Scholarlink

<p>Self-rectifying memristors can significantly suppress the sneak current and lead to the development of a highly dense, power and area-efficient passive crossbar array, facilitating parallel analog computation in neuromorphic computing systems. This study reports a selector-less, forming-free Ag<sub>2</sub>O<sub>1-δ</sub>-based 64 × 64 (4 kb) passive memristive crossbar array. The fabricated memristive cells exhibit self-rectifying (rectification ratio &gt;10<sup>3</sup>) analog bipolar switching characteristics with negligible spatiotemporal variations. Multilevel (&gt;4-bit) conductance states (~3.60 μS to ~22.30 μS at 373 K) were achieved through controlled compliance current (20 μA ≤ I<sub>CC</sub> ≤ 1 mA). Based on the robust synaptic characteristics of the memristors assessed by the excitatory/inhibitory electrical signals, the applicability of the crossbar array was simulated via a multi-layer perceptron (MLP) model on the MNIST dataset, achieving 96.08% recognition accuracy under the device constraints. To further enable accurate and noise-robust convolutional neural network (CNN) deployment on memristor hardware, three training strategies such as Post-Training-Quantisation, Straight-Through-Estimator, and Alpha-Blending were evaluated. Alpha-Blending emerged as the most effective and robust strategy, enabling a ResNet-18 model to achieve an accuracy of 94.08% on the CIFAR-10 task and 75.28% on the CIFAR-100 task, with drops of only 1.07% and 3.2%, respectively, under moderate levels of additive Gaussian noise (σ = 0.10).</p><p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Self-rectifying Ag2O1-δ-based synaptic crossbar array for neuromorphic computing applications

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月2日

查看全文
  • Bishal Kumar Keshari,
  • Chetan Kodand Reddy Madadi,
  • Sanghamitra DebRoy,
  • Akshay Salimath,
  • Venkat Mattela,
  • Parikshit Sahatiya

摘要

Self-rectifying memristors can significantly suppress the sneak current and lead to the development of a highly dense, power and area-efficient passive crossbar array, facilitating parallel analog computation in neuromorphic computing systems. This study reports a selector-less, forming-free Ag2O1-δ-based 64 × 64 (4 kb) passive memristive crossbar array. The fabricated memristive cells exhibit self-rectifying (rectification ratio >103) analog bipolar switching characteristics with negligible spatiotemporal variations. Multilevel (>4-bit) conductance states (~3.60 μS to ~22.30 μS at 373 K) were achieved through controlled compliance current (20 μA ≤ ICC ≤ 1 mA). Based on the robust synaptic characteristics of the memristors assessed by the excitatory/inhibitory electrical signals, the applicability of the crossbar array was simulated via a multi-layer perceptron (MLP) model on the MNIST dataset, achieving 96.08% recognition accuracy under the device constraints. To further enable accurate and noise-robust convolutional neural network (CNN) deployment on memristor hardware, three training strategies such as Post-Training-Quantisation, Straight-Through-Estimator, and Alpha-Blending were evaluated. Alpha-Blending emerged as the most effective and robust strategy, enabling a ResNet-18 model to achieve an accuracy of 94.08% on the CIFAR-10 task and 75.28% on the CIFAR-100 task, with drops of only 1.07% and 3.2%, respectively, under moderate levels of additive Gaussian noise (σ = 0.10).