Diffusive migration behavior of single atoms in aluminum alloy substrates: Explaining machine-learning-accelerated first principles calculations

In this paper, we investigated the diffusion migration behavior of single atoms in an aluminum matrix using a machine-learning (ML)-accelerated first-principles calculation method. Initially, we used density functional theory to investigate the diffusion migration behavior of 30 individual atoms within the aluminum matrix. The interaction energy between alloy atoms and vacancy along with the diffusion potential of alloy atoms in the aluminum matrix are utilized as the output parameters. The intrinsic parameters of the alloy atom such as atomic radius, ionic radius, and first ionization energy are employed as input eigenvalues to construct a mathematical model for ML. The relationship between descriptors and prediction targets is initially determined by correlation analysis, and the number of input features and descriptors for different targets is determined using recursive feature elimination. Subsequently, the sophistication of the chosen model is demonstrated via cross-validation and fine-tuned to optimize its performance. Advanced ML models strive to achieve a balance between generalization ability and accuracy in limited data. To validate its efficiency and accuracy, the CatBoost model has undergone rigorous testing by traditional algorithms. Interpretable ML was also performed in order to understand the prediction process of the model. The CatBoost model was subjected to Shapley additive explanations interpretation analysis and the trained model was used to predict the diffusion migration behavior of other single atoms in the periodic table in the aluminum matrix. The results of ML-accelerated first principles calculations can be interpreted to provide a theoretical basis for further development of novel aluminum alloys. (Figure presented.)