A Physics-Driven Stacked Ensemble Network for Efficient Target RCS Prediction

In this paper, we propose a stacked ensemble network that integrates machine learning (ML) as the base model and deep learning (DL) as the meta-model. A novel loss function is designed to incorporate both data-driven and physics-driven losses by introducing a scattering center model. This integration leverages physical information mechanisms to guide network training, providing interpretability and enabling accurate, quasi-real-time radar cross section (RCS) predictions for targets. Additionally, transfer learning is employed to reduce the required sample size by nearly 50%, enabling the pretraining and fine-tuning of the best-performing model among five stacking methods for more complex conductive and dielectric targets. Numerical results demonstrate that the proposed approach reduces RMSE by 49%-77% and improves R² by 23%-52% compared to using the five ML methods individually before optimization. The training time for the optimal method is only 271.940 seconds, and it can accurately predict the azimuthal RCS of an electrically large and realistic target within 0.078 seconds.