Developing a laser welding process for a new application typically involves many welding trials. Each trial requires material preparation, welding, sectioning and metallographic characterization of weld depth and weld width; spatter is evaluated from high-speed camera recordings. A classical Design of Experiments (DoE) campaign therefore easily grows to several hundred welds before a parameter window satisfying all geometric and defect requirements is established, and pinpointing a setting inside a narrow production tolerance remains difficult.

Bayesian optimization (BO) has been shown to reduce experimental effort in many process development tasks by selecting each new parameter set from what previous trials already revealed, concentrating effort near the most promising regions. However, standard BO starts with no physical knowledge and spends its first trials rediscovering what a welding expert already knows. We present a physics-informed BO in which established laser welding physics, such as the characteristic dependence of weld depth on power, welding velocity and focus diameter, is embedded into the method from the very first trial.

Evaluated against a validated laser welding reference across thirty independent repetitions, conventional DoE found a specification-compliant weld in roughly two thirds of cases after an average of 150 trials, while standard BO required 38. The proposed physics-informed BO identified a compliant weld in every repetition after about 18 trials. This corresponds to roughly an eightfold reduction in experimental effort compared to DoE, and more than twofold compared to standard BO.