Machine Learning-Based\nPrediction of Proton Conductivity\nin Metal–Organic Frameworks

Recently, metal–organic frameworks\n(MOFs) have demonstrated\ntheir potential as solid-state electrolytes in proton exchanged membrane\nfuel cells. However, the number of MOFs reported to exhibit proton\nconductivity remains limited, and the mechanisms underlying this phenomenon\nhave not been fully elucidated, complicating the design of proton-conductive\nMOFs. In response, we developed a comprehensive database of proton-conductive\nMOFs and applied machine learning techniques to predict their proton\nconductivity. Our approach included the construction of both descriptor-based\nand transformer-based models. Notably, the transformer-based transfer\nlearning (Freeze) model performed the best with a mean absolute error\n(MAE) of 0.91, suggesting that the proton conductivity of MOFs can\nbe estimated within 1 order of magnitude using this model. Additionally,\nwe employed feature importance and principal component analysis to\nexplore the factors influencing the proton conductivity. The insights\ngained from our database and machine learning model are expected to\nfacilitate the targeted design of proton-conductive MOFs.