddi
ii
i
i
i
i
i
∆+=
−=∆
=
<
==
=
+
KK
The error at the i-th iteration is given by ( ) dde ii −=
( ) ( ) kii eCe ≤+1 if then we say the algorithm converges with order k.
Newton’s method has quadratic convergence, i.e. k=2
Algorithm: Newton’s Method
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Advantages of Newton’s method
•Optimal (quadratic) convergence close to solution
Disadvantages of Newton’s method
•Poor or no convergence far away from the solution;
•Computation of K(di) is very expensive in the general multidimensional case (K (di) is a matrix).
Algorithm: modified Newton’s method
( )
( )( )
endif end
else
if
econvergenc untilfor
'
,2,1
0
1
0
0
ddddrdrd
dd
toldridd
i
ii
i
i
i
i
∆+=
−=∆
=
<
==
=
+
KK
Modified Newton’s Method
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
( )dR
Similar to the scalar case
( ) ( ) ( ) ( )
( ) ( )( ) ( ) 0
!21
00
000
002
2
000
=∆+=
∆++≅
+∆++∆++=
=
==
ddKdR
ddRdddR
ddRddddR
dddRdR
T
ε
εε
εε
εε
εε
KK
Let be a n-dimensional vector valued nonlinear function of the n-dimensional vector d
( )( )
( )nnn
n
n
dddRR
dddRR
dddRR
,,,
,,,
,,,
21
2122
2111
KK)MM
KK)
KK)
=
=
= ( )
∂∂
∂∂
∂∂
∂∂
∂∂
∂∂
=
∇=+ =
n
n
n
n
nn
n
u
uu
dR
dR
dR
dR
dR
dR
uRudRdd
M
L
MOM
MOM
L
2
1
1
1
2
1
1
1
0εεε
• Directional or Frechet derivative
( )0dK T is called the tangent stiffness matrix
Newton Method: Multidimensional Case
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Remarks:
• The load step is needed since the method might not converge if the entire load is applied at once. Instead the load is applied incrementally. Each increment is converged before the next step is applied.
• Likewise one might have to apply the “g” b.c.’s incrementally. This method is called “displacement control”.
• In practice, the terminology “Newton-Raphson method”is often used to denote algorithms in which a new left hand size matrix is formed for each iteration. If KT is not updated in each iteration, but kept frozen for a couple of iterations, the term “modified Newton” method is used.
( )
( )( ) ( ) ( )
( )( )( ) ( ) ( )
( ) ( )
( )
1
1
:
K
0
0
1
1
1
int11
until
loop) (iteration do
e.g) stepping load ( loop step
+←=
<
+=∆+=
−=
−=∆
=
=
=
+
+
+
++
nndd
RR
iiddd
dFF
Rdd
dd
i
n
in
ii
iii
in
extn
iiiT
ni
ε
Newton-Raphson Method: Algorithm
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Outline
Some MEMS ExamplesMixed-Domain Simulation of electrostatic MEMS and microfluidics
Techniques for interior problems (e.g. FEM)Techniques for exterior problems (e.g. BEM)Algorithms
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
BEM - Introduction
What is Boundary Element Method ?Boundary discretization only Integral based method
Approaches available for solving boundary integral equations
BEM based on CollocationBEM based on Galerkin
Analysis of a turbine blade using FEM and BEM
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Comparison of FEM and BEM
BEM FEM
Fewer packages availableLot of commercial packages available
Unsymmetric, dense and smaller matrices
Symmetric, sparse and large matrices
Boundary mesh (1D/2D)Domain mesh (2D/3D)
Global approach, Integral based
Local approach
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Table 1
ConstantNeumann b.c.
Dirichlet b.c.
