Fast Multi-Physics Simulation of Microwave Filters via Deep Hybrid Neural Network

Abstract

One fundamental difficulty in multiphysics numerical simulation is the complex interactions between different physics domains leading to plenty of computational costs. Although neural networks have recently been introduced in multiphysics simulations, the modeling complexity and the enormous amount of training data required may still pose significant challenges to researchers. In this work, we introduce a low-cost, electromagnetic-centric, multiphysics modeling approach to simulate microwave filters. With ground-truth datasets being generated from the finite element method, a novel deep hybrid neural network (DHNN) model structure is introduced, which uses the sigmoid and the ReLU functions as activators to mimic the diversity of biological neurons. A new, more feasible training algorithm is proposed for the efficient development of the DHNN model. The algorithm adopts the design-of-experiment (DOE) sampling technique and is specifically designed for the simulation of multiphysics responses. The strong approximation ability of the DHNN can lead to high-accuracy modeling with fewer training data and less resource consumption. Another advantage of this approach is that the modeling process is more concise and easier to apply compared with other modeling technologies. Numerical examples show that the DHNN can achieve higher accurate results with much less training data compared to traditional ANNs. The advantages of the proposed method in computational efficiency are more pronounced, especially when the amount of input data increases.