Material Mapping Using Experimental Data and Crystal Graph Neural Networks

Selecting optimal materials from a vast pool of candidates is a key challenge in materials development. Materials informatics, which integrates data science and materials science, facilitates more efficient material discovery. However, experimental data often lack crystal structure information, limiting the application of crystal graph-based methods. In this study, we combine experimental data with machine learning to propose a novel approach to material discovery. We extracted thermoelectric material data from StarryData2, an experimental database of materials properties, and developed a predictive model using Gradient Boosting Decision Trees (GBDT). This model was then applied to data from the Materials Project, creating a database that incorporates both structural information and predicted zT values, which represent the efficiency of thermoelectric materials. By applying crystal graph neural networks, we generated a material map that reflects the relationship between structure and zT predictions. In this case, the message passing neural network (MPNN) demonstrated high expressiveness, effectively capturing material features. This approach, which combines experimental data with machine learning, shows promise for guiding more efficient and intuitive material discovery.