Scalar function
Problems
ResistivityElectric current
Electro-potential
Electric conduction
PermittivityElectric flowField potential Electrostatic
Stream functionIdeal fluid flow
Warping function
Elastic torsion
Thermal conductivity
Heat flowTemperatureHeat Transfer
)0( 2 =φ∇ )( φ )qn
K( =∂φ∂)( φ=φ )K(
.)DegT( ≡ )TT( =)q
nT( =∂∂
λ− )( λ
)(ψ )q)t,rcos(r( =
)sm( 12 −≡φ )( φ=φ )qn
( =∂φ∂
)voltV( ≡)VV( =
)qnV( =∂∂
ε− )( ε
)voltE( ≡)EE( =
)qnE
k1( =∂∂
)k(
rt
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
Laplace equation represents many problems in engineering (Table 1)
v+-
Interior problem Exterior problem
0T2 =∇ 02 =φ∇inside outsideTemperature ‘T’ known on the surface Potential known on the surfaceφ
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
BEM is based on the Second Theorem of Green
Problem definition
Ω
uΓ
n
x
)x(p2 =φ∇ Ω∈x
Governing equation:
Dirichlet boundary condition (b.c):
Neumann boundary condition (b.c):
)y()y( φ=φ uy Γ∈
)y(qn
)x(
yx
=∂φ∂
=
ny Γ∈
nΓ
≠=
Poisson;0Laplace;0
)x(p
y
Definition of the problem
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
Derivation of BIE:Multiplying with and integrating over
( ) 0dp2 =Ωφ−φ∇∫Ω∗
Integrating by parts we get (2D case),
( ) 0dpdn =+Ωφ+φ∇⋅φ∇−Γ⋅φ∇φ ∫∫ Ω
∗∗
Γ
∗
∗φ Ω)p( 2 −φ∇
Integrating by parts the second integral we get,
( ) 0dpdndn 2 =Ωφ−φ∇φ+Γ⋅φ∇φ−Γ⋅φ∇φ ∫∫∫ Ω
∗∗
Γ
∗
Γ
∗
( )1. . .
where n is the outward normal to the boundary Γ
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
Therefore, Equation (1) can be written as,
is the Fundamental Solution of Laplace equation. ∗φ
( ) ( ) 0dn
dn
dpdp 22 =Γ∂φ∂
φ−Γ∂φ∂
φ+Ωφ−φ∇φ=Ωφ−φ∇ ∫∫∫∫ Γ
∗
Γ
∗
Ω
∗∗
Ω
∗
( ) ( )[ ] Γ
∂φ∂
φ−∂φ∂
φ=Ωφφ∇−φφ∇⇒ ∫∫ Γ
∗∗
Ω
∗∗ dnn
d22 ( )2. . .
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
Fundamental Solution for Laplace equation :satisfies Laplace equationrepresents field generated by a concentrated unit charge acting at a
point ‘i’effect of this charge is propagated from ‘i’ to infinity
∗φ
0)j,i(2 =δ+φ∇ ∗ =δ )j,i( Dirac Delta function
Multiplying with and integrating we get,φ
idjid φδφφφ −=Ω−=Ω∇ ∫∫ ΩΩ
∗ )),(()( 2
Therefore,
Γφ
∂φ∂
=Γ
∂φ∂
φ+φ ∫∫ Γ
∗
Γ
∗
dn
dn
i
( )3. . .
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
Fundamental Solution
Equation
Convection/decay Equation
Diffusion Equation
Wave Equation
Helmholtz
Laplace
Fundamental Solutions : One Dimensional Equations
002 =δ+φ∇ ∗
0022 =δ+φλ+φ∇ ∗∗
( ) 0tt
c 02
222 =δδ+
∂φ∂
−φ∇∗
∗
( ) 0ttk
102
2 =δδ+∂φ∂
−φ∇∗
∗
( ) 0txt 0 =δδ+βφ+∂φ∂
φ+∂φ∂ ∗
∗∗
xr,2r
==φ ∗
( )rsin21
λλ
−=φ∗
( )rctHc2
1−=φ∗
H=Heaviside function
( )
−π
−=φ∗
kt4rexp
kt4tH 2
φ
−δ−=φ φβ−
∗ rter
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
Fundamental Solution
EquationFundamental Solutions : Two Dimensional Equations
Navier’s Equation
Plate Equation
Wave Equation
D’Arcy(orthotropic case)
Helmholtz
Laplace 002 =δ+φ∇ ∗
0022 =δ+φλ+φ∇ ∗∗
0dxdk
dxdk 02
2
2
221
2
1 =δ+φ
+φ ∗∗
( ) 0tt
c 02
222 =δδ+
∂φ∂
−φ∇∗
∗
( ) 0tt 0
422
2
=δδ+φ
∇µ−
∂∂ ∗
0x l
j
jk =δ+∂σ∂ ∗
22
21 xxr,
r1ln
21
+=
π=φ ∗
( )( )rHi4
1 20 λ=φ ∗
H0 = Hankel function
+
π−=φ∗
21
2
22
1
21
21 kx
kxln
21
kk1
( )
ππµ+=φ ∗
t4rS
4tH
i
Si = Integral sine function
( )( )222 rtcc2
rctH−π
−−=φ ∗
llkk eU ∗∗ =φ
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
Fundamental Solution
EquationFundamental Solutions : Three Dimensional Equations
Navier’s Equation(Isotropic homogenous)
Wave Equation
D’Arcy
Helmholtz
Laplace00
2 =δ+φ∇ ∗
0022 =δ+φλ+φ∇ ∗∗
0dxdk
dxdk
dxdk 02
3
2
322
2
221
2
1 =δ+φ
+φ
+φ ∗∗∗
( ) 0tt
c 02
222 =δδ+
∂φ∂
−φ∇∗
