GPT-assisted learning of structure-property relationships by graph neural networks: Application to rare-earth doped phosphors

Abstract

Applications of machine learning techniques in materials science are often based on two key ingredients, a set of empirical descriptors and a database of a particular material property of interest. The advent of graph neural networks, such as the Crystal Graph Convolutional Neural Network (CGCNN), demonstrates the possibility of directly mapping the relationship between material structures and properties without employing empirical descriptors. Another exciting recent advancement is in large language models such as OpenAI's GPT-4, which demonstrates competency at reading comprehension tasks and holds great promise for accelerating the acquisition of databases on material properties. Here, we utilize the combination of GPT-4 and CGCNN to develop rare-earth doped phosphors for solid-state lighting. GPT-4 is applied to data-mine chemical formulas and emission wavelengths of 264 Eu(II)-doped phosphors from 274 papers. A CGCNN model is trained on the acquired dataset, achieving a test $R^2$ of 0.77. The model is then used to screen over 40,000 inorganic materials to make predictions on the emission wavelengths. We also demonstrate the possibility of leveraging transfer learning to fine-tune a bandgap-predicting CGCNN model towards the prediction of phosphor emission wavelengths. The workflow requires minimal human supervision, little domain knowledge about phosphors, and is generalizable to other material properties.