Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
A DSMC-based variance reduction formulation for low-signal flows
H. A. Al-Mohssen [email protected] N. G. Hadjiconstantinou [email protected]
Mechanical Engineering Dept., MITSeptember 2009
Motivation
✤ Boltzmann Equation(BE): describes the evolution of PDF f=f(x,c,t)
✤ Direct Simulation Monte Carlo simulates the BE, the uncertainty in "measurement" is:
✤ We want:
✤ Related Work:
๏ Öttinger, 1997: polymer simulation
๏ Chun and Koch, 2005: linearized BE
๏ Hadjiconstantionu, et el. 2004-2009: deviational particles
2
ΣUncertainty �ΣThermal
NSamples⇒ problems in low signal(≡ deviation from equilibrium) flows
(eg. low Ma flows).
ΣUncertainty �Σ�Signal�NSamples
s.t. σ(Signal)→ 0 as Signal → 0
� f� t� c �� f�x� �� f� t�Collision
�12�� � � �∆1� � ∆2� � ∆1 � ∆2�� f1� f2�c12�Σ�����c1��c2
✤ Let 〈R〉 be a property of interest (eg. ux=〈cx〉,〈cx〉 etc.). In general, it can be written as:
Where feq is an arbitrary reference (equilibrium) distribution
✤ An estimate of this quantity (that we will call R ̅) can be calculated by generating samples c i from f(c i)
Notation
3
�R� �� R�c� f �c���c �R�eq �� R �c� feq�c���c and likewise for feq≠f,
� R � 1N��i�1N R�ci�
4
Variance Reduction Approach
I � � � f �c� � g�c�� dc �� g�c� dcI � � f �c� dc � I �
1N �
i�1
Nf �ci�
if � g�c� dc is known deterministically & f �c� � g�c�
(1)
(2)
Variance Reduction Approach
I � � � f �c� � g�c�� dc �� g�c� dcI � � f �c� dc � I �
1N �
i�1
Nf �ci�
if � g�c� dc is known deterministically & f �c� � g�c�
(1)
(2)
0 1 2 3 4 5 60
1
2
3
4
5
� f �c� dc0 1 2 3 4 5 6
0
1
2
3
4
5
� � f �c� � g�c�� dc
� g�c� dc
Variance Reduction Approach
I � � � f �c� � g�c�� dc �� g�c� dcI � � f �c� dc � I �
1N �
i�1
Nf �ci�
if � g�c� dc is known deterministically & f �c� � g�c�
(1)
(2)
(2) can be estimated more efficiently than (1)
Variance Reduction Approach
I � � � f �c� � g�c�� dc �� g�c� dcI � � f �c� dc � I �
1N �
i�1
Nf �ci�
(2) can be estimated more efficiently than (1)
if � g�c� dc is known deterministically & f �c� � g�c�
(1)
(2)
Hadjiconstantinou, Baker, Homolle, Radtke
I � �� f �c� � g�c� dc� �� g�c� dc
Simulate using deviational particles
Variance Reduction Approach
I � � � f �c� � g�c�� dc �� g�c� dcI � � f �c� dc � I �
1N �
i�1
Nf �ci�
(2) can be estimated more efficiently than (1)
if � g�c� dc is known deterministically & f �c� � g�c�
(1)
(2)
Hadjiconstantinou, Baker, Homolle, Radtke
I � �� f �c� � g�c� dc� �� g�c� dc
Simulate using deviational particles
This Work
I � �� f �c� dc� � �� g�c� dc� �� g�c� dc
Unmodified DSMC Auxiliary Weighted DSMC
Illustration
�4 �2 0 2 4
0.1
0.2
0.3
0.4
f
c�4 �2 0 2 4
0.1
0.2
0.3
0.4
feq
c
10 20 30 40 50 60Time
2.07�1010
2.08�1010
2.09�1010
2.1�1010
2.11�1010
2.12�1010�Cx4�
Non-Eq Simulation (Regular DSMC)
9
Illustration
10 20 30 40 50 60Time
2.07�1010
2.08�1010
2.09�1010
2.1�1010
2.11�1010
2.12�1010�Cx4�
10
Non-Eq Simulation (Regular DSMC)
Eq Simulation (Weighted DSMC Auxiliary Simulation)Analytical Eq Value
Non-Eq Simulation (Regular DSMC)
Eq Simulation (Weighted DSMC Auxiliary Simulation)
Non-Eq Simulation (DSMC VR)
RVR � R � Req � �R�eq
11
Illustration
10 20 30 40 50 60Time
2.07�1010
2.08�1010
2.09�1010
2.1�1010
2.11�1010
2.12�1010�Cx4�
Analytical Eq Value
Formulation
✤ Our Formulation:
๏ Use an unmodified DSMC to directly calculate R ̅
๏ Use an auxiliary simulation to calculate R ̅eq. The auxiliary simulation does not perturb the main DSMC simulation and uses the same samples c i
๏ How can we calculate both R ̅ and R ̅eq from the same set of data?
