A DATA-DRIVEN MACHINE LEARNING APPROACH FOR NUCLEAR POWER PLANT APPLICATIONS

Abstract

Stress corrosion cracking (SCC) is a major degradation mechanism in safety-critical components because it results from the combined action of a susceptible material, tensile stress, and a corrosive environment. This study develops a data-driven model for predicting SCC crack growth rate in stainless steel 304 using six input variables: cold work, Vickers hardness, carbon content, temperature, stress intensity factor, and yield strength. The original project compiled 90 experimental data records from published research and technical sources. Because the dataset was small for machine-learning applications, regression-based data extension with bounded random perturbations was used to generate a working dataset of 1090 records while retaining the observed trends. XGBoost was initially evaluated but produced unstable predictions on the available data. Light Gradient Boosting Machine (LightGBM) was subsequently selected because it provided lower prediction errors, faster training, and more consistent behavior. The final model used regression learning with an 80:20 data split, feature standardization, a learning rate of 0.05, 31 leaves per tree, and early stopping based on validation RMSE. Model analysis identified temperature, Vickers hardness, and yield strength as the three most influential variables, with normalized importance values of 19.47%, 17.70%, and 16.90%, respectively. The predicted crack growth rate increased with stress intensity factor and showed a particularly strong temperature dependence. The reported prediction error decreased from initial values of -753.75% and 66.87% to below ±1% during subsequent model training, with a minimum reported error of -0.04%. Although the results demonstrate the potential of machine learning for SCC assessment, the limited experimental database and use of synthetic data mean that further validation using independent experimental and plant-relevant data is required.