Scholarlink

<p>In this study, two types of resistive random-access memory (ReRAM) devices—ZrO<sub>x</sub> single-layer and ZrO<sub>x</sub>/TaO<sub>x</sub> bilayer structures—are fabricated and investigated for neuromorphic computing applications. The ZrO<sub>x</sub> single-layer device shows degraded performance due to limited conductance states. In contrast, the ZrO<sub>x</sub>/TaO<sub>x</sub>-based device, which incorporates a TaO<sub>x</sub> oxygen reservoir, exhibits gradual and self-compliant switching, high endurance exceeding 10⁹ cycles, and stable retention at 85&#xa0;°C. Under ± 1.3&#xa0;V, 10 ns pulse conditions, it demonstrates linear and symmetric long-term potentiation (LTP) and depression (LTD), enabling precise analog weight modulation. A two-layer multilayer perceptron (MLP) trained on the MNIST dataset achieves 93.31% classification accuracy with the ZrO<sub>x</sub>/TaO<sub>x</sub> device, comparable to the software baseline. In off-chip inference, analog weights mapped onto the convolutional layer of an Attention-on-Top Generative Adversarial Network (AOTGAN) yield natural image inpainting and high peak signal-to-noise ratio (PSNR). These findings reveal that TaO<sub>x</sub> integration is crucial for analog precision, stability, and reproducibility, demonstrating the potential of ZrO<sub>x</sub>/TaO<sub>x</sub>-based ReRAM as an energy-efficient building block for next-generation neuromorphic systems.</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

Fast and energy-efficient resistive switching in ZrOx/TaOx bilayer structures under ultra-low voltage operation for next-generation memory applications

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月3日

查看全文
  • Min Kyung Lee,
  • Jae-Seung Jeong,
  • Jaeho Jung,
  • Hyun Kyu Seo,
  • Changhwan Shin,
  • Gun Hwan Kim,
  • Min Kyu Yang

摘要

In this study, two types of resistive random-access memory (ReRAM) devices—ZrOx single-layer and ZrOx/TaOx bilayer structures—are fabricated and investigated for neuromorphic computing applications. The ZrOx single-layer device shows degraded performance due to limited conductance states. In contrast, the ZrOx/TaOx-based device, which incorporates a TaOx oxygen reservoir, exhibits gradual and self-compliant switching, high endurance exceeding 10⁹ cycles, and stable retention at 85 °C. Under ± 1.3 V, 10 ns pulse conditions, it demonstrates linear and symmetric long-term potentiation (LTP) and depression (LTD), enabling precise analog weight modulation. A two-layer multilayer perceptron (MLP) trained on the MNIST dataset achieves 93.31% classification accuracy with the ZrOx/TaOx device, comparable to the software baseline. In off-chip inference, analog weights mapped onto the convolutional layer of an Attention-on-Top Generative Adversarial Network (AOTGAN) yield natural image inpainting and high peak signal-to-noise ratio (PSNR). These findings reveal that TaOx integration is crucial for analog precision, stability, and reproducibility, demonstrating the potential of ZrOx/TaOx-based ReRAM as an energy-efficient building block for next-generation neuromorphic systems.