(364k) Graph Invariants As Rapid Features for Classification and Regression of Bulk Hydrogen Intercalation Energies in Reducible Metal Oxides

Intercalating hydrogen into reducible metal oxides drastically alters the physical and chemical properties of the host oxides. These changes can be directly quantified with quantum chemical calculations like density functional theory (DFT). However, due to the existence of many intercalation sites, this approach becomes tedious, requiring significant computational resources. Machine learning (ML) methods have been applied to these systems to accelerate property predictions, such as intercalation energies, band gaps, and crystal structures. Additionally, ML methods can help interrogate the underlying physics that govern the relationships between intercalation and material properties. The features for these models can be tabulated elemental properties (e.g., electronegativity), making the application of the models general and rapid. Alternatively, DFT-derived quantities, such as electronic band centers and charges, have also been used as features with the additional computational cost of a relatively small number of DFT calculations. Although the use of tabulated features can save computational cost, the derived models often require many features and potentially complex ML architectures for increased accuracy. This may obscure the physical interpretation that could be extracted from simple models that use few descriptors.

In this talk, we will discuss our recent application of graph fingerprints, commonly used in convolutional neural networks (CNNs) for image recognition and molecular ML applications, to predict DFT-calculated hydrogen intercalation energies in bulk metal oxides. The graphs represented the lattice structures of the metal oxides, considering the number of bonds and neighbors at each atom center. Specifically, we considered the graphs as Laplacian matrices of the bulk metal oxides and their intercalated counterparts. Instead of using the full matrices, as is typical for CNNs, we collapsed the full matrices into various single-valued graph invariants. These invariants were then used as features for our ML models that comprise a direct physical link to the geometries of the metal oxides. We demonstrate that a simple logistic regression model can be used to classify metal oxides as either intercalating or non-intercalating oxides. We also show that this classification model can be used to focus on a region of training and improve the performance of regression models that predict the intercalation energy of hydrogen in metal oxides. In summary, this work illustrates a quantitative link between the graph invariants representing metal oxide lattice geometries and changes in material properties associated with bulk hydrogen intercalation. Overall, our ML approach can be applied to rapidly screen energy storage materials and assess the stability and performance of electrocatalysts.