Accelerated Machine Learning Antenna Synthesis Through Binary Metal-Vacuum Mesh Activation and Data Reuse in a MoM Framework

Paper i proceeding, 2025

We propose a novel approach to efficiently generate large datasets for antenna synthesis by addressing challenges in handling arbitrary geometrical configurations. In 30 minutes, a dataset of 3,000,000 antennas, covering a frequency range from 0.1 to 5 GHz, was generated, significantly accelerating the design process through machine learning. This was achieved by adding/removing triangular metal patches inside the antenna mesh domain through efficient manipulation of the rows and columns of a pre-calculated Method-of-Moment (MoM) matrix. Such flexibility in efficiently simulating new designs by reusing precomputed data is not available in commercial electromagnetic (EM) solvers. The generated dataset of antenna meshes is used to train a convolutional neural network (CNN). After training, the CNN accurately predicts the input impedance for arbitrary input meshes, without using an EM solver. The CNN was inte-grated into a genetic optimization algorithm, allowing antenna optimization in minutes instead of hours, with a mean squared error 0.0026.