Reliability of universal ML interatomic potentials in real-world applications

Determine the reliability and effectiveness of universal machine learning interatomic potentials (uMLIPs) when applied to practical, real-world atomistic simulation tasks, establishing whether these models can robustly support downstream simulations and property predictions beyond fitting energies and forces.

Background

Universal machine learning interatomic potentials (uMLIPs) promise near–quantum accuracy with far lower computational cost and broad applicability across materials classes. Yet, most prior assessments emphasize training-aligned metrics (energy/force/stress fitting) rather than end-to-end, domain-relevant tasks encountered in realistic simulations.

This work focuses on challenging nanoporous materials, particularly metal–organic frameworks (MOFs), where porosity, diverse chemistries, and coordination bonds stress-test generalization. The stated open question motivates the development of a task-oriented benchmark (MOFSimBench) to assess whether uMLIPs are ready for practical deployment in such settings.