Low cycle fatigue life prediction for neutron-irradiated and nonirradiated RAFM steels via machine learning

EUROFER97 and other Reduced Activation Ferritic-Martensitic (RAFM) steels are candidate structural materials for fusion reactors. Qualification of these steels requires the assessment of their performance under fatigue loading especially after exposure to neutron irradiation. However, the significantly high costs and engineering complexity of performing such tests make the experiments on irradiated material difficult. Therefore, alternative methods to predict the fatigue life of RAFM steels are deemed beneficial. In this study, machine learning was used to predict the fatigue life of irradiated and non-irradiated RAFM steels utilizing published experimental data. Four algorithms were benchmarked: Random Forest Regression (RF), Support Vector Regressor (SVR), Gradient Boosted Regressor (GBR), and MultiLayer Perceptron (MLP). Predictions were made based on two scenarios: one scenario with yield strength as input along with the fatigue test condition, and the second scenario without the yield strength. The results showed that RF and GBR were the best performing algorithms, with R² score between 0.9522 and 0.9696 on the training set and between 0.9058 and 0.9249 on the validation set. The Mean Absolute Percentage Error (MAPE) score was between 0.3 and 0.43 % on the training set and 0.7 and 1.2 % on the validation set. The quality of prediction of fatigue life without using the yield strength was shown to be almost equal to that in the first scenario. SHAP analysis revealed that strain range was the most influential input feature, while GBR showed greater sensitivity to secondary features such as temperature and diameter. GBR also captured more nuanced interactions between input variables, particularly in irradiated datasets, confirming its superior ability to model complex fatigue behaviour. The ability of these algorithms to predict the fatigue life corresponding to test conditions not used for training was checked using 15 % of the collected database as a prediction set, which included fatigue life of irradiated materials. Results indicated that these algorithms were able to predict the fatigue life of this set with 83-87 % of the points lying within the factor of two from the experimental value.