Multi-timescale temporal processing for ion-motion memristor-based physical reservoir computing
Sung Kwan Hwang, Do Young Yang, H Cho, Jung Ho Yoon
Abstract
Reservoir computing can process sequential and time-dependent data by simplifying training through output-only learning. Physical reservoir computing brings this framework into hardware, achieving faster inference and simpler architecture. Ion-motion memristors can be used to build physical nodes with nonlinear dynamics, time-dependent conductance, low-power switching, and compact size. However, most current implementations operate with a single relaxation time, limiting their ability to capture the diverse temporal features found in natural data. This perspective aims to address this challenge by examining memristor-based strategies for enabling multi-timescale processing. We review existing implementations and their limitations and explore potential approaches to modulate memristor relaxation time through material engineering, structural design, and external circuitry. By presenting insights and design strategies, we highlight the potential development of high-performance physical reservoir computing systems capable of handling complex, temporally diverse tasks.