Interpretable artificial intelligence for modulated metasurface antenna design using SHAP and MLP

Modulated metasurface-based leaky-wave antennas have garnered increasing attention due to their low profile, cost-efficiency, and versatility in applications such as telecommunications, imaging, and radar systems. These antennas offer fine control over critical electromagnetic properties, including leakage factor, linear momentum, and angular momentum states. In this work, we propose an interpretable artificial intelligence framework that integrates SHapley Additive exPlanations (SHAP) with a multi-layer perceptron (MLP) to predict two key radiation metrics: sidelobe level (SLL) and half-power beamwidth (HPBW). Through SHAP's game-theoretic analysis, we identify dominant features and key feature interactions governing antenna behavior. These insights inform targeted feature engineering-specifically, the introduction of interaction terms between the most influential variables. We then develop a multi-task neural network that jointly predicts SLL and HPBW, leading to substantial improvements in predictive accuracy. Our results demonstrate that interpretability can be leveraged not only to understand model decisions but also to refine model architecture and performance. The proposed methodology achieves near-perfect accuracy in SLL prediction and significant enhancement in HPBW estimation, offering a generalizable approach for integrating explainable AI into the modeling and optimization of advanced metasurface antenna systems.