Energy-Efficient Reservoir Computing Based on Solution-Processed Electrolyte/Ferroelectric Memcapacitive Synapses for Biosignal Classification
Sai Jiang, Jinrui Sun, Mengjiao Pei, Lichao Peng, Qinyong Dai, Chaoran Wu, Jiahao Gu, Yanqin Yang, Jian Su, Ding Gu, Han Zhang, Huafei Guo, Yun Li
Abstract
The classification of critical physiological signals using neuromorphic devices is essential for early disease detection. Physical reservoir computing (RC), a lightweight temporal processing neural network, offers a promising solution for low-power, resource-constrained hardware. Although solution-processed memcapacitive reservoirs have the potential to improve power efficiency as a result of their ultralow static power consumption, further advancements in synaptic tunability and reservoir states are imperative to enhance the capabilities of RC systems. This work presents solution-processed electrolyte/ferroelectric memcapacitive synapses. Leveraging the synergistic coupling of electrical double-layer (EDL) effects and ferroelectric polarization, these synapses exhibit tunable long- and short-term plasticity, ultralow power consumption (∼27 fJ per spike), and rich reservoir state dynamics, making them well-suited for energy-efficient RC systems. The classifications of critical electrocardiogram (ECG) signals, including arrhythmia and obstructive sleep apnea (OSA), are performed using the synapse-based RC system, demonstrating excellent accuracies of 97.8 and 80.0% for arrhythmia and OSA classifications, respectively. These findings pave the way for developing lightweight, energy-efficient machine-learning platforms for biosignal classification in wearable devices.