Physics-Constrained Neural Networks for Electromagnetic Surrogate Modeling

Abstract—This paper presents a neural network-powered framework for accelerating electromagnetic simulations for early stage antenna design. The framework addresses the computational bottlenecks associated with conventional surrogate models and interpolation techniques by introducing compact, physics compliant surrogate models capable of predicting spherical wave expansion (SWE) coefficients and scattering parameters with high accuracy and in real-time. By leveraging the SWE representation, the framework ensures Maxwell-consistent predictions while providing memory-efficient field reconstruction at any point within the region of validity. The capabilities of the proposed frame work are demonstrated through a case study involving a multi layer patch antenna designed for Ka-band operation. The study highlights the framework’s ability to deliver accurate, real-time predictions of electromagnetic fields and scattering parameters across a multi-parameter design space, showcasing its potential as a scalable and efficient tool for antenna design engineers.
Index Terms—Machine Learning, Antenna Design, Computational Electromagnetics, Surrogate Modelling, Deep Learning.