∗
0x l
j
jk =δ+∂
σ∂ ∗
23
22
21 xxxr,
r41
++=π
=φ∗
rier4
1 λ−∗
π=φ
21
3
23
2
22
1
21
321 kx
kx
kx
41
kkk1
−
∗
++
π−=φ
r4crt
π
−δ
=φ∗
llkk eU ∗∗ =φ
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
What happens when point ‘i’ is on ?Γ
ε=r
Boundary surface
Boundary point ‘i’
Surface
Γ
εΓ
3D case - Hemisphere around point ‘i’ 2D case - Semicircle around point ‘i’
Boundary point ‘i’
ε=r
Boundary curve Γ
Boundary curve εΓ
Augment the boundary with Hemisphere of radius in 3DSemicircle of radius in 2D
εε
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
Consider equation (3) before any boundary conditions have been applied,
- RHS integral easy to deal (lower order singularity),
Γ
∂φ∂
φ=Γ
∂φ∂
φ+φ ∫∫ Γ
∗
Γ
∗
dn
dn
i
≡
πεπε
∂φ∂
=Γ
πε∂φ∂
=Γφ
∂φ∂
→εΓ→εΓ
∗
→ε ∫∫εε
04
2n
limd4
1n
limdn
lim2
000
- LHS integral behaves as,
φ−=
πεπε
φ−=Γ
πεφ−=
Γ∂φ∂
φ→εΓ→εΓ
∗
→ε ∫∫εε
i2
2
0200 21
42limd
41limd
nlim
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation
Therefore,
Γ
∂φ∂
φ=Γ
∂φ∂
φ+φ ∫∫ Γ
∗
Γ
∗
dn
dn
c i ( )4
for smooth boundaries
. . .
,21c =
πθ
=2
c for corner pointsθ P
Boundary with corner point
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Integral Formulation (contd.)
Exterior Problem - Electrostatics,
A system of ‘Nc’’ ideal conductors
n
n−
∞R
jΓ
∞Γ=Γref
02 =φ∇ outside
Potential known on the surface of each conductorφ
For 3D Electrostatic problem the boundary integral equation is,
∑ ∫=
Γ
∗ Γφ∂φ∂
=φc
j
N
1j
i dn
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Element MethodEquation (4) is discretized to find system of equations
Boundary is divided into N elements
Discretized form of equation (3) at point ‘i’ is given as,
Γφ∂φ∂
=Γ∂φ∂
φ+φ ∑ ∫∑ ∫=
Γ
∗
=Γ
∗
dn
dn
cN
1j
N
1j
i
jj
Constant elements Linear elements Quadratic elementsNodes Element
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Element Method
In matrix form,
[ ] [ ]
∂Φ∂
=Φn
GH
where Hij and Gij are the influence coefficients given as,
‘i’ is the source point (where fundamental solution is acting) ‘j’ is the field point (any other nodes on the boundary)
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Element MethodConstant Elements:
and are assumed to be constant over each element The value of and is assumed equal to that at mid-element node
The influence coefficients, Hij and Gij are given as,
∫Γ∗
Γ∂φ∂
+δ=j
dn
)j,i(21H ij
Γφ= ∫Γ∗dG
j
ij
‘i’ is the source point (where fundamental solution is acting) ‘j’ is the field point (any other nodes on the boundary)
φ
∗φφ∗φ
Node ‘i’
Element ‘j’
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Element Method
Evaluation of integrals:Hij and Gij can be calculated numerically, for the case
For the case , Hij and Gij are evaluated analytically
21
21
21 =Γ
∂∂
∂∂
+=Γ∂∂
+= ∫∫ Γ
∗
Γ
∗
dnr
rd
nH
j
ii φφ
+
π=Γ
π=Γφ= ∫∫ ΓΓ
∗ 12/l
1ln2l1d
r1ln
21dG
ii
ii
ji ≠ji =
Node ‘i’r
n
lElement ‘i’
0
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Element MethodLinear Elements:
and are assumed to vary linearly over each element
Nodal value of Nodal value of
[ ]
φφ
=φ2
1
21 NN
[ ]∫Γ∗
Γ∂φ∂
+δ=j
dn
NN)j,i(21H 21
ij
[ ] Γφ= ∫Γ∗dNNG
j21
ij
φ ∗φ
∗φorφφ ∗φor
l
[ ]
∂φ∂∂φ∂
=∂φ∂
nn
NNn 2
1
21
Therefore,
Node ‘i’Element ‘j’
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Boundary Element Method
Putting all the unknowns on LHS we get,
[ ] FxA =
Note: A is a dense matrix
Dense MatrixA
(N x N) Vec
tor
x(N
x1)
= Ax
(N x
1)
DIRECTO(N3)
ITERATIVEO(N2)
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Fast Integral Equation Solver
