Generalized Machine Learning Potentials for Water Adsorption in Al-Based Metal–Organic Frameworks Considering Framework Flexibility

Metal–organic frameworks (MOFs) are promising materials for gas adsorption and separation owing to their tunable structures and large capacities. Given the ubiquity of water vapor in the environment, predicting water adsorption within MOFs is critical, but the classical force fields, such as the UFF, often fail. To overcome this problem, we propose a strategy to fine-tune a machine learning potential (MLP) from the pretrained MACE model for water adsorption within MOFs. We test our strategy on a dataset of 420 Al-MOFs with 439 distinct linkers. The fine-tuned model can predict heats of adsorption and Henry coefficients of six well-known Al-MOFs, in good agreement with experiments. We also show that these accurate predictions require not only the accurate description of MOF-H2O interactions but also the framework flexibility, which are missing in previous screening studies. From a case study on MIL-160, we established the accuracy thresholds of 8 kJ/mol on total energies and 50 meV/Å on atomic forces for heat of adsorption calculations, and our final model achieves it on 409 Al-MOFs. Using this model, we can escape the local minimum and obtain optimal configurations with significantly lower DFT-energy than DFT-optimized ones. To identify optimal MOFs for water harvesting, we also applied this model to calculate DFT-level heat of adsorption. Compared with these DFT-level results, UFF underestimates the heat of adsorption by more than 3.5 kJ/mol in 92.4% of MOFs. Thus, our work demonstrates that the MLP can provide DFT-level predictions for high-throughput screening studies.