Device-Oriented Design of Periodic Electromagnetic Structures via Physics-Regularized Surrogates

Abstract

This paper presents a device-oriented surrogate method for periodic slow-wave electromagnetic structures. Instead of directly learning a geometry-to-final-response mapping, the proposed method predicts physically meaningful cold descriptors, including the phase constant $\beta(f)$, coupling impedance $K_{z}(f)$, and attenuation-related loss quantities. These descriptors are then used for device-level evaluation through synchronism, coupling strength, and loss criteria. A physics-regularized neural surrogate is trained with a smoothness penalty to suppress non-physical oscillations in frequency-dependent predictions under limited data. The method is demonstrated using a G-band PCM-focused sheet-beam TWT interaction circuit with a staggered double-vane slow-wave structure. Full-wave descriptors extracted from a held-out geometry are used to illustrate the physical quantities learned by the surrogate and to identify the operating region associated with the reported high-power band of the validated G-band TWT. The proposed route links geometry, cold electromagnetic characteristics, and device-level metrics such as synchronism bandwidth, gain trend, and output-power-related performance.