Methodologies developed for dataset preparation and the interpretability of machine learning algorithms used for the prediction of crack growth rate

The rapid development of high-performance computing has made data-driven methods increasingly useful in material science. Predicting fatigue crack growth in additively manufactured alloys is particularly challenging due to the combined effects of process parameters, microstructure, and loading conditions. Traditional analytic models, such as the Paris law, cannot fully capture these interactions, and previous machine learning studies have not explored a multi-material dataset with a robust interpretability framework. This study introduces a methodology for dataset preparation, hyperparameter tuning, model interpretability, and machine learning-based prediction of crack growth rate using experimental data of different alloys. Several algorithms were tested for both pointwise prediction and complete sigmoidal crack growth curves. Model interpretability was enhanced through Shapley value analysis, which highlighted key features and their co-dependency, linking them to underlying material mechanisms. The proposed framework advances predictive accuracy and interpretability, offering practicality for diagnostic applications and structural design of additively manufactured components.