Prediction of formation energy for oxides in ODS steels by machine learning

[Display omitted] •Magpie descriptors and physical features were introduced to the machine learning models.•An effective light gradient boosting machine model was established to predict the formation energy of oxides.•The phase and proportions of various nano-oxides in six oxide dispersion strengthened steels were determined.•The formation mechanisms of different kinds of oxides were analyzed based on the results of machine learning models. The phase and proportion of nano-oxides in oxide dispersion strengthened (ODS) steels are determined by the formation energy and the content of oxide-forming elements, particularly minor reactive elements, which in turn influence the macroscopic properties of the ODS steels. To study the phase, morphology, and interfaces of oxides in six ODS steels, transmission electron microscopy (TEM), scanning transmission electron microscopy (STEM), and high-resolution transmission electron microscopy (HRTEM) techniques were employed. Subsequently, machine learning (ML) methods were used to predict the formation energies of the oxides in the ODS steels. Among six ML models employed, the light gradient boosting machine (LGBM) model shows the highest accuracy in predicting the formation energy of the oxides, with root mean square error (RMSE) values of 0.18 eV/atom and 0.27 eV/atom, mean absolute error (MAE) values of 0.11 eV/atom and 0.18 eV/atom, and coefficient of determination (R2) values of 0.96 and 0.92 for the training and testing sets, respectively. For Magpie descriptors, the most essential feature is “dev_Electronegativity” with an importance score of 444. The method developed in this study can establish an accurate prediction model for the formation energy of oxides. The knowledge obtained in this work guides the future development of high-performance ODS alloys.