Automatic Coupler Design Based on Artificial Neural Network With Self-Adaptive Local Surrogates

Abstract

A new artificial neural network (ANN) inverse modeling method with self-adaptive local surrogates, called ANN-SALS, is presented in this article, for a large-scale beam-forming network (BFN). To get accurate dimensions of the huge number of directional couplers, constituting a large BFN could be computationally expensive. Here, we develop an automated and more efficient algorithm to design multiple couplers with different specifications simultaneously using shared data for local surrogate modeling. While the result predicted by ANN serves as the starting point, a Gaussian process (GP) local surrogate model is built around it for fine-tuning. Then, ANN database is updated by selected GP training samples as well as the fine-tuned solution. After carefully setting the design specification, GP training data can be shared among the design of multiple couplers despite different specifications so that the design of one coupler is improved by the other ones. What is more, the size of region where GP is built is corrected by a quasi-sensitivity analysis method to improve the local optimization success rate. To verify the algorithm, four different couplers with different design requirements are designed, showing the proposed technique could unravel various multiple coupler design problems automatically and is much more efficient compared with existing methods.