Pixels to Physics: Transfer Learning for Deep Learning Surrogate Modeling of MIMO Antenna S-Parameters

Abstract

The conventional design of Multiple-Input Multiple-Output (MIMO) antennas relies heavily on full-wave electromagnetic simulation to evaluate candidate geometries, a process that is computationally expensive and limits the scale of design space exploration. This thesis investigates the use of transfer learning to adapt a pretrained Residual Network with Spatial Attention surrogate model to predict S-parameters of randomly generated MIMO antenna geometries operating in the 2-5 GHz frequency range. A dataset of 100,000 pixelated antenna designs was generated using Method of Moments simulation to facilitate domain adaptation from a source dataset of mixed electromagnetic structures. Two transfer learning strategies, feature extraction and fine-tuning, are systematically evaluated across multiple learning rate and batch size configurations, including a differential learning rate experiment applying distinct rates to the pretrained body and output head. Fine-tuning with a learning rate of 1x10-5 and batch size of 4096 was identified as the champion configuration, achieving a test set mean squared error of 0.03474 and a normalized root mean squared error of 9.4%. Analysis of the performance gap relative to the pretrained baseline suggests that dataset size disparity, domain shift, and architectural capacity constraints are likely contributing factors. The per-frequency error analysis reveals increasing prediction error toward higher frequencies, consistent with the transition from electrically small to distributed electromagnetic behavior. These findings motivate future directions including expanded and more balanced dataset generation, architectural refinements, and the incorporation of coordinate-aware and graph-based modeling approaches toward a surrogate model capable of supporting practical MIMO antenna design workflows.