A Materials Map Integrating Experimental and Computational Data via Graph-Based Machine Learning for Enhanced Materials Discovery

Materials informatics (MI), emerging from the fusion of materials science and data science, has the potential to greatly accelerate material development and discovery. Although MI relies on data from both computational and experimental studies, their integration remains challenging. In our previous study, we addressed this challenge by training a machine learning model on experimental data and applying it to compositional entries in a computational database, thereby creating a unified dataset. In this study, we use these integrated datasets to construct material maps that visualize the relationships between material properties and structural features. The goal is to provide experimental researchers with a practical tool for exploring structurally similar compounds and thus their associated routes. We generate the materials map using the MatDeepLearn (MDL) framework, which represents crystal structures as graphs and employs deep learning for property prediction. Statistical analyses reveal that the MDL equipped with a message passing neural network (MPNN) architecture efficiently captures features related to the structural complexity of materials. Interestingly, this representational advantage does not always lead to higher accuracy in property prediction. We attribute this finding to the strong learning capacity of MPNN, which contributes primarily to the organization of data points within the materials map rather than to incremental gains in predictive precision.