Descriptor Adequacy as a Materials-Regime Principle in MOF Adsorption Machine Learning

Machine learning is changing how metal--organic frameworks (MOFs) are screened for gas storage and separation, but its reliability depends on a basic question: does the descriptor contain the physics that controls the adsorption property? Here we show that descriptor adequacy is not a universal property of a feature set, but a property of the materials regime being studied. Using a topology-held-out benchmark of 255939 MOFs and 27 adsorption targets covering uptake, working capacity, and selectivity, we identify when simple pore descriptors are sufficient and when local chemical information is needed. Pore geometry is highly transferable for high-pressure, volume-controlled adsorption, especially methane storage and hydrogen-rich pre-combustion conditions. Local chemistry becomes important for low-pressure CO 2 and CH 4 uptake and for separation-derived targets, where framework--guest interactions control adsorption preferences. We also find that good regression accuracy does not always give reliable discovery: post-combustion selectivity can show high R 2 while still missing many top-ranked candidates. These results provide a process-resolved guide for matching descriptor complexity to adsorption mechanism, topology transfer, and screening objective. More broadly, the study establishes descriptor adequacy as a materials-regime principle for MOF informatics.