Scholarlink

<p>To address the challenges of physical interpretability and scalability in artificial neural networks, physical reservoir computing (PRC) has emerged as a promising research direction in contemporary machine learning. PRC tackles these challenges by leveraging simple and efficient physical systems to emulate neural network dynamics. In this work, we develop a theoretical dynamic model of coupled magnetic dipoles driven by an external magnetic field and systematically analyze its dynamic behaviors. We employed this model to construct a standard physical reservoir computing architecture, and revealed the mechanism of magnetic dipole system for computing. Subsequently, the computational performance is evaluated using regression and classification benchmarks. Our results demonstrate that the models’ rich nonlinear dynamic responses provide a robust foundation for achieving high accuracy in both chaotic time-series prediction and spoken digit recognition. The proposed contactless and nonlocal coupling architecture offers valuable theoretical insights for the advancement of PRC hardware development.</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

Nonlinear spatiotemporal dynamics in magnetic dipole systems for physical reservoir computing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月19日

查看全文
  • Liangshu He,
  • Yabin Jin,
  • Shulamit Edelstein,
  • Pedro David García,
  • Timon Rabczuk,
  • Xiaoying Zhuang,
  • Dani Torrent

摘要

To address the challenges of physical interpretability and scalability in artificial neural networks, physical reservoir computing (PRC) has emerged as a promising research direction in contemporary machine learning. PRC tackles these challenges by leveraging simple and efficient physical systems to emulate neural network dynamics. In this work, we develop a theoretical dynamic model of coupled magnetic dipoles driven by an external magnetic field and systematically analyze its dynamic behaviors. We employed this model to construct a standard physical reservoir computing architecture, and revealed the mechanism of magnetic dipole system for computing. Subsequently, the computational performance is evaluated using regression and classification benchmarks. Our results demonstrate that the models’ rich nonlinear dynamic responses provide a robust foundation for achieving high accuracy in both chaotic time-series prediction and spoken digit recognition. The proposed contactless and nonlocal coupling architecture offers valuable theoretical insights for the advancement of PRC hardware development.