Neural space mapping optimization for EM design

Space mapping is a recognized method for speeding up electromagnetic (EM) optimization. This paper describes a neural space mapping optimization method. In this method, neural networks are used as mapping function instead of linear mapping function. The surrogate model developed in each iteration is trained to match the fine model at single point. Through training process, Neural networks can give the model more flexibility and better accuracy than the linear mapping function. This method can use several iterations to make the fine model satisfy the specifications. The number of fine model evaluation of this method is much less than that of direct EM optimization. This technique is illustrated by a microwave filter examples.