An adaptive deep learning for cross‐condition adsorption prediction in metal‐organic frameworks

Abstract Accurate prediction of gas adsorption across diverse scenarios is essential for accelerating the discovery of metal–organic frameworks (MOFs) tailored for gas separation, capture, and storage. However, existing machine learning models face significant limitations in cross‐condition prediction and generalization capabilities. This work establishes Adp‐DeepSorption, a novel adaptive adsorption prediction framework for accurate multi‐task adsorption tasks. Adp‐DeepSorption incorporates environmental information into a Transformer‐based MOFs encoding module, first achieving a deep coupling between the material and conditions, and improving cross‐condition prediction capabilities. Comprehensive validation on a self‐built database demonstrates Adp‐DeepSorption's robust cross‐domain generalization, achieving R 2 of 0.86–0.96 across six extrapolation scenarios and uniquely maintaining R 2 ≥0.85 in all cases. Atomic‐level interpretability analysis shows that Adp‐DeepSorption will adjust its attention according to conditions. This study establishes a prediction method for dynamically sensing the adsorption condition for intelligent screening of multi‐condition adsorbents and provides a generalizable approach for incorporating environment into materials informatics.