Prediction and exploration of proton conductivity using machine learning in proton-conducting ceramics

Abstract

Proton-conducting ceramic fuel cells, a type of solid-oxide fuel cell, are being considered for use at 300°C-600°C due to their electrolyte properties. Therefore, we used machine learning to identify new materials that exhibit high proton conductivity at low temperatures. We constructed a database comprising proton (726 data points from 76 perovskite-type oxide compositions) and hole conductivity data (738 data points from 103 perovskite-type oxide compositions). Using gradient boosting, we predicted proton conductivity based on various descriptors, including ionic radius and electronegativity. The structure of the virtual composition was perovskite-type (ABO3), and the host elements of the A and B sites were assumed to be Ba, Sr, and Ca and Co, Mn, Ni, and Fe, respectively. The additive elements of the B site were assumed to be Er, Gd, In, Lu, Ru, Sc, Y, and Yb. Hypothetical compositions, BaMn(1-x)/2Y(1-x)/2InxO3 (x = 0.1, 0.3, 0.5, 0.7), based on predicted high proton and low hole conductivities were synthesized using the liquid phase method. The isotope effect on electrical conductivity was confirmed at 800°C-400°C, using 1.9%H2O (or D2O)-1%O2-Ar, and observed at low temperatures (400°C) for x = 0.3, 0.5, and 0.7. Therefore, BaMn(1-x)/2Y(1-x)/2InxO3 (x = 0.3, 0.5, 0.7) was found to be a proton conductor. However, the proton conductivity was found to be low, below the predicted value.