From Fixed to Arbitrary Frequency: A Scalable LLM-Enhanced Framework for Electromagnetic Device Design

Abstract

Conventional electromagnetic (EM) design is typically restricted to predefined frequency ranges. This article presents a language-driven generative framework for high-fidelity EM synthesis at arbitrary frequencies, incorporating physics-scaling-guided surrogate model (PSGSM) with vector to interval line preprocessing (VILP) and variable-controlled data collection (VCDC). The approach is defined by three methodological pillars. First, frequency scaling invariance is leveraged to decouple multiband responses into a normalized domain, enabling robust cross-frequency data/model reuse. Second, the VILP method transforms stochastic user vectors into structured, piecewise linear intervals, allowing the surrogate model to perceive continuous scaling trajectories rather than isolated points. Third, the VCDC strategy stabilizes the low-frequency resonance ( $f_{1}$ ) to isolate high-frequency variations ( $f_{2}$ ), ensuring data alignment with scaling logic and significantly enhancing sampling efficiency. The framework is validated through the design of a bandpass filter (BPF) and dual-passband frequency-selective surfaces (DP-FSSs). Experimental results from a fabricated DP-FSS prototype demonstrate high precision and flexibility in frequency-independent EM synthesis.