Machine learning highlights chemistry as the key factor in metal–organic frameworks for atmospheric water harvesting

Abstract

Atmospheric water harvesting (AWH) using metal–organic frameworks (MOFs) offers a promising route to address freshwater scarcity in arid and off-grid environments. Yet, the structural and chemical factors that govern MOF performance remain insufficiently understood. Here, we combine high-throughput Grand Canonical Monte Carlo (GCMC) simulations with interpretable machine learning to study the structure–property relationships driving water uptake in MOFs. A chemically and structurally diverse set of 2600 frameworks was selected from the ARC-MOF database, and water uptake capacities were computed at 100% and 30% relative humidity. Among several regression models, Light Gradient Boosting Machine (LGBM) achieved the highest predictive accuracy. SHapley Additive exPlanations (SHAP) and correlation analyses identified adsorption energetics, local electrostatics (oxygen and hydrogen partial charges, metal electronegativity), and framework density as the dominant factors, with geometry acting as a secondary modulator. To provide an explicit analytical form for rapid screening and hypothesis generation, we constructed a second-order polynomial regression model using the top SHAP-ranked features. These results advance the fundamental understanding of water adsorption in MOFs and establish a scalable, data-driven framework for the rational design of high-performance materials for AWH.