Reproduce “Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers” 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

Luis Medrano-Navarro, Giacomo Baldan, Qiang Liu, Benjamin Holzschuh, Jan Hagnberger, Mathias Niepert, Nils Thuerey

GMP provides two pre-training strategies for engineering prediction: generated geometric descriptors for steady problems and online PDE solutions for transient ones, rather than stored collections of solved examples. On DrivAerML, GMP-Net reaches relative L2 error 0.346 versus 0.412 from scratch with 16 training samples. The gains vary by model and data regime, and complex low-data settings can still overfit.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.