Scholarlink

In this work, we introduce Poisson Hamiltonian Neural Networks (PHNNs) as an extension of Hamiltonian Neural Networks to better capture the dynamics of Poisson-Hamiltonian systems. By incorporating structure-preserving numerical methods, PHNNs can learn a wider range of dynamical systems beyond traditional symplectic models. We explored different training strategies, comparing Explicit Euler (EE) and Poisson-Hamiltonian Integrators (PHI). Our results showed that, while Euler-trained models offer better short-term accuracy, PHI-trained models stand out for their long-term stability and preservation of geometric structures. A hybrid approach – training with EE and testing with PHI – proved to be the best balance between accuracy and stability. These results highlight the potential of combining machine learning with geometric numerical methods to model complex dynamic systems without the need for explicit governing equations.

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

Poisson Hamiltonian Neural Networks: Structure-Preserving Learning of Dynamical Systems

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Adérito Araújo,
  • Gonçalo Inocêncio Oliveira,
  • João Nuno Mestre

摘要

In this work, we introduce Poisson Hamiltonian Neural Networks (PHNNs) as an extension of Hamiltonian Neural Networks to better capture the dynamics of Poisson-Hamiltonian systems. By incorporating structure-preserving numerical methods, PHNNs can learn a wider range of dynamical systems beyond traditional symplectic models. We explored different training strategies, comparing Explicit Euler (EE) and Poisson-Hamiltonian Integrators (PHI). Our results showed that, while Euler-trained models offer better short-term accuracy, PHI-trained models stand out for their long-term stability and preservation of geometric structures. A hybrid approach – training with EE and testing with PHI – proved to be the best balance between accuracy and stability. These results highlight the potential of combining machine learning with geometric numerical methods to model complex dynamic systems without the need for explicit governing equations.