Are Atoms Enough? An Explainable Graph Neural Network for Carbon Capture and Gas Separation in Metal-Organic Frameworks

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:A novel GNN architecture with pore-aware representation and a trust framework that significantly improves prediction accuracy and interpretability for MOF gas separation properties.

🏢実務担当者:The public web tool enables rapid screening of MOFs for CO2 capture without deep ML expertise, and the trust labels guide experimental validation efforts.

🏛政策担当者:The research supports the development of advanced carbon capture materials, aligning with global CCUS deployment goals and Japan's Green Growth Strategy.

Are atoms enough to represent metal-organic frameworks (MOFs) in a graph neural network? Dominant crystal graph neural networks assume so, encoding each MOF entirely through its atoms and leaving pore space to emerge on its own. We introduce PoreGCN, a heterogeneous graph neural network in which Voronoi-derived pore vertices sit alongside atoms via dedicated atom-to-pore edges, giving the model a direct view of the cavities that govern adsorption and separation. Adding pore vertices produces large gains on the properties that matter most for screening. On the 2,737 real CoRE MOFs, the largest cavity diameter R2 rises from 0.07 under an atom-only baseline to 0.95 under PoreGCN. On the 51,163 hypothetical hMOF structures, PoreGCN reaches R2 between 0.91 and 0.99 across five geometric properties and R2 = 0.86 on log10(CO2/N2) selectivity. Retaining atoms alongside pores also unlocks two complementary attribution channels, the pore branch identifies which cavities drive a prediction, and the atom branch traces those cavities back to the metal nodes and linker chemistry responsible for them, something a pore-only model cannot do. Agreement between these two channels, combined with ensemble consensus, defines a four-scenario trust framework that separates individual predictions worth acting on from those requiring further validation, regardless of the model's population-level accuracy. Trustworthy selectivity predictions are 83.8% accurate at the 20% relative-error threshold against 8.3% for the untrustworthy subset, a roughly tenfold precision lift on the same model and test partition. The signal survives distribution shift, the condition that matters most in practice, maintaining a 1.3 to 2.0-fold enrichment on real CoRE MOFs as an independent external dataset, even where population-level calibration has been lost entirely. PoreGCN also recovers established chemistry independently, ranking fluorinated linkers and Zr-cluster secondary building units highest for CO 2 /N 2 selectivity. Predictions, trust labels, and per-atom attributions are available through a public web tool. https://huggingface.co/spaces/catenate/PoreGCN

• openalex https://doi.org/10.26434/chemrxiv.15004189/v1first seen 2026-06-04 05:01:52