Metal–Organic Frameworks for Water Harvesting: Machine Learning-Based Prediction and Rapid Screening

Atmospheric
water harvesting based on metal–organic frameworks
(MOFs) is an emerging technology to potentially mitigate water scarcity.
Because of the tremendously large number of existing MOFs, it is challenging
to find suitable candidates. In this context, a data-driven approach
to identify top-performing MOFs represents an important direction.
Herein, we develop a machine learning (ML) method to predict water
adsorption in MOFs and screen out top-performing MOFs for water harvesting.
First, experimental water adsorption isotherms in MOFs are collected
and water adsorption properties are extracted. Quantitative structure–property
relationships are analyzed in terms of pore structure and framework
chemistry, providing task-specific design principles. Then, ML models
are trained and interpreted to predict water adsorption properties
by using structural and chemical features, as well as operating conditions
as descriptors. The transferability of the ML models is validated
by out-of-sample predictions in seven newly reported MOFs. Finally,
the ML models are applied to screen ∼8000 “Computation-Ready,
Experimental” (CoRE) MOFs. Top-performing candidates are identified
including 149 MOFs with the maximum adsorption capacity ≥35
mmol/g, 39 MOFs with working capacity ≥10 mmol/g in a relative
pressure window 0.1–0.3, and 139 MOFs with working capacity
≥8.7 mmol/g in a relative pressure window 0.6–0.9. The
developed ML-based method would advance task-oriented design and rapid
discovery of reticular materials for energy and environmental applications.