Scholarlink

<p>With the advancement of machine learning and deep learning, Physics-Informed Neural Networks (PINNs) have emerged as a prominent approach for solving partial differential equation (PDE) problems. In this article, we introduce a novel distillation framework specifically designed for PINNs, termed Self-Knowledge Distillation for PINNs (SKD-PINNs). Within this framework, knowledge distillation techniques are integrated into PINNs. Unlike traditional knowledge distillation approaches, the method is focused on regression tasks and transfers knowledge only when the predictions of the teacher model are reliable. Additionally, the SKD-PINNs framework utilizes self-distillation techniques, which iteratively extract knowledge from the model. This process not only enhances supervised information during training but also significantly speeds up network convergence. Notably, the method provides a versatile approach that enhances various existing PINN methodologies. The experimental results show that the method can significantly improve prediction accuracy and accelerate network convergence.</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

Self-knowledge Distillation Enhanced Universal Framework for Physics-Informed Neural Networks

  • 期刊论文

  • 发表日期：

    2025年1月30日

查看全文
  • Ying Li,
  • Jiawei Yang,
  • Dong Wang

摘要

With the advancement of machine learning and deep learning, Physics-Informed Neural Networks (PINNs) have emerged as a prominent approach for solving partial differential equation (PDE) problems. In this article, we introduce a novel distillation framework specifically designed for PINNs, termed Self-Knowledge Distillation for PINNs (SKD-PINNs). Within this framework, knowledge distillation techniques are integrated into PINNs. Unlike traditional knowledge distillation approaches, the method is focused on regression tasks and transfers knowledge only when the predictions of the teacher model are reliable. Additionally, the SKD-PINNs framework utilizes self-distillation techniques, which iteratively extract knowledge from the model. This process not only enhances supervised information during training but also significantly speeds up network convergence. Notably, the method provides a versatile approach that enhances various existing PINN methodologies. The experimental results show that the method can significantly improve prediction accuracy and accelerate network convergence.