Scholarlink

<p>This work presents an electrically tunable physical reservoir computing (RC) system using two-terminal Pt/HfO<sub>2</sub>/TiN memristors. The device’s intrinsic relaxation time is effectively increased by over two orders of magnitude by modulating the base voltage between data pulses, enabling robust temporal data encoding for pulse intervals ranging from 0.14 to 100 ms. Enhanced separability is experimentally demonstrated through the classification of the Neuromorphic-MNIST (N-MNIST) dataset, surpassing the performance of conventional RCs without base voltage modulation. Frequency-domain analysis using multiple superimposed oscillator inputs shows controllable low-pass filtering characteristics. At the same time, the chaotic time-series prediction of the Mackey–Glass dataset achieves a normalized root mean square error of 0.27 without input data modulation and teaching signal for 200 timesteps. These results illustrate a feasible pathway toward tunable RC systems that achieve high performance while keeping the simplicity of a two-terminal device architecture.</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

Electrically tunable physical reservoir computing on two-terminal devices

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月27日

查看全文
  • Soo Hyung Lee,
  • Sungho Kim,
  • Janguk Han,
  • Sung Keun Shim,
  • Kunhee Son,
  • Dong Hoon Shin,
  • Hanyong Jeong,
  • Jinwoo Park,
  • Seong Jae Shin,
  • Yoon Ho Jang,
  • Cheol Seong Hwang

摘要

This work presents an electrically tunable physical reservoir computing (RC) system using two-terminal Pt/HfO2/TiN memristors. The device’s intrinsic relaxation time is effectively increased by over two orders of magnitude by modulating the base voltage between data pulses, enabling robust temporal data encoding for pulse intervals ranging from 0.14 to 100 ms. Enhanced separability is experimentally demonstrated through the classification of the Neuromorphic-MNIST (N-MNIST) dataset, surpassing the performance of conventional RCs without base voltage modulation. Frequency-domain analysis using multiple superimposed oscillator inputs shows controllable low-pass filtering characteristics. At the same time, the chaotic time-series prediction of the Mackey–Glass dataset achieves a normalized root mean square error of 0.27 without input data modulation and teaching signal for 200 timesteps. These results illustrate a feasible pathway toward tunable RC systems that achieve high performance while keeping the simplicity of a two-terminal device architecture.