Reservoir‐Driven Neuromorphic Computing Based on Composite Rare‐Earth/Transition Metal Oxide Memristor

Résumé fourni par la source

ABSTRACT The growing demand for efficient hardware platforms for artificial intelligence has motivated the exploration of memristive devices capable of combining memory and computation. Here, a composite oxide memristor based on an Ag/Gd 2 O 3 :Nb 2 O 5 /Pt structure is fabricated via co‐sputtering and investigated for neuromorphic and reservoir‐computing demonstrations. The device exhibits stable bipolar resistive switching with a large memory window of ∼10 9 and relatively low operating voltages, showing average SET and RESET voltages of +0.16 ± 0.03 V and −0.07 ± 0.05 V, respectively. Pulse‐based electrical measurements show progressive conductance modulation and enable the emulation of synaptic plasticity behaviors, including long‐term potentiation (LTP), long‐term depression (LTD), paired‐pulse facilitation (PPF), and post‐tetanic potentiation (PTP). A 4‐bit pulse encoding scheme generates sixteen distinguishable electrical states that remain reproducible over repeated cycles and across multiple devices. The experimentally measured conductance responses are further incorporated into a device‐aware simulation framework for image classification using the CIFAR‐100 dataset. When combined with a CNN feature extractor, the reservoir‐enhanced model achieves 79% classification accuracy, approaching the 81% accuracy of the ideal digital baseline. These results show that experimentally measured memristor states can be incorporated into device‐aware learning frameworks, providing a practical approach for exploring memristor‐assisted neuromorphic computing strategies.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.