Beam Coefficient Prediction for Antenna Arrays Using Physics-Aware Convolutional Neural Networks

This work introduces a novel approach for predicting the complex element weights of a rectangular planar phased array for a desired beam pattern using physics-aware convolutional neural networks (PA-CNN). An image representation of the desired radiation pattern is provided as input to the proposed network architecture, where PA-CNN learns to recognize spatial features and map input patterns to the amplitude and phase of the antenna element weights. In addition to the classical mean squared error (MSE) loss for regression to true weights, a novel physics-aware loss function is designed to jointly force the PA-CNN to estimate true weight while generating physically consistent beam patterns. The performance of the proposed model is evaluated using metrics such as MSE of estimated weights and traditional radiation pattern metrics such as accuracy of peak direction. The obtained results demonstrate that the models successfully learn the nonlinear relation between antenna parameters and desired beam patterns, where the physics-aware approach outperforms the base CNN cases with an improvement of 6.2% in the MSE between radiation patterns. This approach provides a promising solution for learning complex weights of arrays for synthesizing desired patterns while ensuring physical consistency.