Reliable methane storage screening and inverse design of metal organic frameworks under distribution shift

Abstract Adsorbed natural gas storage in metal-organic frameworks (MOFs) is a leading route to energy-dense methane storage, but the design space is too large for exhaustive simulation. We study the methane deliverable (working) capacity across a public database of 252,353 frameworks described by 14 geometric and 148 chemical descriptors, with uptake from grand canonical Monte Carlo simulation at 5.8 and 65 bar. The database is 97.58% computationally enumerated and 2.42% derived from experimentally reported crystal structures, a composition that shapes every result. A gradient boosting regressor reaches $$R^{2}=0.9963$$ (MAE 0.329 mmol/g) under a random split. Accuracy degrades only modestly under eight structure-aware protocols (topology-grouped $$R^{2}=0.9877$$ ), but training on enumerated frameworks and testing on experimental ones reduces the Spearman correlation from 0.998 to 0.879. Ablation shows 14 geometric descriptors alone recover $$R^{2}=0.9903$$ under a grouped split, so the chemical descriptors buy mainly ranking quality. Split-conformal intervals attain nominal marginal coverage (90.1 and 95.1%) but only 74.7% in the top capacity percentile, and 54.8−87.3% when a whole library is held out. Mondrian and shift-weighted variants restore near-nominal coverage within a fixed split. Reporting all three engineering units together, no framework meets the 263 cm 3 (STP) cm −3 volumetric target, while 1405 (0.56%) exceed 0.5 g g −1 . Benchmarked against direct screening, descriptor-space optimisation falls short: ranking recovers the 57.9 mmol/g record holder in a top-ten shortlist, whereas optimise-then-retrieve returns 17.7−28.8 mmol/g and none of the true top ten. Candidates are database-enumerated; no experimental validation is reported.