Rapid Full-Wave Training-Data Synthesis for Deep-Learning Surrogates

A strategy for efficiently generating high-quality training datasets for forward electromagnetic (EM) surrogates that predict the performance of pixelated microwave passive devices is presented. A precomputed numerical Green’s function (PNGF) based full-wave EM solver enables ultrafast simulation of large numbers of training examples with only a minimal one-time precomputation overhead. The PNGF solver is combined with a direct binary search (DBS) algorithm to enhance the quality of the training dataset for a given computational budget. This combination of fast EM simulations and DBS-guided sample generation substantially lowers the overall computational costs of training surrogate models. To demonstrate the efficacy of the proposed methods, a proof-of-concept deep-learning model is constructed that accurately predicts the performance of pixelated metallic antennas with dielectric substrates. It is trained on 6 million full-wave EM simulations that fully include the dielectric substrate, produced with 3600 speedup relative to Ansys HFSS, thereby removing need for transfer learning.