Knowledge-based coarse and fine mesh space mapping approach to EM optimization

Abstract

Space mapping is an effective method for speeding up EM optimization. The method normally requires an equivalent circuit as the coarse model. This paper addresses the situation when an equivalent circuit coarse model is not available. We establish our coarse model using a lookup table to store the data of coarse mesh EM simulations and its derivatives, avoiding the EM re-simulations w.r.t. the same values of design variables. In the proposed method, the surrogate model is developed using knowledge-based neural network (KBNN) combining the coarse model with a neural network. Our technique uses mostly coarse mesh EM evaluation and occasionally fine mesh EM evaluation to achieve optimal EM solutions with fine mesh accuracy. This technique is illustrated by two microwave filter examples.