@

초록

Neuromorphic computing aims to emulate the neural information-processing mechanisms of the human brain to achieve highly efficient and energy-conserving computation. Emerging non-volatile memory technologies have gained considerable attention as promising platforms for implementing artificial synaptic functionalities. In this work, a ReRAM 16 × 16 crossbar arrays based on a bilayer Ag/V2O5/SiO2/Pt/Ti device structure was fabricated using Photolithography and RF/DC magnetron sputtering techniques. The fabricated device exhibits low-power, high energy efficiency, and reliable synaptic functionality, which are essential for neuromorphic computing. To enhance the electrical performance, the device demonstrates stable and highly reproducible analog resistive switching (ARS) across repeated cycles, with SET and RESET operations at +1V and −1V, respectively. The synaptic functionality was systematically evaluated by applying successive potentiation and depression voltage pulses, using a read voltage of 0.1V and a pulse width of 50 μs. We propose a novel synapse-inspired neural network framework, SynaptiNet, enabling device-aware training of ReRAM crossbar arrays for neuromorphic computing. The proposed SynaptiNet-based neuromorphic CNN, equipped with the learning algorithm, demonstrates strong learning capability, achieving classification accuracies of 96.10% on the MNIST dataset and 76.15% on the Fashion-MNIST dataset. These results confirm the effectiveness of the adopted synaptic learning strategy and highlight the robustness of the proposed framework in providing reliable inference across both relatively simple and more challenging visual recognition tasks. Furthermore, the proposed synaptic devices provide a scalable and energy-efficient pathway for realizing on-hardware learning and cognitive functionalities in brain-inspired computing systems. © 2026 Elsevier Ltd and Techna Group S.r.l. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

키워드