Parallel Decomposition Approach to Gradient-Based EM Optimization

Abstract

Electromagnetic (EM) optimization is an essential part of an EM design cycle. In this paper, for the first time, a decomposition method is formulated to address the challenges of EM optimization with many design variables. Single large EM optimization is decomposed into multiple smaller suboptimizations to improve the optimization efficiency. In practical EM optimizations, there are no two design variables whose effect on the output response is completely independent. We propose a decomposition method, which uses the second-order derivative information to measure the level of dependence (measure of interaction) between the design variables and thereby identify, separate, and group the design variables to form multiple smaller suboptimizations. Each suboptimization includes generating multiple fine model responses for constructing the surrogate model and its optimization using trust region. Multiple suboptimizations are formulated to be independent, so that parallel computations can be exploited to evaluate concurrent suboptimization updates. The resultant vector, a combined vector of suboptimization updates, will be closer to the optimal solution. Furthermore, the step size of the optimization update in the proposed approach is much larger in comparison with the step size of the optimization update without decomposition. Using this approach, the overall large optimization update decreases the value of the objective function more efficiently and quickly. Two examples of EM optimization are used to illustrate the proposed technique.