Crystal Level Features Developed Using Edge Prediction on Graphs Derived from Crystals

Deep learning techniques that use graph representation learning have shown promising results in property prediction of molecules and materials. Instead of feature engineering, inputs to these models more closely resemble the atomic structure of the data as is. However, it is well-known that deep learning techniques require a lot of data to reach up to their potential, which is the case for a variety of different classes, including hybrid organic-inorganic perovskites (HOIP). These HOIP materials, also known as metal halide perovskites, are highly promising solar cell materials that have very high solar efficiency, and can be fabricated largely at room temperature. They could become an attractive alternative to silicon solar cells, once processing and reliability issues are addressed. To circumvent the dearth of data, we used a pre-trained Crystal Graph Convolutional Neural Network (CGCNN) model to predict band-gap values. We further trained this model using transfer learning on new HOIP crystals with different compositions obtained from a dataset of Density Function Theory calculations that the model had not seen before. Furthermore, we also trained models only on HOIP crystals; as well as another model that was trained on HOIP crystal graphs that included an extra edge-feature that distinguished between covalent and ionic interactions within a HOIP crystal. We compared the performance of these models with the transfer learning model to investigate whether models could be transferable to HOIPs, opening the door for similar studies on other hybrid materials.