Session: SYMP 6-10: Reservoir Computing and Control

Paper Number: 140364

140364 - A Physical Memcapacitive Reservoir Computing System for Energy-Efficient Temporal Data Processing 

Reservoir computing (RC) is an emerging, brain-inspired machine learning architecture that retrieves features in temporal input signals and maps them into higher dimensional state space for facilitated temporal data processing. While realized and extensively studied a-silico, physical reservoir layers have been implemented using various technologies such as ferroelectric transistors, silicon photonic modules, spintronic oscillators, atomic switch networks, and, most notably, volatile memristors. Although these devices have shown excellent performance in various tasks, including hand-written and spoken digit recognition, chaotic time series prediction, etc., their resistive nature leads to inherent energy dissipation, resulting in increased power and energy consumption. Alternatively, capacitive memory devices, also known as memcapacitors, offer a more energy-efficient approach due to their energy-storage nature. In this work, we leverage volatile lipid membrane-based memcapacitors, which closely mimic fundamental short-term synaptic plasticity functions, as reservoirs to solve classification tasks and analyze time series data, both experimentally and in simulation. Upon transmembrane potential stimulation, these ionically charged lipid bilayers vary their capacitance through geometrical changes resulting from electrowetting and electrocompression, i.e., an increase in the interfacial bilayer area and a decrease in the hydrophobic thickness, respectively. Our RC system accomplishes a testing dataset accuracy of 99.6% for classifying English spoken digits (‘0’ – ‘9’) and a prediction error of 0.000781 for predicting a second-order nonlinear dynamical system’s response. Furthermore, to demonstrate the device’s temporal data processing capabilities for real-time applications, we deploy our PRC system for epilepsy detection in an electroencephalography signal and achieve a testing dataset accuracy of 100%. In addition, we demonstrated that, in achieving these remarkable results, each memcapacitor maintains an average power of 415 fW for a pulse duration of 100 ms and consumes no more than 41.5 fJ of energy per spike, irrespective of the selected input voltage pulse duration. These energy consumption metrics are orders of magnitude lower than those achieved by state-of-the-art memristors deployed as physical reservoirs. In an era of rising demands for energy-efficient systems, such low-energy devices can be considered a vital asset and pave the way for energy-efficient solutions to temporal and classification problems. Finally, we believe that the softness, biocompatibility, and energy efficiency of our memcapacitors deem them highly suitable for signal processing and computing implementations in biological environments. The integration of these memcapacitors into real-world applications holds promise not only for enhancing computational efficiency but also for advancing our understanding of biological processes through the lens of innovative machine learning architectures.

 

Presenting Author: Ahmed Salah Mohamed The Pennsylvania State University

Presenting Author Biography: I am a fourth-year mechanical engineering PhD student and graduate research assistant at the mechanical engineering department at Penn State University. My work focuses on the development and use of bio-molecular materials for neuromorphic computing research and applications.

Authors:

Ahmed Salah Mohamed
Md Razuan Hossain
Nicholas X. Armendarez
Joseph S. Najem
Md Sakib Hasan