Structure Optimization of Large-Scale MEMS Micro-Hotplates Based on Finite Element Simulation and Machine Learning

Large-scale MEMS micro-hotplates for gas sensors face significant challenges in achieving uniform thermal distribution and low power consumption due to sensitivity to structural parameters. This study proposes an integrated approach combining finite element simulation with machine learning for parameterized design optimization. Initially, a thermo-electromechanical coupled model is constructed using COMSOL Multiphysics to perform gridded simulations on key parameters: cantilever beam width ($50-200 \mu ~\mathrm{m}$), support layer thickness (1-$10 \mu ~\mathrm{m}$), and heating electrode linewidth ($5-50 \mu ~\mathrm{m}$). This generates a comprehensive dataset encompassing temperature fields, power consumption, and stress displacement across hundreds of instances. A random forest regressor is employed to extract nonlinear parameter correlations and sensitivities, revealing dominant influences such as support layer thickness on temperature uniformity. Subsequently, a genetic algorithm facilitates dynamic multi-objective optimization, minimizing power ($<15 ~\text{mW}$) and temperature standard deviation ($<2 ~\mathrm{K}$) while constraining stress displacement ($<2 \mu ~\mathrm{m}$). The optimized parameters are validated using a multi-layer perceptron network trained with mean squared error loss augmented by L1 regularization (weight 0.01) for noise robustness. Results demonstrate optimized configurations (e.g., width $150 \mu ~\mathrm{m}$, thickness $5 \mu ~\mathrm{m}$, electrode $20 \mu ~\mathrm{m}$) yielding 25 % power reduction and 60 % improved uniformity, with efficiency gains of $\text{1 0 - 2 0}$ times over traditional methods and high accuracy (MSE $<1.0$). This closed-loop methodology enhances thermal and mechanical performance in simulation environments, offering a scalable framework for advanced MEMS design.