Machine learning potential for modelling dynamic hydrogen bond networks in MOF MIL-120

Metal-organic frameworks (MOFs) are porous materials with the potential for gas adsorption and separation technologies due to their tunable structural and chemical characteristics. However, simulating gas adsorption isotherms in MOFs with DFT-level accuracy remains a key challenge. In this work, we take MIL-120 as a case study and propose a complete computational workflow to fine-tune a pre-trained MACE potential, enabling accurate machine-learning interatomic potentials tailored for frameworks with dynamic structural behavior. Using this ML potential to accelerate sampling, we uncover a strong coupling between CO2 adsorption and the complex dynamic hydrogen-bond network on the MIL-120 pore surface. The presence of CO2 induces local configurational rearrangements that reshape the pore environment. This insight provides a new perspective for the structural design of MOFs for direct air capture (DAC), and offers a generalizable strategy for simulating adsorption in other flexible MOF systems.