Graph representational learning for bandgap prediction in varied perovskite crystals

Abstract

Perovskites are an important class of materials that are actively researched for applications in solar cells and other optoelectronic devices due to their ease of fabrication and tuneable bandgaps. High throughput computational techniques like Density Functional Theory (DFT) and Machine Learning (ML) are viable methods to accelerate discovery of new perovskite materials with favourable properties. ML specifically is faster and requires lesser computational power. We recognized the importance of having robust datasets for ML and hence collated a dataset of varied perovskite structures along with their indirect bandgaps. We employed a graph representational learning technique and trained a model that predicted bandgaps for all types of perovskites. The model has a mean absolute error of 0.28 eV and can predict bandgap in a few milliseconds. The metric of generalization gap is introduced to quantify the performance of ML models. This metric will help in building more generalized models that can predict properties for novel materials. Furthermore, we believe that these computational techniques should be user-friendly to those less experienced in the field. Hence, for researchers unacquainted with DFT or ML, we built a pipeline that abstracts the specific processes. This makes it easier for material scientists to quickly screen viable inorganic perovskite compounds allowing them to synthesize and experiment on the more promising compounds. © 2021 Elsevier B.V.