Predicting oxygen gap formation using machine learning

We have developed a new transfer learning approach for predicting the energy to form oxygen vacancies for some perovskites, and used it to identify new potential materials for fuel cell cathodes.

Reducing CO2 emissions is a top priority for economies around the world in order to meet environmental commitments. Therefore, the development of new, cost-effective energy conversion technologies is of great importance. One of these technologies is the solid oxide cell (SOC), which either converts chemical energy into electrical energy with high efficiency (SOFC: Solid Oxide Fuel Cell) or uses renewable energy to produce hydrogen or hydrogen-based synthetic fuels (SOEC: Solid Oxide Electrolysis Cell). In addition, the SOC in a device can also be operated in both modes alternately (rSOC: reversible solid oxide cell). Major energy sources include hydrogen, natural gas, biogas and other renewable fuels, and the resulting reduction in CO2 emissions is an important argument for large-scale deployment of this technology. However, the degradation effects in both modes of operation remain an important research topic. In SOFC operation in particular, degradation of the air electrodes has emerged as a major issue limiting stack lifetime and durability and posing a critical challenge for expanded application.

(La,Sr)(Co,Fe)O3-δ (LSCF) is one of the most important air electrode materials for SOC applications. However, Sr is a very reactive element and contributes to various degradation problems. Therefore, it is desirable to find new Sr-free air electrode materials. The potential materials should be chemically and mechanically compatible with adjacent components, have good tolerance to impurities in the surrounding atmosphere, and have low energy to form oxygen vacancies to ensure rapid surface oxygen exchange and diffusion. Perovskite oxides (ABO3) with cubic symmetry have shown high potential as air electrode materials for SOC applications. In general, the large number of candidates for potential air electrode materials makes a systematic search for new and better materials difficult or even impossible. Therefore, guidance by machine learning tools, especially artificial neural networks (ANNs), to optimize material properties is a promising approach.

We have developed a new transfer learning approach for predicting the energy to form oxygen vacancies for some ABO3 perovskites, from a two species doped system to a four species doped system. An artificial neural network is used for this purpose. A training data set with two species is used to train prediction models for determining the energy of oxygen vacancy formation. A formally similar feature space is defined for predicting the oxygen vacancy formation energy of perovskites doped with four elements. The transferability of the prediction models between physically similar but different data sets, i.e., training and test data sets, was validated by further statistical analysis of the residual distributions. The proposed approach is a valuable tool in the search for new energy materials.

Reference: Yin, X.; Spatschek, R.; Menzler, N.H.; Hüter, C. A Pragmatic Transfer Learning Approach for Oxygen Vacancy Formation Energies in Oxidic Ceramics. Materials 2022, 15,2879. https://doi.org/10.3390/ma15082879