This paper presents a novel strategy for developing highly accurate and computationally efficient machine-learned force fields (MLFFs) for metal-organic frameworks (MOFs). The approach utilizes existing moment-tensor potentials (MTPs) and kernel-based potentials (VASP MLPs) and an active learning parametrization method. Benchmarking against DFT calculations and experimental data for various MOFs demonstrates near-DFT accuracy in predicting forces, structural parameters, elastic constants, phonon band structures, and thermal conductivity, while maintaining high computational efficiency. This advancement has the potential to significantly improve the computational modeling of MOFs.

Sandro Wieser, Egbert Zojer