Generalized Machine Learning Potentials for Predicting Low-Pressure Water Adsorption in Flexible Al-Based Metal-Organic Frameworks.

Abstract

Metal-organic frameworks (MOFs) are promising materials for adsorption and separation, but accurately predicting water adsorption remains a major challenge in molecular simulation. Classical force fields, such as UFF, often fail to capture the strong, directional hydrogen-bonding interactions between water and MOFs, while high-throughput density functional theory (DFT) calculations are computationally prohibitive. Here, we develop a transferable machine-learning potential (MLP) for water adsorption in Al-based MOFs by fine-tuning the pretrained MACE foundation model. Using a data set of 420 Al-MOFs with diverse chemical environments, we show that accurate prediction of water adsorption thermodynamics requires not only an improved description of water-framework interactions but also explicit treatment of framework flexibility. The resulting model reproduces experimental heats of adsorption and Henry coefficients for a series of benchmark Al-MOFs and achieves DFT-level accuracy across more than 400 Al-MOFs. Analysis of MIL-160 establishes practical accuracy requirements of approximately 10 kJ mol-1 in total energies and 40 meV Å-1 in atomic forces for reliable adsorption thermodynamics. Compared with DFT-level predictions, UFF underestimates water adsorption enthalpies by more than 10 kJ mol-1 for 74% of the materials studied. In addition, the MLP identifies lower-energy framework configurations than previously reported DFT-optimized structures, highlighting the importance of enhanced configurational sampling. These results demonstrate that foundation-model-based machine-learning potentials can enable DFT-accurate, high-throughput screening of water adsorption in flexible MOFs.