Comparing classical and machine learning force fields for modeling deformation of solid sorbents relevant for direct air capture

Direct air capture (DAC) with solid sorbents such as metal-organic frameworks (MOFs) is a promising approach for negative carbon emissions. Computational materials screening can help identify promising materials from the vast chemical space of potential sorbents. Experiments have shown that MOF framework flexibility and deformation induced by adsorbate molecules can drastically affect adsorption properties such as capacity and selectivity. Force field (FF) models are commonly used as surrogates for more accurate density functional theory (DFT) calculations when modeling sorbents, but most studies using FFs for MOFs assume framework rigidity to simplify calculations. Although flexible FFs for MOFs have been parameterized for specific materials, the generality of FFs for reliably modeling adsorbate-induced deformation to near-DFT accuracy has not been established. This work benchmarks the efficacy of several general FFs in describing adsorbate-induced deformation for DAC against DFT. Specifically, we compare a common classical FF (UFF4MOF) with several machine learning (ML) FFs: M3GNet, CHGNet, MACE-MP-0, MACE-MPA-0, eSEN, and the Equiformer V2 model developed from the recent Open DAC 2023 dataset. Our results show that current classical methods are insufficient for describing framework deformation, especially in cases of interest for DAC where strong interactions exist between adsorbed molecules and MOF frameworks. The emerging ML methods we tested -- particularly CHGNet, MACE-MP-0, and Equiformer V2 -- appear to be more promising than the classical FF for emulating the deformation behavior described by DFT but fail to achieve the accuracy required for practical predictions.