• Use weights!
12
R �1N��i�1
NR�ci�
Auxiliary Simulation Using Weights
✤ Likelihood ratios (Wi≡W(c i)≡feq(c i)/f(c i)):
✤ As a result:
13
�R�eq �� R �c� feq�c���c �� R�c�� feq �c�f �c� � f �c���c �� R�c��W�c� f �c���c
� Req � 1N��i�1N Wi�R�ci�
RVR � R � Req � �R�eq � 1N �i�1
N �1 �Wi� R�ci� � �R�eq
Rejected
Advection Step
Collision Step
Accepted
Initalize N particles
Select i & jAccept with probability
cr� � cr �MX
Sample equlibrium andnon-equlibrium properties
ci � ci, c j � c j
ci � ci�, c j � c j�
Algorithm Overview
Regular DSMC
Rejected
Advection Step
Collision Step
Accepted
Initalize N particles
Select i & jAccept with probability
cr� � cr �MX
Sample equlibrium andnon-equlibrium properties
ci � ci, c j � c j
ci � ci�, c j � c j�
at ci & Wi � ?? ??
Wi� � ?? ??& Wj
� � ?? ??
Wi� � ?? ??& Wj
� � ?? ??
Algorithm Overview
Regular DSMC
+
VR DSMC
✤ For every particle at c i there will be on average particles at c i. If we have x particles at c i there will be (at c i)
๏ particles in an Equilibrium simulation
๏ particles in a Non-equilibrium simulations
✤ Since we advance particles per the Non-equilibrium rules each particle at needs to simulate:
✤ The final collision rules
๏ Accepted
๏ Rejected
Conditional Weight Update Rules
16
x Pci�ci�x Peq:ci�ci�
Pci�ci��
Peq:ci�ci�
Pci�ci�
Wi &Wj � Wi Wj
Wi � Wi1 �Wj c�r1 � c�r
&Wj � Wj1 �Wi c�r1 � c�r
�
✤ Problem:
๏ These weight update rules are not stable ⇒ loss of Variance Reduction
✤ Solution:
๏ From definition Wi=feq(c i)/f(c i) ⇒ we need knowledge of PDF
๏ Re-construct the PDF from samples, this is a standard numerical method known as Kernel Density Estimation. Specifically, for every particle at c replace
✤ Implementation:
๏ For accepted particles
Stability
17
Wi � W�i �
1�� Si �� �j�Si Wj (average weights within a sphere of radius ε in velocity space)
f �ci� � f��ci� � � K�c� � c� f �c�� d c� f
��ci� � f �ci� as K��c� � ∆��c�since
Rejected
Advection Step
Collision Step
Accepted
Initalize N particles
at ci & Wi �feq�ci�f �ci�
Select i & jAccept with probability
cr� � cr �MX
Sample equlibrium andnon-equlibrium properties
Wi� �Wi
�1�Wj cr� ��1�cr� � & Wj
� �Wj�1�Wi cr� ��1�cr� �
ci � ci, c j � c j
ci � ci�, c j � c j�
Wi � Wi�Wj�
, Wj � Wi�Wj�
Final Algorithm
(1)W�i �
1�� Si �� �j�Si Wj
estimate Wi�
& Wj�
using (1)
�
�
�
�
�
��
�
0.10 0.500.20 0.300.15�
0.20
0.10
0.05
0.02
0.01
ErrrelError
Results: Bias (error) vs. ε in 0D
19
ε/Cmp
NCell �3
Cmp3 >>1
0.0 0.5 1.0 1.5 2.0Time �Τ�
1.005
1.010
1.015
1.020
1.025
1.030�cx4�
�
�
Error �� ��Exact
Appx. Sol.
