Fast Optimization of Arbitrary-shaped Antennas Using a Deep Neural Network Model Trained Once by an Efficient Electromagnetic Field Solver

Artikel i vetenskaplig tidskrift, 2025

Fast optimization of arbitrary-shaped antennas is enabled by a neural network model, trained by a Method of Moments (MoM) framework capable of evaluating large sets of pixel-based antenna metal layouts. The MoM matrix equation is constructed once for a fully metalized pattern. Matrix rows and columns are selectively removed to reflect the absence of metal pixels. Fixed regions, such as the ground plane, dielectric, and meshed port are accounted for through the Schur complement. Using this framework, a dataset of 2,000,000 antenna configurations is generated in 19 hours—a speedup of 13.5 times compared to a plain MoM approach. Meshing is done only once, as opposed to commercial solvers, including meshing the speed advantage is 270 times. A convolutional neural network is trained on this dataset and combined with a genetic algorithm to synthesize various triple-band Wi-Fi 7 antennas, which are experimentally validated. These results demonstrate the realworld applicability of the proposed MoM framework for MLbased optimization of arbitrary-shaped antennas.