Space group predictions in inverse materials search

Inverse materials search aims to search of wanted material system with desired properties that is often done with the help of modern searching algorithms. However, the searching speed is often limited by the huge parameter space. In this work, we proposed a machine-learning-assisted approach to enhance the material search speed. The dataset contains 51000 theoretically calculated electronic band structures of different crystals that includes metals, semiconductors, and insulators from Materials Project. A supervised learning is adopted to train neural networks to give predictions of 7 crystal systems and 230 space groups with bands degeneracies as input. As a result, over 90% and over 70% prediction accuracies are obtained for 7 crystal systems and 230 space groups respectively. Thus the predicted crystal system and space groups information are as extra input parameters to accelerate material search. The factors that affects neural network training are mainly due to orbital degeneracies and the lifted degeneracies by spin-orbital coupling.