Intelligent prediction of crack stress intensity factors for nuclear-grade pressure vessels based on XFEM-PSONN collaboration

Abstract

The reactor pressure vessels (RPVs) are key components in nuclear power plants, and their structural integrity assessment is of great significance for the safe and stable operation of nuclear power plants. To address issues such as low computational efficiency and limited applicability of existing assessment methods, this study proposes an innovative collaborative prediction method based on the extended finite element method (XFEM) and the particle swarm optimization neural network (PSONN). This method enables rapid and accurate prediction of stress intensity factors (SIFs) under the combined influence of multiple parameters including crack geometric parameters, container structure dimensions and internal pressure. Firstly, a parametric model including typical crack configurations such as beltline shells and nozzle corners is established using XFEM, and a comprehensive database of SIFs is constructed. By systematically comparing the predictive performance of eight machine learning (ML) algorithms, a neural network model based on Particle Swarm Optimization is developed. And K-fold cross-validation and grid search techniques are adopted to optimize the model's hyperparameters. The interpretability analysis of SHAP indicates that internal pressure and crack inclination Angle are the most critical parameters affecting the prediction accuracy. By effectively integrating the physical accuracy of XFEM with the computational efficiency of PSONN, the proposed method provides a practical tool for rapid and accurate safety assessment upon crack detection in in-service inspections.