Intelligent Prediction of EM Responses in mmWave Acoustic Resonators via Prompt-to-Structure Large Language Model and Regression Models

Higher-order asymmetric Lamb wave mode resonators employing interdigital transducers (IDTs) on LiNbO₃ thin films are promising candidates for mmWave acoustic filters. However, impedance matching in 50 Ω systems requires large static capacitance (C₀) and inevitably introduces structural inductance, leading to electromagnetic (EM) self-resonances that degrade selectivity and mode purity. While full-wave solvers such as HFSS can capture these effects, they are computationally expensive and demand substantial expertise. In this work, a lightweight, prompt-to-structure design framework is introduced based on a fine-tuned T5-small model trained on 420 fabricated A-mode resonators. The model converts natural-language prompts into device parameters, which are then processed by an ExtraTrees regressor to predict key EM responses, including self-resonance frequency (SRF). Validation on 85 Y-128° LiNbO₃ devices shows close agreement between predictions and measurements, with an average relative error of 0.03. The framework accelerates design iteration, reduces reliance on EM experts, and supports scalable mmWave acoustic filter development.