Scholarlink

<p>Quantum machine learning in data-intensive science is limited by the difficulty of interfacing high-dimensional, chaotic classical data with resource-limited quantum processors. To address this issue, we study a physics-informed Koopman-quantum hybrid framework motivated by a representation-level correspondence between Koopman operator evolution and quantum evolution. On this basis, we build a compact hybrid pipeline in which Koopman analysis distills raw waveforms into low-dimensional residual statistics that are then processed by a modular parallel quantum neural network. We test the framework on 4763 labeled channel sequences from 433 tokamak discharges. The main PQNN benchmark uses classical preprocessing followed by noiseless circuit simulation and achieves about 97.0% accuracy in screening corrupted diagnostic data while using far fewer trainable parameters than a deep CNN baseline. Additionally, we complement this benchmark with two simplified, hardware-motivated noise sources: finite-shot/readout uncertainty and effective perturbations to PQNN expectation values. These checks indicate qualitative stability under moderate device-relevant perturbations, while also highlighting the remaining gap to platform-specific hardware validation.</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

Validating a Koopman-quantum hybrid paradigm for diagnostic denoising of fusion devices

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Tie-Jun Wang,
  • Run-Qing Zhang,
  • Ling Qian,
  • Yun-Tao Song,
  • Ting Lan,
  • Hai-Qing Liu,
  • Keren Li

摘要

Quantum machine learning in data-intensive science is limited by the difficulty of interfacing high-dimensional, chaotic classical data with resource-limited quantum processors. To address this issue, we study a physics-informed Koopman-quantum hybrid framework motivated by a representation-level correspondence between Koopman operator evolution and quantum evolution. On this basis, we build a compact hybrid pipeline in which Koopman analysis distills raw waveforms into low-dimensional residual statistics that are then processed by a modular parallel quantum neural network. We test the framework on 4763 labeled channel sequences from 433 tokamak discharges. The main PQNN benchmark uses classical preprocessing followed by noiseless circuit simulation and achieves about 97.0% accuracy in screening corrupted diagnostic data while using far fewer trainable parameters than a deep CNN baseline. Additionally, we complement this benchmark with two simplified, hardware-motivated noise sources: finite-shot/readout uncertainty and effective perturbations to PQNN expectation values. These checks indicate qualitative stability under moderate device-relevant perturbations, while also highlighting the remaining gap to platform-specific hardware validation.