Reproduce “HotPOT: Operator Transformer with Heterogeneous Physics Experts for Large-Scale PDE Pre-training” 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
Jiayi Li, Guancheng Zhou, Yisi Luo, Ruihua Chen, Penghao Jiang, Deyu Meng, Flora D. Salim, Piotr Koniusz
HotPOT predicts physical evolution across equation families using a shared history representation and family-specific learned predictors paired with fixed numerical updates. It trains these pairs together with next-state supervision. Across six settings and two activated-parameter scales, HotPOT beats MoE-POT in all 24 comparisons, with mean error reductions of 47.1% after pre-training, 54.3% after fine-tuning, and 50.7% overall. Tests cover regular grids and supported families, not arbitrary unseen equations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.