Interpretable machine learning for accelerating the discovery of metal-organic frameworks for ethane/ethylene separation

Abstract

Interpretable machine learning (ML) is applied to accelerate the discovery of promising metal–organic frameworks (MOFs) for the selective separation of ethane (C2H6) and ethylene (C2H4). Based on molecular simulation data, ML models are first trained and tested to classify MOFs into C2H4-selective and C2H6-selective categories using different types of material descriptors and fingerprints as input. It turns out that the model developed with the force-field inspired descriptors shows a higher classification accuracy. However, the PubChem fingerprint-based model is more interpretable because its binary variables directly indicate the existence or nonexistence of various substructures. Adopting this interpretable ML model, insightful structural characteristics are obtained and applied for efficient MOF screening to identify C2H6-selective candidates by means of substructure matching. Based on molecular simulation, 93.8% of the identified candidates are verified to be C2H6-selective with the best MOF presenting a high C2H6/C2H4 selectivity of 6.46. This work demonstrates the great potential of interpretable ML in accelerating the discovery of high-performance MOFs.