Toward Intelligent Design and Measurement of MEMS Acoustic Wave Resonators

Surface acoustic wave (SAW) resonators are a cornerstone in 5G/6G radio frequency (RF) front ends owing to their compact form factor, strong electromechanical coupling, and compatibility with massive wafer-level manufacturability. Nevertheless, conventional design–test workflows remain constrained by extensive reliance on full-wave multiphysics simulations, iterative tape-out cycles, and exhaustive frequency-domain sweeps. Such processes incur high cost, long iteration cycles, and limited scalability, while spurious modes and parasitic effects are difficult to capture accurately. To overcome these challenges, this work establishes a data-driven design–test prediction framework that unifies modeling, validation, and optimization of SAW resonators within a single computational hierarchy. Training and evaluation use 14 883 industry-verified devices, 10 000 simulated structures, and 283 fabricated samples under a fixed split protocol. The system attains$R^{2} \gt 0.99$on key metrics and$10^{-3}$-level mse for full admittance spectrum reconstruction from structure via ensemble regressors. Furthermore, a tailored masked U-Net-1D performs sparse-spectrum recovery from 16 uniformly spaced points (a 98% reduction from 1024 points) while preserving$R^{2} \gt 0.98$on the 25% held-out test cohort. Beyond prediction, a physics-guided surrogate library enables inverse design. Given targets, the engine synthesizes previously unseen, spurious-suppressed resonators whose predicted spectrum aligns with measurements. Collectively, these elements bridge design and fabrication, jump out of physics and mathematical equations, sharply reduce simulation and test burden, and provide a scalable, physics-aware pathway to AI-driven automation of such microelectromechanical systems (MEMS) acoustic devices for next-generation RF applications.