AI-driven discovery of high-performance metal–organic frameworks for next-gen atmospheric water harvesting

Abstract

Metal–organic frameworks (MOFs) offer unique advantages for atmospheric water harvesting (AWH) due to their high porosity and tuneable water adsorption behavior. However, their vast structural diversity hinders efficient candidate selection. Here, a machine learning (ML) framework is developed for predictive modeling and high-throughput screening of MOFs, leveraging over 500 000 data points. Specifically, a training set of 88 experimental water isotherms across 64 unique MOFs was curated, and it was complemented with 3015 adsorption data points for nitrogen, carbon dioxide, and methane drawn from open databases. These gas adsorption metrics act as proxies for pore hydrophilicity, allowing the model to learn transferable structure–property relationships. Our ensemble meta-model, combining diverse base learners, achieves R² ≈ 0.96–0.99 on internal cross-validation and, under the stricter leave-one-MOF-out protocol, retains R² = 0.93–0.98 across the ten targets. Against three independently measured external frameworks, the meta-model achieves R² = 0.95 and MAE = 2.87 mmol g⁻¹, providing a realistic estimate of transferability to unseen MOFs together with a residual-based uncertainty envelope of approximately ±8 mmol g⁻¹. We further translate the model into an open-access tool that enables researchers to predict water uptake for user-defined MOFs and conditions. This approach significantly reduces the experimental and computational burden of MOF screening and accelerates AWH material discovery. By bridging data-driven predictions with practical usability, our work provides a scalable platform for advancing MOF-based water harvesting technologies.