Abstract: With the widespread use of low-observable dielectric objects such as drones, composite structures, and radomes in engineering, the demand for electromagnetic scattering modeling for these targets is rapidly increasing. However, due to the multiple reflection and transmission effects of electromagnetic waves within dielectric objects, traditional numerical methods typically require significant time and computing resources for high-precision calculations, making them difficult to meet real-time requirements. To address this issue, this paper proposes a two-branch neural network that integrates physical information to construct a parameterized electromagnetic scattering model for three-dimensional dielectric targets. By incorporating geometric prior information and phase modulation, combined with a loss function designed based on amplitude-phase consistency, this method achieves efficient learning of scattering characteristics while ensuring physical consistency. This model enables rapid prediction of target radar cross sections at the decibel scale. To validate the effectiveness of this method, numerical experiments were conducted on a commercial NVIDIA GPU platform for targets with various geometric shapes and dielectric parameters, including a frustum model and a fixed-wing drone. Results demonstrate that the proposed model maintains prediction accuracy consistent with full-wave simulations while achieving millisecond-scale inference times, demonstrating its potential for real-time application in complex electromagnetic environments.