Reproduce “DeltaPhi: Physical States Residual Learning for Neural Operators in Data-Limited PDE Solving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xihang Yue, Yi Yang, Linchao Zhu
DeltaPhi is a framework that improves neural operators for solving partial differential equations by learning residuals between similar physical states rather than direct input-output mappings. This method leverages the stability of physical systems to implicitly augment training data, leading to enhanced generalization and prediction accuracy, particularly in scenarios with limited available data.
Key insights
- Leveraging the stability property of physical systems allows for the generation of physically valid residual training pairs, effectively expanding the training data distribution.
- Learning solution residuals is a smoother and more robust task for neural operators compared to learning absolute solutions, enhancing tractability.
- The approach implicitly augments the training distribution and discourages memorization of direct mappings, thereby mitigating overfitting in data-limited scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.