A Deep Learning-Based Method for Predicting Surface Currents on Targets With Large Electrical Size Variations

Efficient electromagnetic simulation has long been a challenging task. Traditional computational electromagnetics methods, such as full-wave numerical methods (FWNMs) and high-frequency asymptotic methods (HFAMs), face significant limitations in handling large-scale problems. While FWNMs offer high accuracy, they are computationally expensive; conversely, HFAMs are faster but less accurate for complex scatterers. This paper presents a deep learning-based approach to solve electromagnetic scattering problems involving large electrical sizes. The proposed method integrates two key techniques: the induced current prediction method (ICPM), which employs deep learning techniques to predict induced currents, and the induced current splicing method (ICSM), which applies similar techniques to splice induced currents. The proposed method addresses large electrical size scattering problems by systematically decomposing them into smaller, more manageable subproblems. ICPM and ICSM are then employed to combine the solutions of the induced currents from the subproblems into a cohesive solution for the original large-size scattering problem. This approach significantly improves computational efficiency while maintaining high accuracy. Compared to FWNMs, ICPM improves speed of acquisition of accurate induced currents, while ICSM effectively merges the induced currents from multiple sub-targets into the induced current of the larger target. Extensive numerical tests show that ICPM provides high prediction accuracy and outperforms traditional methods in terms of both efficiency and hardware requirements. Furthermore, ICSM significantly improves the generalization ability of the method for solving scattering problems at large electrical sizes. The combined use of ICPM and ICSM offers a robust, resource-efficient solution for tackling large-scale electromagnetic scattering problems. This approach provides a promising direction for future applications in electromagnetic simulation and optimization.