Machine Learning (ML) is increasingly used in advanced manufacturing, yet its application to predictive weld quality assurance remains limited, particularly for high-precision techniques such as electron beam welding (EBW). This study presents the first validated ML model for predicting EBW success rate prior to execution, using real manufacturing data from the International Thermonuclear Experimental Reactor (ITER) Vacuum Vessel, a nuclear-grade component with strict quality requirements. A supervised learning model was developed using 1,802 EB welds, with input features including weld geometry, process parameters, and component configuration. Multiple algorithms were evaluated—Random Forest, XGBoost, K-Nearest Neighbors, and Artificial Neural Networks—with Random Forest achieving the best balance of accuracy and interpretability. The model predicted weld conformity with 83% accuracy and defect root cause and length with over 90% accuracy. Feature importance analysis identified weld length and VV sector sequence as dominant predictors, consistent with expert knowledge and first-of-a-kind learning curves. The model was adapted to output probability scores, enabling conservative decision-making in a nuclear context. Predictions for remaining welds showed accurate classification of true positives and true negatives, demonstrating strong generalization within the dataset. This approach allows proactive quality assurance, reducing deviation from the planned path and minimizing impacts on resources, documentation, and schedule. This work fills a critical gap in predictive quality assurance for EBW and supports Industry 4.0 goals by embedding ML-driven decision support into manufacturing workflows. The methodology is transferable to other fusion and high-value manufacturing programs, with future work focused on generalization and integration of explainable AI tools.
Zuniga et al. (Sun,) studied this question.