Machine learning for imbalanced datasets: Application in prediction of 3d-5d double perovskite structures

3d-5d double perovskite is one kind of the most promising materials due to its novel properties and excellent performance with strong spin-orbit coupling. We used machine learning (ML) to search new 3d-5d perovskites. Three new methods derived from rescaling strategy in training models are used to solve the imbalanced problem which is more effective and more accurate than conventional machine learning models. Out of 664 scanned candidates of A2BB'O6 double perovskite (A is nonmagnetic metal, B and B' are 3d and 5d transition metals, respectively), we find 166 compounds can be stable 3d-5d double-perovskite structures, where A cations have to be +3 or +2 changed. Our results may throw light on the development of perovskite-based devices with strong spin-orbital couplings.