Machine learning for bead geometry prediction in 5083-H112 aluminum laser welding considering a challenging dataset

Laser welding of 5083-H112 aluminum alloy presents several challenges due to the material’s high reflectivity and thermal conductivity. Consequently, it becomes difficult to establish a processing window that properly associates parameters with the desired weld bead geometry. This study aims to analyze and compare the predictive performance of five widely used machine learning regression techniques (Support Vector Regression, Decision Trees, Random Forest, K-Nearest Neighbors, and Artificial Neural Networks) for estimating weald bead geometry, considering a challenging dataset. It has small size, high process variance, and a discontinuity corresponding to the conduction-to-keyhole transition, thereby providing an important benchmark for assessing the models' capability of calibration on limited datasets. Weld bead width and penetration depth were selected as outputs, while scanning speed and power as inputs. A total of 99 bead-on-plate experiments were performed by combining 11 levels of laser power (1000 to 4600 W) and three levels of scanning speed (0.5, 1.0, and 4.0 m/min), within each parameter combination, process repeatability was accessed. The dataset was split into training (80%) and testing (20%) subsets. The best hyperparameters for each machine learning technique were determined by using GridSearch and five-fold cross-validation. The K-Nearest Neighbors technique resulted as the most recommended due to its low mean absolute percentage error, 4.68% for width and 8.35% for penetration depth, combined with the generation of smoother curves for physical plausibility. Feature importance analysis highlighted laser power as predominant factor. The proposed techniques demonstrate strong potential for industrial implementation, specially assisting technical personnel in defining process windows. • Comparison of five ML regression techniques for aluminum laser welding; • Challenging dataset with low repeatability and discontinuities; • KNN achieved low MAPE and produced smoother weld bead predictions; • Laser power identified as the dominant factor in weld bead geometry; • ML techniques show strong potential for industrial real-time implementation.