Laser beam welding is widely used in manufacturing due to its ability to create high-quality joints. However, optimizing process parameters is challenging because of the complex and interconnected physical phenomena involved. As a result, developing the process usually requires extensive experimentation and expert knowledge, which can slow down scaling and delay wider industrial adoption. In this work, we demonstrate how machine learning can support welding engineers in designing welding processes more efficiently. 

This is achieved by using a validated high-fidelity numerical simulation to generate a synthetic dataset for aluminum spanning key process parameters, including laser power, welding velocity, and beam profile characteristics. The data, in a reduced format, is then used to train machine learning models that directly predict key output parameters, such as weld pool or keyhole geometry, from input parameters within seconds. To provide more than just prediction, SHAP (SHapley Additive exPlanations) analysis is applied to identify the most influential process parameters. This allows engineers to focus on the variables that most significantly impact the weld, thereby supporting process optimization.

Finally, we address the inverse problem by identifying parameter sets that produce the target weld geometry. Given the high dimensionality and complexity of the parameter space, we demonstrate how a dimensionality reduction technique can help visualize and efficiently navigate feasible process areas.