Accelerating CO2 direct air capture screening for metal-organic frameworks with a transferable machine learning force field

Abstract

Direct air capture (DAC) of CO₂ is necessary for climate change mitigation, but it faces challenges from low CO₂ concentrations and competition from water vapor. Metal-organic frameworks (MOFs) hold significant promise for DAC owing to their high surface area and adsorption-based capture processes. However, identifying optimal MOFs is hindered by structural complexity and vast chemical diversity. Here, we introduced a machine learning force field (MLFF) tailored for CO₂ and H₂O interactions in MOFs by fine-tuning a foundation model. To address smoothing issues and catastrophic forgetting, we curated the GoldDAC dataset and introduced a continual learning scheme. We further developed DAC-SIM, a molecular simulation package integrated with MLFF, including a Widom insertion. Then, we screened an extensive MOF database, uncovering high-performing MOFs and identifying chemical features for DAC applications. This approach overcomes prior limitations in describing MOF-CO₂ and MOF-H₂O interactions, providing a scalable and accurate framework for DAC research of porous materials.