Bayesian-Inspired Sampling for Efficient Machine-Learning-Assisted Microwave Component Design
Zhao Zhou, Zhaohui Wei, Jian Ren, Yu‐Xiang Sun, Yingzeng Yin, Gert Frølund Pedersen, Ming Shen
Abstract
Machine learning (ML) has demonstrated significant potential in accelerating the design of microwave components owing to its great ability to approximate the projection between geometric parameters and electromagnetic (EM) responses. A well-trained ML model can predict the EM responses of a microwave component with unseen geometric parameter settings accurately, or determine the parameter settings based on desired EM constraints in a matter of milliseconds. However, this ML-based design process often requires heavy simulation to collect a large amount of training data. To mitigate this issue, this article proposes an efficient Bayesian-inspired sampling-assisted ML method for the design of microwave components. In contrast to typical ML-based design methods which use uniform and arbitrary sampling to extensively represent the entire parameter space, necessitating intensive simulation for generating training data, the proposed Bayesian-inspired sampling strategy efficiently represents the entire parameter space by recognizing and emphasizing more promising parameter settings. This is achieved by defining a Bayesian-based expression for evaluating the probability of the outcome of adding a new data sample in a specific parameter area. During each iteration of the design process, new data is always added in the area with the highest probability of beneficial outcomes. Therefore, it optimizes the distribution of training data and reduces the amount of required training data and simulations. Results from three design case studies demonstrate that the proposed method can significantly reduce the number of required data and simulation by around 40% for the same model performance. This validates that the proposed Bayesian-inspired sampling-aided ML method significantly improves overall efficiency.