Matrix-Vector multiplication: O(N(logN)O(N(logN)22))Storage: O(N(logN)O(N(logN)22))
Results : 2-Conductor Problem
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Fast Integral Equation SolverResults : Mirror Problem
Matrix-Vector multiplication: O(N(logN)O(N(logN)22))Storage: O(N(logN)O(N(logN)22))
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Fast Integral Equation SolverResults : Comb-Drive Problem
Matrix-Vector multiplication: O(NlogNO(NlogN))Storage: O(NlogNO(NlogN))
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
References
P.K. Banerjee, The Boundary Element Method in Engineering, Mc-Graw Hill, 1994
G. Beer, Programming the Boundary Element Method, Wiley, 2001
C.A. Brebbia and J. Dominguez Boundary Elements An Introductory Course, Mc-Graw Hill, 1996
J.H. Kane, Boundary Element Analysis in Engineering Continuum Mechanics, Prentice Hall, 1994
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Outline
Some MEMS ExamplesMixed-Domain Simulation of electrostatic MEMS and microfluidics
Techniques for interior problems (e.g. FEM)Techniques for exterior problems (e.g. BEM)Algorithms
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Coupled Electromechanical Analysis
We need to self-consistently solve the coupled electrical and mechanical equations to compute the equilibrium displacements and forces. Three approaches –
Relaxation techniqueFull-Newton methodMulti-level Newton method
Solution of elastostatic equations is represented by
))(( qPRu M=
Solution of electrostatic equations is represented by
),( VuRq E=
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Relaxation Technique
Simplest black-box approachData is passed back and forth between black-box electrostatic and elastostatic analysis programs until a converged solution is obtained
))(()1( kM
k qPRu =+
)( kE
k uRq =Repeat
Compute
Compute
0;1 == kuk
1+= kk
εε ≤−≤− ++ 11 kkkk qquuUntil
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Relaxation Technique
AdvantagesVery quick implementation based on black-boxesExisting mechanical and electrical solvers can be used
DisadvantagesFails to converge for strong coupling between electrical and mechanical domains
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Multi-Level Newton Algorithm
Matrix-free approaches: Matrix-vector product involving a Jacobian and a vector can be computed as
εε )()( uRuuRu
uR −∆+=∆∂∂
Define a new residual
−−= )(
)(),( qRuuRqquR
M
E
The Jacobian of the residual is given by
∂∂−
∂∂−
=
∂∂
∂∂
∂∂
∂∂
=I
qR
uRI
uR
qR
uR
qR
quJM
E
22
11
),(
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Multi-Level Newton Algorithm
),(),( kkkk quRuqquJ −=
δδ
Repeatsolve
0;0;1 === kk quk
1+= kk
εε ≤−≤− ++ 11 kkkk qquuuntil
set uuu kk δ+=+1
qqq kk δ+=+1set
use an iterative solver
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Iterative Solution of Linear Systems
Lets say we need to solve Key steps in GMRES algorithm
pPq =
make an initial guess to the solution, 0qset k = 0do
compute the residual, kk Pqpr −=
if ,tolrk ≤ return as the solutionkqelse
choose and ins'α βk
k
j
jj
k rqq βα += ∑=
+
0
1
to minimize 1+krset 1+= kk
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Multi-Level Newton Algorithm
θθ )()( uRruRr
uR −∗+
=∗∂∂
[ ]
[ ]
−+−
−+−=
∂∂−
∂∂−
=
)()(1
)()(1
),(qRqqRu
uRuuRq
uq
Iq
Ru
RI
uqquJ
MM
EE
M
E
θδθ
δ
θδθ
δ
δδ
δδ
ru
arusign ∗∗= )(θ
)5.0,01.0(∈a
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Multi-Level Newton Algorithm
AdvantagesBlack box based approachSuperior global convergence
