Reproduce “On the Theoretical Limits of Learning Unconstrained Interatomic Potentials” 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
Prajwal Ananth, Shuwen Yue
The authors construct worst-case landscapes testing whether small target-weighted energy and force errors ensure molecular dynamics samples the correct distribution. They show kinetic trapping and wrong learned equilibria can yield total variation distance at least 1−e^−Ω(d), despite exponentially small errors. In harmonic argon at N=128, all four unconstrained trajectories leave the crystal basin; this does not establish practical MLIP failures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.