Mesh-Based Deep Learning Surrogate Model for Rapid Wideband 3-D Electromagnetic Scattering Computation

Three-dimensional (3D) wideband electromagnetic scattering underpins applications from radar cross section (RCS) analysis to stealth design, yet existing numerical methods often demand prohibitive computational resources. Drawing on recent advances in deep learning, we present GTC-3DEM, a mesh-based network that integrates graph neural networks, Transformers, and convolutional decoders to rapidly predict scattering fields from triangular surface meshes. By injecting global incident-wave parameters through a hierarchical conditioning control mechanism, GTC-3DEM captures both local and global scattering phenomena with high fidelity. Compared to conventional simulations, our approach reduces per-sample RCS computation time from tens of seconds to under 0.2 seconds, while preserving accuracy across diverse object geometries and frequencies. Further extended studies on frequency scalability, mesh sensitivity, polarization diversity, and total/complex electric-field prediction demonstrate the robustness and adaptability of the proposed framework. These results establish GTC-3DEM as a powerful surrogate model for fast, accurate, and scalable 3D electromagnetic scattering computations. Our code is publicly available at https://github.com/cytosineXT/GTC-3DEM.