Deep Learning Approach to Material Properties

Abstract: In this thesis, we consider a deep learning approach to predict material properties. Primarily we study artificial neural networks (ANN), which predict the energy distance to the convex hull (measure of stability) of perovskites. Further, we explore if the networks can be generalised to predict band gaps and unit cell volume. We also demonstrate total energy calculations using density functional theory as it is essential for building the datasets required for ANN studies. More detailed, we re-implement an ANN architecture proven successful at predicting formation energies and reproduce the corresponding results. This ANN model is then directly applied to predict the distance to the convex hull for the perovskites. We explore further improvements to the input representation and the network itself. In the following we use transfer learning to generalise the model to predict band gaps and unit cell volume. While reproducing the results of the re-implemented model, the results are considerably improved. For the perovskites, we find that our deep learning model achieves significantly better performance than conventional machine learning methods. This improved model allowed for a high-throughput search of tetragonal quaternary perovskites, which resulted in identifying 437 new potentially stable quaternary perovskites. In line with previous research, we find that the models performance varies over the periodic table. While the transfer learning for the unit cell volume is successful, band gaps were more difficult to predict.