Prediction of post-irradiation swelling rate of 316L stainless steel based on Variational Autoencoders and interpretable machine learning

Abstract

In the field of materials science, accurately predicting the swelling rate of materials in irradiated environments is crucial for ensuring safety and reliability. This study aims to enhance the predictive accuracy of the swelling rate of irradiated 316L stainless steel, particularly in high-tech applications such as nuclear energy. By comparing various machine learning models, it was found that the Extreme Trees Regression (ETR) model performed best on the test set, achieving an R² of 0.79 and a Root Mean Square Error (RMSE) of 1.65 %. Although it demonstrated strong generalization capabilities, the limited data volume restricted its predictive accuracy. To address this issue, the study employed Variational Autoencoders (VAEs) for data augmentation, generating an additional 400 synthetic data points to expand the original dataset. This enhancement increased the R² on the test set to 0.91 and reduced the RMSE to 1.11 %. Following data augmentation, feature selection was conducted, resulting in Si, C, IrF, T, and Dd being identified as the optimal feature combination. SHapley Additive exPlanations (SHAP) was then utilized for interpretability analysis, revealing the significant effects of these features on the swelling rate. The findings provide essential insights for understanding and optimizing the swelling behavior of materials following irradiation.