Scholarlink

<p>High-fidelity field reconstruction has been a focal point for many research studies, as the measured sensor data are often sparse and incomplete in both time and space. Physics-informed neural networks (PINNs) have been proposed to reconstruct fields using imperfect data, as they incorporate physical principles and thereby reduce reliance on the known sensor data. However, the placement of sensors remains crucial for optimizing PINNs, and existing studies have not sufficiently considered this aspect. Therefore, developing algorithms that intelligently improve sensor placement is of significant importance. In this study, we introduce a general approach that employs differentiable programming with attention modules to optimize sensor placement during the training of a PINNs model in order to improve field reconstruction. We evaluate our method using three distinct cases: the Allen-Cahn equation problem, the lid-driven cavity flow problem, and the cylinder flow problem to demonstrate our approach effectiveness in flow field inference, system identification, and its capability for multi-condition generalization. The results indicate that our method improves test scores and effectively learns the optimal layout of sensors for various Reynolds numbers, which advances our understanding of the relationship between sensor placement and reconstruction precision using PINNs.</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

Reconstruction of fields based on physics-informed neural networks with sensor placement optimization

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Bingteng Sun,
  • Shengze Cai,
  • Qiang Du,
  • Zinan Wang,
  • Yiling Chen,
  • Zhen Tian,
  • Junqiang Zhu

摘要

High-fidelity field reconstruction has been a focal point for many research studies, as the measured sensor data are often sparse and incomplete in both time and space. Physics-informed neural networks (PINNs) have been proposed to reconstruct fields using imperfect data, as they incorporate physical principles and thereby reduce reliance on the known sensor data. However, the placement of sensors remains crucial for optimizing PINNs, and existing studies have not sufficiently considered this aspect. Therefore, developing algorithms that intelligently improve sensor placement is of significant importance. In this study, we introduce a general approach that employs differentiable programming with attention modules to optimize sensor placement during the training of a PINNs model in order to improve field reconstruction. We evaluate our method using three distinct cases: the Allen-Cahn equation problem, the lid-driven cavity flow problem, and the cylinder flow problem to demonstrate our approach effectiveness in flow field inference, system identification, and its capability for multi-condition generalization. The results indicate that our method improves test scores and effectively learns the optimal layout of sensors for various Reynolds numbers, which advances our understanding of the relationship between sensor placement and reconstruction precision using PINNs.