An Efficient Generalized Regression Neural Network for Automatic Coupler Structure Optimization

When focusing on a large-scale Beam Forming Network (BFN), the design efficiency problem of multiple similar couplers is hard to deal with in traditional ways. In this paper, by utilizing Generalized Regression Neural Network (GRNN) and latest transfer learning method, a new promising framework for coupler structure optimization is proposed. Firstly, the six submodules with various functions establish the framework. Secondly, the input data are classified into qualified and unqualified groups, in which the connection knowledge between physical dimensions and electrical performance will be learned and transferred. Thirdly, cooperated with Gaussian Process, the suggestions are given for unqualified samples. Finally, in the optimization experiment of 75 four-port waveguide couplers, the framework achieves 91.67% improvement with less running time. The results show that this method has great potential for efficient design of multiple couplers.