Nonlinear Thermal Drift Compensation IN Hybrid On-Chip Mems Piezoelectric-Capacitive Sensors Using Machine Learning-Assisted Modeling

Abstract

Thermal drift remains a critical challenge in MEMS sensors, where nonlinear material responses under wide temperature ranges introduce bias and instability. This work presents a simulation-to-system framework for hybrid piezo-capacitive MEMS sensors, combining finite element multiphysics modeling with machine learning–based drift compensation. FEM simulations incorporating nonlinear temperature dependencies of silicon and AIN generated datasets across –40 °C to +125 °C, capturing capacitance variations, piezoelectric charge responses, and resonance frequency shifts. Extracted features were used to train ridge regression, support vector machines, and shallow neural networks, with performance validated on unseen thermal cycles.

Results show that ridge regression achieved stable but less accurate compensation (R² = 0.91, RMSE = 0.38), SVMs provided improved nonlinear mapping (R² = 0.95, RMSE = 0.24), and shallow neural networks delivered the highest accuracy (R² = 0.98, RMSE = 0.12). System-level FPGA/ASIC simulations confirmed feasibility of real-time compensation under resource constraints. The framework eliminates dependence on early hardware prototypes, reducing cost and accelerating design cycles. This study establishes FEM-driven datasets as a reliable foundation for compensation model design, paving the way for future hardware validation and cross-material generalization.