Adsorption regime changes the information needed for data-driven metal–organic framework screening
Year: 2026
Type: preprint
Abstract: How much fitting evidence, and which materials descriptors, are needed to prioritize porous materials? We analyze archived computational uptake labels for 279,010 metal–organic frameworks at four CO2/CH4 conditions (298 K). In retrospective experiments using target-informed example selection, each condition reaches at least 90% of its B = 1000 top-5% recall endpoint by B = 500. Local-chemistry RAC descriptors improve methane top-5% recall by 0.175 at 5.8 bar but only 0.039 at 65 bar, consistent with a pressure-dependent balance between chemical and geometric signals. These values describe predictor-fitting budgets, not total acquisition cost: the random-split setting also uses approximately 44,642 labeled structures for conformal calibration. Missing historical split identities and reference-informed candidate selection constrain transfer and shortlist claims. Thus, adsorption conditions appear to modify representation utility, while prospective label efficiency and physical attribution require independent, leakage-audited validation.
Source: ChemRxiv
Institutions: University of Cambridge, University of Tehran
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221605497