DisadvantagesCan be sensitive to the choice of the matrix-free parameter
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Full-Newton Technique
Represent the mechanical and electrical equations as
0)()(),( int =−= qfufquR extM
0)(),( =−= VquPquRE
Let and be self-consistent solutionsu q0),( =quRM
0),( =quRE
0..),(),( 00 =+∆∂∂+∆
∂∂+= tohq
qRu
uRquRquR MM
MM
Let and be some initial guess0u 0q
0..),(),( 00 =+∆∂∂+∆
∂∂+= tohq
qRu
uRquRquR EE
EE
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Full-Newton Technique
Neglecting h.o.t
−=∆
∆
∂∂
∂∂
∂∂
∂∂
),(),(
00
00quRquR
qu
qR
uR
qR
uR
E
M
EE
MM
),( 00 quRqq
Ruu
RM
MM −=∆∂∂
+∆∂∂
),( 00 quRqq
Ruu
RE
EE −=∆∂∂+∆
∂∂
In matrix form
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Full Newton Algorithm
Repeat
solve
0;0;0 )()( === ii qui
1+= ii
εε ≤∆≤∆ )()( ii quuntil
set )()1()( iii uuu ∆+= −
)()1()( iii qqq ∆+= −set
−=
∆∆
∂∂
∂∂
∂∂
∂∂
−−
−−
),(),(
)1()1(
)1()1(
)(
)(
iiE
iiM
i
i
EE
MM
quRquR
qu
qR
uR
qR
uR
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Full Newton Algorithm
→∂
∂→∂∂
uuf
uRM )(int
entirely elastostatic part
→=∂−∂→
∂∂ P
qVPq
qRE )( entirely electrostatic part
→∂
∂→∂∂
qqf
qR ext
M )( electrical to mechanical coupling term
→∂
∂=∂−∂→
∂∂ q
uuP
uVPq
uRE )()( mechanical to electrical coupling
term
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Microfluidics: Gas Flows
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Microfilter properties:Openings of various shapesThickness between 1 and 5µmOpening size as small as 2nmHigh burst pressure achieved
Introduction to Microfilters
Design issues:Flow profilesEstimation of flow rateDependence of flow rate on:
geometrysurface propertiespressure difference
Rarefaction effects observed due to smalldimensions
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Characteristics of Flows in Micro-Channels
Typical Characteristics:
• Compressible
• High Kn #
• Small Re #
• Small Ma #
• Wide range of Kn #
• Reacting
Effects of high Knudsen Number:
• Slip velocity
• Thermal jump
• Strong interaction with walls
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
DSMC Flow Chart
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Micro-Filter Elements
1x1 1x5 0.2x1 1x10 0.05x1 0.2x2hc (µm) 1 1 0.2 1 0.05 0.2lc (µm) 1 5 1 10 1 2hp (µm) 5 5 1 5 1 1lin (µm) 4 6 4 4 4 4lout (µm) 7 7 5 7 7 5Kn 0.054 0.054 0.27 0.054 1.1 0.27
hp
lc
hc
lin lout
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
1µmX1µm Filter Element
X Velocity Y Velocity
TemperaturePressure
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Knudsen Number and Length Effects
Effect of Kn:Slip velocity increases with Kn
Effect of Length:As lc/hc increases, 2D channel approximationholds good for smaller Kn
@ y=hc/2
@ x=lc/2
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Effect of Surface Accommodation1x1 µm filter
Smaller accomodation coefficients:Strong increase in slip velocityTemperature drop increases
@ x=lc/2Velocity
@ y=hc/2
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Flow Rate vs. Pressure Difference
Dependence of flow rate onpressure is linear Qualitative behavior iscaptured by 2D channel formula + 1st order slip BC(Arkilic & Breuer, 1997)Good agreement for large lc/hcEffective length can be used for smaller lc/hc
University of Illinois at Urbana-ChampaignBeckman InstituteComputational MEMS/NEMS
Conclusions
MEMS design is still an artCritical issues
Mixed-domain simulation toolsMultiscale approachesSystem level modeling tools
Need fast and radically simpler techniques for MEMS modeling