Deep Learning and Optimization Approaches in Improving the Thermo-Electro-Mechanical Responses of MEMS Resonant Accelerometers via a Novel Bidirectional Long Short-Term Memory: A Review

Abstract

Micro-Electro-Mechanical Systems (MEMS) resonant accelerometers have emerged as highly sensitive devices for precision sensing applications, including aerospace navigation, structural health monitoring, and consumer electronics. However, their performance is significantly affected by thermo-electro-mechanical (TEM) coupling effects, which introduce nonlinearities, drift, and instability under varying environmental conditions. Traditional compensation techniques based on analytical modeling and calibration often fail to capture complex temporal dependencies and nonlinear interactions inherent in such systems. Recent advancements in deep learning, particularly Bidirectional Long Short-Term Memory (BiLSTM) networks, offer promising solutions for modeling time-dependent behaviors in MEMS sensors. This review paper explores the integration of deep learning and optimization techniques to enhance the performance of MEMS resonant accelerometers by mitigating TEM-induced distortions. It provides a comprehensive analysis of existing literature focusing on machine learning-based compensation methods, optimization frameworks, and hybrid modeling approaches. Furthermore, the paper discusses the advantages of BiLSTM in capturing bidirectional temporal dependencies, improving prediction accuracy, and enabling real-time compensation. Challenges such as data scarcity, model generalization, and computational complexity are also examined. The study concludes by highlighting future research directions in combining physics-based models with data-driven approaches for robust and adaptive MEMS sensor systems.