Data driven modeling of magnetism in dilute magnetic semiconductors: correlation between the magnetic features of diluted magnetic semiconductors and electronic properties of the constituent atoms

We propose an efficient machine learning based approach in modeling the magnetism of diluted magnetic semiconductors (DMSs) leading to the prediction of new compounds with enhanced magnetic properties.

The approach combines accurate ab initio methods with statistical tools to uncover the correlation between the magnetic features of DMSs and electronic properties of the constituent atoms to determine the underlying factors responsible for the DMS-magnetism. Taking the electronic properties of different DMS systems as descriptors to train different regression models allows us to achieve a speed up of several orders of magnitude in the search for an optimum combination of the host semiconductor and the dopants with enhanced magnetic properties. We demonstrate this by analyzing a large set of descriptors for a wide range of systems and show that only 30% of these features are more likely to contribute to this property.

We also show that training regression models with the reduced set of features to predict the total magnetic moment of new candidate DMSs has reduced the mean square error by about 20% compared to the models trained using the whole set of features. Furthermore, our results indicate that the predictive power of our method can be improved even more by extending our descriptor set. (paper)