Explainable machine learning accelerated density functional theory prediction for diffusive transport behavior of elements in aluminum matrix and graphene/aluminum interface

In this paper, the diffusive migration behavior of alloy atoms in aluminum matrix and different types of graphene/aluminum interfaces is systematically investigated by using a machine learning accelerated density functional theory. A small sample dataset is established by first principles calculation, the types of input and output eigenvalues are determined by feature engineering, and the number of input features for perfect interfaces, defective interfaces, and aluminium matrix are finally determined to be 6, 5, and 4 by taking into account the effects of model complexity and prediction accuracy. With a five-fold crossover and by comparing more than a dozen machine learning models, the CatBoost algorithm possesses the lowest error as well as a better coefficient of determination. We further optimized the CatBoost algorithm with further parameters and adjusted the regularization term coefficients to avoid the risk of overfitting. The impact of each feature on the model prediction results was quantitatively described by constructing a matrix of SHAP values. The best performing Catboost model was used to predict the full periodic table data, which in turn was used to screen out the elemental species that are easy to move towards the graphene/aluminum composite interface. Those alloying elements are beneficial for modifying the defective graphene in the composite by comparing the results of elemental diffusive migration in the aluminum matrix as well as at different graphene/aluminum interfaces. The results of machine learning accelerated first principles calculations can provide a theoretical basis for further development of new aluminum alloy composite.