This presentation was given at a workshop in Stuttgart on 10 February 2026 as part of the Lasers4MaaS project (EU Horizon Europe, grant agreement 101178719). It demonstrates how high-fidelity multi-physics simulations of laser beam welding, validated against experimental data, can be used to generate training datasets for data-driven reduced-order models (ROMs). Based on a dataset of 73 CFD simulations of 316L stainless steel covering laser power, welding velocity, and different beam shapes, the presentation shows how machine learning models (Neural Networks, Gaussian Process Regression, and Support Vector Regression) can solve both the forward problem of predicting the weld geometry (depth, width) from process parameters, and the inverse problem of identifying suitable welding parameters for a target weld geometry. Both a PDF version and the original PowerPoint file (including animations and videos) are provided.