Smoothness Errors in Dynamics Models and How to Avoid Them
Edward Berman, Luisa Li, Jung Yeon Park, Robin Walters · 2026
Read Smoothness Errors in Dynamics Models and How to Avoid Them by Edward Berman, Luisa Li, Jung Yeon Park, Robin Walters (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-05352.txt.
Description
Neural networks that learn physics on curved surfaces work with meshes and graph convolutions, and those convolutions have a smoothing habit. Edward Berman and 3 co-authors study that habit in this 2026 machine learning paper, presented in its ICML 2026 camera-ready form. A known fix, unitary convolutions, preserves features exactly by construction, but the authors question whether preservation is what physics wants.
For diffusion-like processes, smoothness naturally grows over time, so a convolution forbidden from smoothing is overconstrained. The authors prove unitary convolutions hurt on such dynamics tasks, then propose relaxed unitary convolutions that balance preservation against the smoothing a system physically needs. Both versions are generalized from graphs to meshes. On heat and wave equations over complex meshes, and on weather forecasting, the relaxed method beat strong baselines including mesh-aware transformers and equivariant networks.
Catalog record
| CREATOR | Edward Berman, Luisa Li, Jung Yeon Park, Robin Walters |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 82 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-05352