A Hybrid Mechanism and Data-Driven Approach for Predicting Fatigue Life of MEMS Devices by Physics-Informed Neural Networks

Abstract

Our research group has focused on the long-term performance degradation of micro-electro-mechanical systems (MEMS) devices under mechanical, thermal, and electrical stresses. We previously used physics-based models for performance prediction, but the diverse materials, structures, and environmental conditions of MEMS devices affect their long-term stability. Traditional physics-based models struggle to account for these factors, limiting practical application. To address this, we employ data-driven methods to include more variables. This paper introduces a hybrid approach combining physical principles with data-driven techniques, specifically physics-informed neural networks (PINN), to model and analyze performance degradation in MEMS resonators. We compare this method with previous physics-based and data-driven approaches (support vector machine and random forest) under identical conditions. The hybrid method not only provides accurate predictions but also considers more operating conditions, enhancing practical application.