Accelerated prediction of perovskite material properties with classical machine learning and graph neural network

Abstract

Perovskite materials, possessing a plethora of exceptional properties, have garnered significant attention. Nonetheless, owing to their intricate structure and chemical composition, several obstacles remain in the preparation, characterization, and application of perovskite materials. The rapid advancement of machine learning technologies has expedited research on perovskite materials in recent years. This technique aids researchers in rapidly screening and optimizing the properties of perovskite materials, while also uncovering hidden patterns and trends from vast amounts of experimental and computational data. In this paper, we designed traditional machine learning models built upon manual descriptors, as well as structure-based graph neural network (GNN) models, to precisely forecast various characteristics of perovskite materials. Our results demonstrate that the end-to-end GNN model performs exceptionally well for various properties when sufficient data is available. This highlights the versatility and value of the method in expediting the development of novel perovskite materials.