| Abstract Scope |
The increasing adoption of sensing and data-driven methods in welding and additive manufacturing has enabled advances in process monitoring and quality assurance. However, deploying machine learning (ML) in production remains challenging due to limited labeled data, domain shifts between lab and plant settings, and variability across machines and applications. This presentation introduces a unified perspective on developing credible, scalable, and generalizable ML for welding processes.
The first part highlights progress in applicable ML, including virtual sensing, multi-modal sensing, and robust pipelines for defect detection and quality prediction. Emphasis is placed on self-supervised learning, enabling effective use of unlabeled plant data and improving robustness. The second part discusses a vision for generalizable AI across machines and applications within the same process family. Physics-informed surrogate modeling and representation learning are outlined to enable transfer across geometries, materials, and equipment with minimal recalibration. |