Reproduce “Group Averaging for Physics Applications: Accuracy Improvements at Zero Training Cost” 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
Valentino F. Foit, David W. Hogg, Soledad Villar
Researchers empirically showed that applying group averaging as a post-processing step significantly boosts the accuracy of pre-trained machine learning surrogate models in physics simulations. This technique, requiring no model re-training or architectural changes, reduced prediction errors by up to 37% across various physical systems, enhancing visual quality and consistency.
Key insights
- Group averaging can confer the proven generalization benefits of equivariance to any pre-trained model for free, effectively projecting non-equivariant predictions onto the space of equivariant functions.
- Significant accuracy improvements are achievable even when approximating continuous symmetries (like S1 or T2) with a very small number of Monte Carlo samples, incurring minimal computational overhead.
- The effectiveness of group averaging is contingent on the underlying physical problem possessing the specified symmetry and the training data distribution respecting that symmetry.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.