A Physics-Informed Dual-Branch Neural Network for Scattering from 3D Dielectric Objects

The scattering center (SC) model provides a compact, physically interpretable, and efficient representation of electromagnetic (EM) scattering by approximating the total field as a superposition of dominant contributors characterized by location, amplitude, and phase. However, extending the SC model to three-dimensional (3D) dielectric objects remains challenging due to complex scattering mechanisms inherent to 3D dielectric objects. This paper presents a physics-informed dual-branch convolutional learning (PhiDCL) neural network for efficient, accurate, and interpretable EM scattering from arbitrarily shaped 3D dielectric objects. The proposed approach employs a dual-branch convolutional neural network (CNN) to learn the real and imaginary components of the complex scattering weights associated with equivalent scattering centers. By incorporating wavelength-scale geometric embedding and directional feature encoding, the network captures physically meaningful scattering behaviors while requiring only a limited set of full-wave simulations for training. Extensive validation on canonical and realistic objects, including cones, cubes, SLICY-like structures, and UAVs, demonstrates that PhiDCL achieves accurate radar cross-section (RCS) predictions with consistently low mean absolute errors (typically below 1 dB) across different geometries, material parameters, and incident angles. Compared with traditional numerical full-wave methods, it improves the computational efficiency by more than two orders of magnitude. The resulting parametric model serves as a scalable surrogate for rapid RCS estimation, target recognition, and inverse scattering analysis, enabling real-time EM signature prediction in resource-constrained applications.