Machine Learning-Assisted Design of the W-band Via-holes Transition Based on Multilayer LCP Substrate

To reduce the computational burden with full-wave electromagnetic (EM) simulations, this paper presents a W-band via interconnect design methodology guided by a collaborative approach that integrates Gaussian Process Regression (GPR) and Genetic Algorithm (GA). By using training data samples from HFSS simulations with different dimensions of via-hole transition, a GPR surrogate model is initially built. In order to obtain a viahole transition with low reflection coefficient, both the GA and the GPR surrogate model is then employed to obtain the solution(ie, the optimal dimension). During the optimization process, the solution obtained from each optimization iteration is validated using full-wave EM simulations. If it does not meet the requirements, it will be used as training data to retrain the surrogate model, with the aim of improving the model’s accuracy. The optimal solutions shows that the via-holes transition has a return loss S11 of less than - 20dB and an insertion loss S21 of better than 2 dB in the range of $75-110 \mathrm{GHz}$.