Physical reservoir computing system fully implemented using a single flash memory device via tailored decay pulse modulation

Abstract

With the rapid expansion of artificial intelligence (AI) applications, developing energy-efficient hardware capable of processing temporal data has become increasingly critical. In this work, we present a physical reservoir computing (RC) system fully implemented using a single TiN/Al₂O₃/Si₃N₄/SiO₂/poly-Si (TANOS) flash memory device. Unlike prior approaches that rely on multiple or heterogeneous devices, our system uniquely realizes both the reservoir and readout functionalities within a single device platform. By applying a tailored decay pulse scheme, we induce short-term memory (STM)-like dynamics in a device traditionally known for long-term memory (LTM), enabling dynamic reservoir state evolution essential for temporal signal encoding. The TANOS device demonstrates excellent endurance (>10⁵ cycles), low gate leakage (~10.06 nA), and high device uniformity, supporting reliable and low-power operation, with the operation possessing the highest energy consumption (erase) consuming only 513.1 pJ per pulse at room temperature. When integrated into a CNN-based RC framework, the system achieves a high classification accuracy of 88.38% on the Fashion MNIST dataset and maintains strong performance in a fully hardware-oriented MNIST simulation. These results highlight the potential of standard silicon memory technology for building compact, energy-efficient, and fully self-contained neuromorphic computing systems, paving the way for scalable and CMOS-compatible AI hardware using a single memory device.