Unraveling the Humidity-Induced Phase Transition in CALF-20 via Machine Learning Potentials

CALF-20 is a metal-organic framework (MOF) that is known for its exceptional CO₂ selectivity under humid conditions. However, the molecular mechanism underlying its humidity-induced structural phase transitions and how water and CO₂ compete at the molecular level remain unclear. In this work, we developed a machine learning potential (MLP) with first-principles accuracy to investigate how water dynamics drive the transition from the open-pore (OP) to the closed-pore (CP) phase. We identify that water coordination at the Zn node forms hydrogen-bonded water dimers and trimers along the [011] diffusion channel, which are critical for the OP-CP phase transition. In the CP structure, these Zn-bound water networks exhibit residence times exceeding hundreds of picoseconds and occupy approximately half of the Zn nodes, reproducing previous experiments for the first time. Remarkably, preadsorbed CO₂ disrupts this water network. This effect is further supported by diffuse reflectance infrared Fourier transform spectroscopy, which reveals suppressed O-H stretching signals in CO₂ preloaded samples, indicating the inhibition of hydrogen-bonded water cluster formation. These findings provide a mechanistic explanation for the experimentally observed delay in the water uptake under competitive CO₂/H₂O adsorptions. Moreover, our MLP simulations accurately reproduce the water adsorption isotherm and experimental X-ray diffraction patterns of the α, β, τ, and γ phases, establishing a direct link between the microscopic structure and the macroscopic phase behavior observed in experiments. This study provides molecular-level insights into humidity-induced transitions in flexible MOFs and demonstrates a simulation framework for modeling guest-responsive behavior in soft porous materials.