Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Unisolver: PDE-Conditional Transformers Towards Universal Neural PDE Solvers

Created by
  • Haebom

Author

Hang Zhou, Yuezhou Ma, Haixu Wu, Haowen Wang, Mingsheng Long

Outline

This paper presents Unisolver, a general-purpose neural network-based PDE solver capable of solving a wide range of partial differential equations (PDEs). Existing neural network-based PDE solvers are limited to specific PDEs or a limited set of coefficients, resulting in poor generalization performance. Unisolver leverages the mathematical structure of PDE solutions to flexibly integrate PDE components, such as equation symbols and boundary conditions, into a transformer model as domain- and point-specific conditions. Trained on diverse data, Unisolver demonstrates state-of-the-art performance and generalization capabilities on three large-scale benchmarks. The source code is available on GitHub.

Takeaways, Limitations

•
Takeaways:
◦
Overcoming the limitations of existing methods by presenting a general-purpose neural network-based PDE solver for various PDEs.
◦
Improving the generalization performance of models by leveraging the mathematical structure of PDE solutions.
◦
Achieving cutting-edge performance in large-scale benchmarks.
◦
Increased reproducibility and scalability of research by providing open source code.
•
Limitations:
◦
Potential increase in computational cost due to the complexity of the transformer model.
◦
The generalization performance of a model depends heavily on the diversity of the training data.
◦
Performance for extremely complex or special PDEs requires further study.
👍
Made with Slashpage