Temporal processing with a four-transistor one-capacitor analog short-term memristor for neuromorphic reservoir computing

Abstract

Temporal signal processing plays a vital role in modern artificial intelligence systems, where analysing dynamic patterns over time is essential for classification and prediction tasks. Reservoir computing (RC) provides an efficient neuromorphic framework for such applications by exploiting transient system dynamics while significantly reducing training complexity. In this work, we present a novel four-transistor analogue device that emulates short-term memory (STM) behaviour, similar to volatile memristors, and serves as a physical reservoir node. The device supports 16 discrete conductance states, each encoded over 4 bits, and its state evolution is governed by charge accumulation and natural decay, modulated via a complementary metal-oxide-semiconductor (CMOS)- compatible enable signal. Through comprehensive circuit-level implementation and modelling, we confirm that the device exhibits fading memory and nonlinear conductance dynamics essential for temporal encoding. We demonstrate the utility of this device in two benchmark artificial intelligence applications. First, we implement a hardware-inspired reservoir computing system for digit classification using temporally encoded five-by-four binary images. The reservoir state is generated using only five memristors, and classification is performed using a simple logistic readout, achieving 100% classification accuracy. Second, we apply the same device to predict the chaotic behaviour of the Lorenz attractor. Temporal pulse streams derived from quantized attractor states are processed through a 300-memristor reservoir, and a linear readout accurately forecasts future system dynamics, achieving a normalized mean square error (NMSE) of 0.031. The system generalises well beyond the training horizon, highlighting the capability of the proposed memristor-based reservoir in modelling complex temporal sequences.