Modeling CO2 Adsorption in Flexible MOFs with Open Metal Sites via Fragment-Based Neural Network Potentials

Metal-organic frameworks (MOFs) with open metal sites (OMS) are among the most promising porous materials for gas adsorption and separation, owing to their strong and selective interactions with guest molecules. However, simulating adsorption in such systems with high accuracy and efficiency remains a key challenge due to the need to model complex guest–MOF interactions and framework flexibility. Classical force fields often lack the precision to capture these effects, while ab initio methods are computationally prohibitive for large-scale, long-timescale simulations. In this work, we develop a neural network potential (NNP) trained on highly accurate DFT (PBE-D4/def2-TZVP) level data derived from a single representative fragment of the Mg-MOF-74 framework, a prototypical OMS-containing MOF, with CO2 molecules. Despite the limited training domain, the NNP accurately captured both intra- and intermolecular interactions in the CO2–Mg-MOF-74 system, including those involving the open metal sites. We integrated this NNP into a hybrid molecular dynamic (MD) and grand canonical Monte Carlo (GCMC) simulation workflow, enabling accurate modeling of CO2 adsorption in flexible MOFs. This approach allows us accounting for both framework dynamics and complex host–guest interactions with chemical accuracy and computational efficiency. Our results highlight the crucial role of framework flexibility in adsorption behavior and demonstrate that fragment-based NNP, when combined with advanced simulation techniques, offer a powerful and efficient approach for realistically modeling adsorption processes in MOFs with open metal sites.