Energy-efficient reservoir computing with 10 x 10 crossbar array memristor for high performance multitask recognition

Abstract

Memristors hold significant potential for developing energy-efficient artificial intelligence (AI) hardware through parallel in-memory computing, thereby overcoming the long-standing von Neumann bottleneck. However, their widespread adoption is hindered by pronounced cycle-to-cycle (C2C) and device-to-device (D2D) variability. This study presents a novel approach to addressing key challenges in memristor-based artificial intelligence devices. We developed a 10 x 10 crossbar array of Fe50W50 hybrid nanocomposite memristors, demonstrating forming-free operation, low variability, and high reliability with low power consumption. The devices exhibit forming-free, low-variability, and highly reliable switching with ultra-low power consumption. The aligned grain boundaries within the nanocomposite enable well-controlled filament formation, ensuring consistent resistive switching characteristics. Leveraging these features, a reservoir computing (RC) architecture is implemented, demonstrating robust performance characterized by 4-bit input separability, short-term (fading) memory, and a strong echo-state property. The system achieves outstanding pattern-recognition accuracies of 98.79% for handwritten character recognition, 88.92% for garment classification, and 91.51% for digit recognition, along with 87.82% accuracy in multi-attribute classification and 98.62% in gesture recognition, underscoring its versatility in spatiotemporal processing. This material algorithm co-design framework not only enhances computational efficiency but also addresses core reliability challenges in memristor-based AI systems, paving the way toward scalable and energy-efficient neuromorphic computing architectures.