A data-driven approach, called materials informatics (MI) method, is developed to maximize the inference ability even using a small training dataset. The idea is to use a joint representation with the three descriptors to describe physical and chemical multifaceted perspectives of materials. This ensemble-based machine learning was trained with only 29 training data. Experiments confirmed that the virtual-screening process successfully discovered five oxygen-ion conductors, that have not been reported.
- Seiji Kajita
- Nobuko Ohba
- Ryoji Asahi
