Title: Singularitytreated quadratureevaluated method of moments solver for 3D capacitance extraction
1Singularity-treatedquadrature-evaluated method
ofmoments solver for3-D capacitance extraction
- Jinsong Zhao
- June 08, 2000
Cadence Design Systems, Inc.
2Outlines
- Introduction
- QMMS
- Implications
- Numerical examples
- Conclusions and future work
3Introduction
- Extraction
- critical piece in physical design and
verification - on-the-fly extraction
- database building
- Requirement
- accuracy
- speed
- capacity
4Research progress
- Classic partial differential equations
- Finite-element
- Finite-difference (absorbing boundary condition,
MEI) - Solve in the 3-D space, hit the curse of
dimension - Integral equation method
- Integral equation boundary condition
- Smaller (compared to domain method) but full
matrix - Fast integral equation solver
5Fast integral equation solvers
- Multipole-accelerated FastCap (K. Nabors, 1991)
- Pre-corrected FFT technique (J. Phillips, 1995)
- SVD-compressing IES3 (S. Kapur, 1996)
- Monopole-based (W. Shi, 1998)
- Wavelet-type multiscale (J. Tausch, 1999)
means acceleration method is Kernel-independent
Now we are able to solve 100000 unknowns!! ...But
WHY??
6Reason for the difficulty
- Why the numerical extraction is so difficult?
- Hand-calculation examples
- plate capacitor
- a metal ball
- A cube becomes a hard problem
- Convergence is slow
- Because corners and edges create SINGULARITIES!
7Singularity structure
- Luckily, we know the singularity structure!
q
Which reminds us to use Gaussian-Jacobi
quadrature to evaluate the integral
8Gaussian Quadrature
- Gaussian quadrature is the best way to evaluate
the integral of a smooth function.
W(x) can be considered as a Precondition
function that best handles singularities
9QMMS
- Quadrature evaluated method of moments solver
- Sample points, rather than coefficients, are
unknowns. This is known as pseudo-spectral
method, as powerful as high-order, yet much
simpler to implement. - Testing functions handle the singularity in the
Greens function quadrature handles the
singularity in the solution space.
10Comparison to Nystrom method
- Nystrom method (Kapur Long, 1998)
- For each r, special quadrature rules are created
- collocation on both r and r
- singularity solution space is not handled
- QMMS
- For each testing function, a general
Gaussian-Jacobi rule is used. Testing function
weakens the singularity in Greens function - collocation on r only
- Implications
- Gaussian nodes are the best ways to allocate
nodes and discretization.
11A square plate example
Smooth-part ofcharge distribution
True charge distribution
12Convergence comparison
- Gaussian node allocation high-order
0.1
uniform
Uniform(2)
Uniform(3)
0.01
0.001
Gaussian(1)
0.0001
Gaussian(2)
Gaussian(3)
350
400
50
100
150
200
250
300
To get 1 accuracy, we only need 16 points for a
plate and 24 points for a cube. It takes no time
to compute.
13Interdigitated capacitor
3000
1000
500
With 500 points, results match 21908 adaptive
panels!
Each panel is discretized based on Gaussian nodes
14Some numerical observations
- SVD-compression technique is used on top of QMMS
- So far, NO preconditioner is found to be better
than without preconditioner - high-order testing degrades convergence and a
better preconditioner is needed for really large
problems.
15Conclusion
- QMMS is presented
- Charge singularity structure at corners and edges
- Method of Moments testing handles Greens
function singularity, Guassian-Jacobi quadrature
handles solution singularity - High-order testing and collocation in solution
- Numerical implication
- Within 1 accuracy, regular LU is good and fast
- More accuracy is achieved through SVD-accelerated
method - Future work
- Inductance
- full-wave