Advanced Bayesian-Inspired Multilayer Effective Parameter Determination Method for Automated ANN Model Generation of Microwave Components

Artificial neural networks (ANNs) have revolutionized microwave computer-aided design by leveraging machine-learning techniques to tackle complex problems. One critical aspect of this process is the automatic modeling of the neural network structure. The conventional approach, which involves qualitative adjustments to prevent underfitting and overfitting, often leads to inefficiencies when the initial structure significantly deviates from the optimal one. This article presents a groundbreaking approach to address this issue, proposing an automatic modeling algorithm for neural networks. This algorithm employs Bayesian theory to optimize the ANN model structure. Based on the Bayesian theory, the formula for calculating effective parameters in a multihidden layer neural network is derived, allowing the initial structure to adopt any form and permitting efficient, quantitative adjustments. This innovative approach enables the computation of effective parameters under any given structure. A higher maximum number of effective parameters in multihidden layer ANN has been obtained compared to single-hidden layer ANN, thus improving modeling accuracy. Compared with the existing Bayesian-based automated ANN model generation methods, the proposed approach significantly enhances both modeling accuracy and speed. The effectiveness of this method is verified through the application of three microwave components.

I. Introduction

Artificial neural networks (ANNs) have brought revolutionary changes to microwave computer-aided design [1], [2], contributing to various tasks such as S-parameter parameterization [3], microwave imaging [4], and transistor modeling [5], [6], [7], [8], [9], [10]. ANNs have also been employed in microwave design optimization, employing strategies such as semi-supervised neural networks [11], hybrid sampling [12], and regularized deep learning [13]. ANN-assisted surrogate modeling has been used for microwave filter optimization [14], further enhanced by dimensionality reduction [15] and the use of self-adaptive local surrogates [16]. Machine learning using ANN in microwave components modeling has been researched [17], [18], [19], [20], [21]. Expedited variable-resolution surrogate models have been applied for miniaturized microwave passives [22], alongside multifidelity electromagnetic simulations and selective Broyden updates for cost-effective design [23]. Concurrently, homotopy optimization was used in conjunction with ANN modeling for millimeter-wave substrate integrated waveguide cruciform couplers [24]. Deep neural networks (DNNs) have facilitated rapid simulation of planar microwave circuits based on their layouts [25], while generative adversarial networks have been innovatively utilized for extrapolating load-pull data [26]. These efforts underscore the many researchers working to bridge the microwave and neural network fields, applying advanced neural network and machine learning methods to address microwave problems.