Surrogate-Driven Electromagnetic Geometry Learning with Measurement-Constrained Model Refinement

Abstract

One of the most significant tasks of modern radio-frequency (RF), microwave, and terahertz systems and systems is the efficient design, optimization, and inverse-derivation of structures and systems that critically depend on an accurate representation of higher-dimensional and complex electromagnetic (EM) geometries. Despite their high fidelity solutions, full-wave numerical solvers cannot be computationally characterised by high fidelity solutions because they are prohibitively expensive in terms of their practical use in iterative geometry discovery and real-time optimization. This paper would provide a solution to this problem by proposing a surrogate assisted electromagnetic-based geometry learning model augmented with measurement-constrained model refinement. The suggested system combines geometry-based learning, physics-aware surrogate modeling, and feedback to determine a rapid and redeemable substitute of traditional simulation-based processes. The deep surrogate model is conditioned to learn the nonlinear parametric extremes mapping between the representation of parametric geometry and the broadband EM responses, therefore, accomplishing the rapid prediction of important performance locations, including scattering parameters, impedance characteristics, and radiation behaviour, with substantially less computing burden. To reduce the effects of fabrication errors, material errors, and environmental interactions which contribute to the simulationmeasurement error, a measurement-constrained refinement algorithm is presented, where experimentally observed responses are enforced by constrained loss regularisation and adaptively updating the surrogate model. This is a refinement mechanism that gives physical consistency to the process and enhances model robustness without affecting the efficiency of the data. The refined surrogate is also used to learn and optimise surrogate-guided inverse geometry and surrogate-guided inverse geometry learning, which enables making direct synthesis of EM structures fulfilling given performance goals without redoing full-wave simulations. Massive numerical experiments and hardware-based validation outcomes show that the presented framework can drive significant decreases in the computational cost at a high predictive accuracy and a high level of the generalisation ability in unknown geometries. The given methodology provides scalable and practically feasible treatment of AI-assisted electromagnetic creation, as it allows developing antennas, metasurfaces and RF elements in the next-generation wireless and sensing models and prototyping them faster.