Longitudinal Prediction of WAAM Bead Geometry from 3D Point-Cloud Linear Measurements Using Machine Learning

Wire arc additive manufacturing (WAAM) bead geometry is commonly represented using global descriptors, even though local bead dimensions can vary substantially along the deposition path. This study presents a spatially explicit framework for extracting and predicting longitudinal bead geometry from three-dimensional laser-scanned point clouds. Fifty single bead deposits were produced over a full factorial parameter range of 65–155 A welding current and 20–40 cm/min travel speed. A dedicated point cloud workflow was used to isolate individual beads, establish local substrate references, extract transverse profiles, and calculate bead width, robust height, and cross-sectional area at multiple longitudinal positions. Machine learning models were utilized to model welding current, travel speed, and normalized longitudinal geometric position as predictors. Linear ridge regression, quadratic ridge regression, and Gaussian process regression were evaluated using repeated nested grouped cross-validation, with complete beads retained as independent validation units. The selected framework achieved mean held-out bead R2 values of 0.981 for width, 0.965 for robust height, and 0.985 for cross-sectional area. Pipeline-level holdout evaluation showed that prediction at entirely unseen torch travel speed levels was more difficult than prediction of unseen beads within the sampled process domain, particularly for width and robust height at the domain boundaries. The results demonstrate that WAAM bead geometry can be represented and predicted as a longitudinally varying quantity rather than a single global value, providing a basis for more spatially resolved process planning, geometric simulation, and future adaptive control.