Bi-Mode Inverse Design of 3D Structures for MEMS Resonators

This work presents an automated algorithm for MEMS resonator structure generation based on inverse design, integrating deep learning and neural networks to predict key physical properties, including resonance frequency (f), quality factor of thermoelastic damping (Q<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${}_{\textit {TED}}$ </tex-math></inline-formula>), and motional impedance (Rx). Unlike traditional methods relying on finite element analysis (FEA), this approach leverages a database-driven deep learning model, achieving prediction speeds 9,740 times faster than the conventional FEA software with an average accuracy of 97.5%, 96.5%, 96.4 for f, Q<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${}_{\textit {TED}}$ </tex-math></inline-formula> and Rx,respectively. The algorithm supports flexural and Lamé modes and could generate resonators with a broad frequency range from ~8 to ~63 MHz, significantly surpassing existing methods. By efficiently predicting seed structures, the method guides the inverse design process, generating high Q, and low Rx resonator structures within 10 minutes. The generated devices exhibit deviations of less than 3% from target performance metrics. Simulations and experimental results validate the feasibility and effectiveness of the proposed algorithm, highlighting its potential for accelerating MEMS design with enhanced performance and precision.