SS
Results: 1D Transient Couette Flow
20
0.2 0.4 0.6 0.8 1.0
x
L
0.990
0.995
1.000
1.005
1.010
1.015
T
T0
0.2 0.4 0.6 0.8 1.0
x
L
�0.020
�0.015
�0.010
�0.005
qy
qy,0
0.2 0.4 0.6 0.8 1.0
x
L
0.990
0.995
1.000
1.005
1.010
1.015
1.020
Ρ
Ρ0
0.2 0.4 0.6 0.8 1.0
x
L
�0.004
�0.002
0.002
uy
c0
0.2 0.4 0.6 0.8 1.0
x
L
0.990
0.995
1.000
1.005
1.010
1.015
T
T0
0.2 0.4 0.6 0.8 1.0
x
L
�0.020
�0.015
�0.010
�0.005
qy
qy,0
0.2 0.4 0.6 0.8 1.0
x
L
0.990
0.995
1.000
1.005
1.010
1.015
1.020
Ρ
Ρ0
0.2 0.4 0.6 0.8 1.0
x
L
�0.004
�0.002
0.002
uy
c0
0.2 0.4 0.6 0.8 1.0
x
L
0.990
0.995
1.000
1.005
1.010
1.015
T
T0
0.2 0.4 0.6 0.8 1.0
x
L
�0.020
�0.015
�0.010
�0.005
qy
qy,0
0.2 0.4 0.6 0.8 1.0
x
L
0.990
0.995
1.000
1.005
1.010
1.015
1.020
Ρ
Ρ0
0.2 0.4 0.6 0.8 1.0
x
L
�0.004
�0.002
0.002
uy
c0
NCell=500
Kn=1.0
~1% Relative Error
Relative Sampling Uncertainty
21
� � � � � � � � � � ��
�
�
�
�
�
�
�
�
�� ��
1�10�4 5�10�4 0.001 0.005 0.010 0.050 0.100
Uwall
c0
0.1
1
10
100
1000
Σu
DSMC
VR-DSMC
≈ 5 cm/s
1,000,000 less samples for the same uncertainty at 5cm/s!
Stability
22
Σ2 �Wi� �Σ2 �Wi�Kn�10,��Si ���0
�� Si ��
10 10 10 10 10 10 10 10 10 10 10 10 10
3
33 3 3 3 3 3 3 3 3 3 3
1
1
11 1 1 1 1 1 1
11
0.3
0.3
0.30.3
0.30.3
0.30.3
0.3
0.1
0.1
0.10.1
0.1
0.1
0.15 10 15 20
1.0
2.0
3.0
1.5
For stability at a fixed ε Kn� � �� Si �� �
� NCell
Conclusions
✤ Variance reduction using likelihood ratios is viable and exciting
๏ Main advantage: the DSMC simulation is never perturbed
✤ Small increase in computational cost
๏ Need to find NN of some particle ⇒ total cost scales as O(NCell Log(NCell))
✤ Stability Issues:
๏ KDE introduces bias that is a function of ε
๏ for low Kn NCell↑ for a Stable and Accurate solution
✤ Looking forward:
๏ Other collision Models
• BGK
• Maxwell
๏ Improve bias for a given NCell
23
Q&AMore info including sample code:http://web.mit.edu/husain/www
Appendix
DSMC with weights: Scattering probability
26
✤ DSMC is a set of probabilistic steps
✤ Start by selecting the same number of candidate particles:
Candidates= NEffNCel(NCell-1)MXσΔt/VCell
✤ if we choose particles of velocity classes c i and c j with weights Wi and Wj respectively there will be:
(NEff)2WiWj CijσΔt/VCell Collisions
✤ to correctly account for this we use the following collision probabilities:
Pi �Wj cijMX
Pj �Wi cijMX