Coupler Structure Optimization Using A Learning- transferred Modular Neural Network

In microwave components, efficient structure optimization of directional couplers is a challenging task. Traditional methods, like space discretization and surrogate-based optimization, are time-consuming and lack of generalization ability. In this paper, by combing two latest AI techniques, namely transfer learning and modular neural network, a new approach is proposed. First, the framework with five submodules is designed, each submodule has a specific function. Second, neural networks in submodules are trained using abundant simulated coupler data, and the deep relations between performance and structure for ideal couplers are learned. Third, the learned knowledge from submodules 2/4 is transferred to submodule 3/5, to provide optimization suggestions for couplers with unsatisfied performances. At last, an experiment using 75 four-port waveguide directional couplers is designed, the result show that the proposed method can effectively learn from the data and output the optimized structure with 91.67% improvement. It also indicates that the proposed method can outperform some other comparable methods.