A Pragmatic Transfer Learning Approach for Oxygen Vacancy Formation Energies in Oxidic Ceramics

Lower oxygen vacancy formation energy is one of the requirements for air electrode materials in solid oxide cells applications. We introduce a transfer learning approach for oxygen vacancy formation energy prediction for some ABO3 perovskites from a two-species-doped system to four-species-doped system. For that, an artificial neural network is used. Considering a two-species-doping training data set, predictive models are trained for the determination of the oxygen vacancy formation energy. To predict the oxygen vacancy formation energy of four-species-doped perovskites, a formally similar feature space is defined. The transferability of predictive models between physically similar but distinct data sets, i.e., training and testing data sets, is validated by further statistical analysis on residual distributions. The proposed approach is a valuable supporting tool for the search for novel energy materials.