em optimization using coarse and fine mesh space mapping
DESCRIPTION
EM Optimization using Coarse and Fine Mesh Space Mapping. Chao Zhang. Introduction. Space mapping is a recognized engineering optimization methodology in the microwave area. The coarse models : empirical functions , equivalent circuits computationally efficient low accuracy - PowerPoint PPT PresentationTRANSCRIPT
EM Optimization using Coarse and Fine Mesh Space Mapping
Chao Zhang
IntroductionSpace mapping is a recognized engineering optimization methodology in the microwave area. The coarse models : empirical functions , equivalent circuits computationally efficient low accuracyThe fine models: provided by an electromagnetic (EM) simulator accurate computationally intensiveSpace mapping combines the computational efficiency of coarse models with the accuracy of fine models.
Algorithm
The fine model: fine mesh EM simulation high accuracy achieved by the mesh convergence process
The coarse model: coarse mesh EM local meshing without the mesh convergence process.
Models
Algorithm
the neural networkweight parameters in the neural networkthe optimal parameters of the neural network
The fine model
The coarse model
The surrogate model
The response
Input variables
The optimal solution
( )f fR x
fx
*fx
( )c cR x
cx
( )s sR x
sx
*cx *
sx
NNPw
*w
Notions
AlgorithmFunctions
The original EM optimization:
(1)Build an input space mapping neural network:
(2)Establish a surrogate model:
(3)Unit mapping: train (2) to learn .Finding the optimal solution:
(4)Error function:
(5)Termination condition: (6)
* = arg min ( ( ))f
f f fUx
x R x
= ( , )c NN s x P x w
( ) ( ) ( ( , ))s s c c c NN s R x R x R P x w=c s x x
* = arg min ( ( ))s
s s sUx
x R x
= ( ( )) d f fE U R x
dE
AlgorithmFunctions
Optimize our surrogate model to match the fine model data. The error function for this optimization is defined as
(7)The solution of the optimization:
(8)Next we perform the neural network training,
(9)Perform the surrogate design optimization. Then perform fine model evaluation and repeat (3)-(9). When termination condition is satisfied, the algorithm stops.
2( ) ( ) ( )c f f c ce x R x R x
* = arg min ( )c
c cex
x x
2* * *= arg min ( , )NN s c w
w P x w x
AlgorithmFlowchart
Start
Stopyes
dE no
DESIGN OPTIMIZATION WITH MAPPED COARSE MODEL:We perform the surrogate design optimization in EM simulator to
find , by solving .
COARSE MODEL OPTIMIZATON: Initialize the surrogate model with unit
mapping and find , by solving .
*sx
FINE MODEL EVALUATION:Evaluate the at ,
and calculate the design error .
*=f s x xUPDATE MAPPING
PARAMETERS BY COARSE MODEL OPTIMIZATION:
Optimize the design variables , by solving .
Then find the optimal , by neural network training
.
*cx
*w
*sx
( )f fR xdE
* = arg min ( ( ))s
s s sUx
x R x * = arg min ( ( ))s
s s sUx
x R x
* = arg min ( )c
c cex
x x
2* * *= arg min ( , )NN s c w
w P x w x
ExamplesA. Bandpass HTS Microstrip Filter
The bandpass microstrip filter example: the coarse mesh and fine mesh EM simulations of this structure are used as the coarse and fine models, respectively.
Space mapping optimization on bandpass filter. (a) The comparison of the coarse and fine mesh
models at initial point. (b) Responses of the fine model from each iteration of
the space mapping optimization.
3.8 3.9 4.0 4.1 4.20
0.2
0.4
0.6
0.8
1
Frequency(GHz)
|S21
|
Coarse MeshFine Mesh
3.8 3.9 4.0 4.1 4.20
0.2
0.4
0.6
0.8
1
Frequency(GHz)
|S21
| InitialIteration1
(a) (b)
ExamplesA. Bandpass HTS Microstrip Filter
COMPARISON OF DIRECT FINE MODEL OPTIMIZATION AND PROPOSED SPACE MAPPING OPTIMIZATION
Optimization algorithm
Number of fine model evaluations
Optimization time Final design error
Direct fine model optimization 71 23h34min 0.0783
Proposed space mapping 2 1h46min 0
ExamplesB. Two-Section Lowpass Elliptic Microstrip Filter
(a) (b)The lowpass elliptic microstrip filter example: the coarse mesh and fine mesh EM simulations of this structure are used as the coarse and fine models, respectively.
Space mapping optimization on lowpass filter. (a) The comparison of the coarse and fine mesh
models at initial point. (b) Responses of the fine model from each iteration
of the space mapping optimization.
1 2 3 40
0.2
0.4
0.6
0.8
1
Frequency(GHz)
|S21
|
Coarse MeshFine Mesh
1 2 3 40
0.2
0.4
0.6
0.8
1
Frequency(GHz)
|S21
|
Initial Iteration1
ExamplesB. Two-Section Lowpass Elliptic Microstrip Filter
COMPARISON OF DIRECT FINE MODEL OPTIMIZATION AND PROPOSED SPACE MAPPING OPTIMIZATION
Optimization algorithm
Number of fine model evaluations
Optimization time Final design error
Direct fine model optimization 69 25h13min 0.0001
Proposed space mapping 2 3h44min 0
Conclusion
1. The method uses coarse and fine mesh EM evaluations.2. The coarse model: coarse mesh EM simulation.3. Useful when equivalent coarse model is not available.4. Much more efficient than direct fine mesh EM optimization.