Hysteresis-aware MEMS neuromorphic networks for embedded sensing and computation

Area of Science:

• Neuromorphic Engineering
• Materials Science
• Computer Science

Background:

• Microelectromechanical systems (MEMS) integrate sensing, memory, and computation on a single substrate.
• Nonlinear bistability in MEMS is often a limitation, hindering reliable device operation.
• Developing efficient analog computing architectures is crucial for advanced applications.

Purpose of the Study:

• To present a hysteresis-aware MEMS-based Continuous-Time Recurrent Neural Network (MEMS-CTRNN).
• To leverage nonlinear bistability as a computational resource rather than a limitation.
• To demonstrate the efficacy of MEMS devices as analog recurrent processors.

Main Methods:

• Embedding hysteresis into the training process of the MEMS-CTRNN.
• Utilizing nonlinear bistability inherent in MEMS devices for computation.
• Validating the MEMS-CTRNN using insect-scale robotic flight data for collision detection.

Main Results:

• The MEMS-CTRNN achieves temporal memory and noise robustness.
• It requires threefold fewer parameters compared to digital Long Short-Term Memory (LSTM) models.
• Demonstrated 93% accuracy in collision detection using robotic flight data.

Conclusions:

• MEMS devices are viable analog recurrent processors.
• The proposed framework enables MEMS-based neuromorphic architectures unifying sensing and computing.
• This approach has potential applications in microrobotics and in situ biomedical monitoring.