Directional Prediction Errors Reveal Systematic Underprediction of Low-Pressure Carbon Dioxide Adsorption in Compact Metal–Organic Frameworks
Year: 2026
Type: preprint
Abstract: A prediction residual has both magnitude and sign, yet materials-screening studies usually treat only its magnitude as informative. Here we show that residual sign localizes systematic underprediction of low-pressure carbon dioxide (CO2) adsorption to a compact, low-accessibility region of metalorganic-framework (MOF) descriptor space. We analyze 7,306 MOFs in a topology-held-out test set, keeping related framework topologies out of training. A one-sided 90% calibration defines the expected underprediction tail; critically, that tail is not distributed uniformly across materials space. Strong-underprediction incidence falls from 27.6% to 4.0% across accessible-volume-fraction quintiles and from 26.5% to 2.7% across accessible-surface-area quintiles, while rising from 4.2% to 24.7% with density. Individual discrepancies reach 6.35 mmol g−1. Literature-derived structural annotations of leading Mg-and Ca-containing cases motivate the testable hypothesis that global geometric descriptors can miss favorable first adsorption environments in compact frameworks; they do not establish a distinct microscopic adsorption mechanism. A calibrated lower-bound screen provides a separate practical corollary for conservative candidate prioritization. The central result is not the expected size of a positive-error tail, but its structured location in descriptor space. Retaining residual direction therefore provides a chemically interpretable diagnostic of model failure and identifies the next discriminating calculation or experiment.
Source: ChemRxiv
Language: English
Subfield: Inorganic Chemistry
Field: Chemistry
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221534640