Examining proton conductivity of metal-organic frameworks by means of machine learning

The tunable structure of metal-organic frameworks (MOFs) is an ideal platform to meet contradictory requirements for proton exchange membranes: a key component of fuel cells. Nonetheless, rational design of proton-conducting MOFs remains a challenge owing to the intricate structure-property relationships that govern the target performance. In the present study, the modeling of quantities available for hundreds of MOFs was scaled up to many thousands of entities using supervised machine learning. The experimental dataset was curated to train multimodal transformer-based networks, which integrated crystal-graph, energy grid, and global-state embeddings. Uncertainty-aware models revealed superprotonic conductors among synthesized MOFs that have not been previously investigated for the application in question, thus highlighting magnesium-containing frameworks with aliphatic linkers as high-confidence candidates for experimental validation. Furthermore, classifiers trained on the activation energy threshold effectively discriminated between well-known proton conduction mechanisms, thereby providing physical insights beyond the black-box routine. Thus, our findings prove high potential of data-driven materials design, which is becoming a valuable addition to experimental studies on proton-conducting MOFs.