Benchmarking Universal Machine-Learned Interatomic Potentials for High-Temperature Metal-Organic Framework Chemistry

TL;DR

This paper benchmarks various machine-learned interatomic potentials for high-temperature metal-organic frameworks, revealing their strengths and limitations in simulating extreme thermal conditions.

Contribution

It introduces a high-temperature AIMD dataset for nine MOFs and evaluates the performance of five leading uMLIPs, highlighting current model limitations.

Findings

ORB-v3 and fairchem OMAT achieve lowest errors across temperatures

All models show significant errors at high temperatures

Long-timescale MD with ORB-v3 reveals generative errors exceed static validation losses

Abstract

Universal machine-learned interatomic potentials (uMLIPs) offer a promising approach to performing atomistic simulations at near-DFT accuracy with greatly reduced computational cost. Here, we present a new high-temperature benchmarking dataset of 40~ps ab~initio molecular dynamics (AIMD) trajectories simulated at 300, 1000, and 2000 K for nine zinc- and zirconium-based metal-organic frameworks (MOFs): ZIF-8, CALF-20, MOF-10, MOF-5, MIP-206, UiO-66, UiO-67, UiO-66-NH2, and NU-1000. These trajectories capture equilibrium dynamics, thermally induced distortions, and early-stage decomposition events, including linker degradation and metal node aggregation. Subsequently, we use this dataset to benchmark five leading uMLIPs: ORB-v3, MACE-MP-0a, MACE-MPA-0, fairchem ODAC23, and fairchem OMAT. Our results reveal that ORB-v3 and fairchem OMAT achieve the lowest energy, force, and stress errors across all temperatures. However, all models exhibit significant error under high-temperature conditions. Long-timescale molecular dynamics simulations produced with ORB-v3 demonstrate that the generative error of uMLIPs far exceeds model losses captured during static validation, highlighting the limitations of current universal models for simulating high-temperature MOF dynamics. This work provides a benchmark for assessing the robustness of uMLIPs in extreme regimes and guides future development of potentials capable of accurately modeling the chemistry of high-temperature MOF dynamics.