Explainable formation energy prediction for uncovering the relationship between the electronic structure and stability of the heterogeneous catalyst

In recent studies, machine learning (ML) applications in the heterogeneous catalysis field for material properties prediction accelerate the catalyst discovery with desired properties. However, due to its high complexity, most ML models suffer from the black-box problem, which cannot provide a basis for prediction. Thus, reliable application and physical insight generation are challenging with conventional black-box models. Here, we developed an ML model that predicts formation energy (Ef) from the density of states (DOS). More importantly, by interpreting the model, we also confirmed the possibility of uncovering the relationship between the electronic structure of materials and their stability. Our model achieves successful performance demonstrating its superior capability of DOS featurization.