Machine Learning for Accelerated Metal Oxide Thermochemical Predictions

A machine learning (ML) framework is developed to predict normalized clustering energies (NCE) of metal(II) oxides in order to extrapolate them to predict cohesive bulk energies. The NCE is the permonomer averaged energy difference between the cluster and the corresponding isolated monomers. A data set of DFT-optimized M(II)O containing four alkaline earth and predominantly 3d transition metals (M = Mg, Ca, Sr, Ba, Sc, Ti, V, Cr, Mn, Co, Cu, Zn, Cd) is encoded with Smooth Overlap of Atomic Positions (SOAP) descriptors and used to train kernel ridge regression (KRR), a feed-forward artificial neural network (ANN), and tree-based models (Random Forest, AdaBoost, XGBoost). On a held-out test set, KRR achieves the lowest error (MAE = 0.619 kcal/mol, R² = 0.994), followed by ANN (MAE = 1.250 kcal/mol, R² = 0.986), while tree ensembles perform less well. By fitting the NCEs, bulk cohesive energies for closed-shell oxides (MgO, CaO, SrO, BaO, ZnO, and CdO) are predicted to within ∼1.5 kcal/mol of the corresponding DFT values. Adding higher-oxidation-state structures (e.g., TiO₂ with Ti(IV)) improves ANN generalization within the Ti subset, suggesting that geometry and oxidation-state diversity improve model transferability. These results demonstrate that ML trained on modest cluster data can deliver accurate predictions of NCEs and bulk thermochemistry at a fraction of the cost of electronic-structure calculations. Furthermore, the model's ability to replicate DFT accuracy suggests that the complex electronic structure of metal oxides can be encoded into a suitable input vector for a simple ML regimen without a substantial increase in error.