Explainable Gradient Boosting Approach for S11 Analysis in Rectangular Microstrip Patch Antennas

Abstract

Machine learning (ML) has emerged as a promising approach for accelerating the design and analysis of microstrip patch antennas by replacing computationally expensive full-wave electromagnetic simulations. However, many existing studies emphasize prediction accuracy while providing limited model interpretability and insufficient validation on unseen antenna geometries. This work presents an explainable Light Gradient Boosting Machine (LightGBM) framework for predicting the reflection coefficient (S11) of a rectangular microstrip patch antenna using geometric dimensions, feed-line dimensions (feed_x and feed_y), and operating frequency. A dataset comprising 60,760 antenna samples generated using ANSYS HFSS was employed for model development, and the LightGBM hyperparameters were optimized using GridSearchCV with five-fold cross-validation. The optimized model achieved an R2 score of 0.8184, a mean absolute error of 0.3244 dB, and a root mean square error of 1.1326 dB on previously unseen test data. SHapley Additive exPlanations (SHAP) were employed to provide global and local interpretations of the trained model. The analysis revealed that operating frequency and feed-line length are the most influential parameters governing impedance matching, while the contributions of the remaining design variables were consistent with established antenna theory. The framework was further validated using independently simulated HFSS antenna geometries with parameter combinations excluded from the training dataset, demonstrating good agreement between predicted and simulated S11 values. A model-guided antenna optimization study at 2.4 GHz improved the predicted S11 from -6.37 dB to -12.03 dB, with independent HFSS validation achieving -13.54 dB. The proposed explainable LightGBM framework therefore provides an accurate, interpretable, and computationally efficient surrogate modeling approach for rapid antenna analysis, design exploration, and optimization.