AI and the World's Hard Problems · Chemistry & materials · arXiv · 2026-10-20
Challenging Atomistic Foundation Models on Disordered Carbons
The authors test six foundation machine-learned interatomic potentials on disordered carbons, including versions with oxygen and hydrogen, to check how reliable they are outside their training domain. They screen the models on DFT energy and force errors, then run melt-quench-anneal molecular dynamics on four of them. MACE-MH-1 was the most consistently robust, and failures such as unphysical extrapolation appeared that low DFT errors did not reveal. All four models produced nanoporous carbon structures that match the main experimental signatures.