PhaseNet: A Deep Learning Framework for Reflectarray Antenna Gain Prediction by Integrating 2D Phase Maps and Angular Embeddings

Reflectarray antenna design traditionally depends on computationally intensive full-wave simulations and experimental measurements, which significantly increase design time and cost. To address these limitations, we propose PhaseNet, an end-to-end deep learning framework that leverages phase maps and radiation angles as inputs to predict reflectarray antenna gain values. PhaseNet integrates spatial features extracted by a convolutional neural network (CNN) backbone with an angle-embedding module, and employs a regression head to enable efficient forward prediction. After training, the model achieves near-real-time inference with a single forward pass, facilitating rapid exploration of high-dimensional design spaces and providing immediate design feedback. In addition, we introduce two novel evaluation metrics, Half-Power Beamwidth (HPBW) and Main-Lobe Root Mean Square Error (RMSE), which allow for a multidimensional evaluation of prediction performance across the full radiation pattern and within the critical main-lobe region. These metrics provide refined criteria for reflectarray antenna design optimization beyond conventional error measures. In particular, PhaseNet achieved the best performance compared to existing models on these newly proposed evaluation metrics, recording up to 0.45 lower HPBW RMSE than existing methods, thereby validating both the relevance of the metrics and the effectiveness of the model. To further enhance practicality, we present a rapid generation of diverse bit-encoded datasets, substantially reducing the time and cost associated with data acquisition. Overall, the proposed framework effectively reduces prediction errors in reflectarray antenna design.