Real-world reliability of universal MLIPs

Determine the reliability and effectiveness of universal machine learning interatomic potentials (uMLIPs) when applied to practical, real-world atomistic simulations, including applications to chemically diverse nanoporous materials such as metal–organic frameworks, covalent organic frameworks, and zeolites.

Background

Universal machine learning interatomic potentials (uMLIPs) promise quantum-level accuracy at significantly reduced computational cost, enabling large-scale atomistic simulations. Despite impressive results on benchmark fitting tasks and selected demonstrations, their performance and robustness in realistic deployment scenarios—especially on out-of-distribution materials and downstream tasks—require systematic evaluation.

The paper targets nanoporous materials, particularly MOFs, which present challenges stemming from diverse chemistries, large unit cells, coordination bonds, and porosity. The authors introduce MOFSimBench to assess uMLIPs on tasks central to practical modeling, including structural optimization, molecular dynamics stability, bulk property prediction (bulk modulus and heat capacity), and host–guest interactions. The stated open question motivates the need for such domain-specific benchmarking to inform adoption and development.