Laser welding of sheet metals induces steep, highly localized thermal gradients that drive residual stress formation and out-of-plane distortion. While thermo-mechanical finite element (FE) welding simulations can predict these effects, their industrial deployment for tolerance-driven distortion compensation scenarios, like in automotive assembly, is limited by demanding model setup and high computational cost. This contribution targets these bottlenecks through a workflow that couples validated thermo-mechanical simulation with physics-guided surrogate modeling for rapid distortion prediction in sheet-metal laser welding. An experimental-numerical data generation pipeline is established to ensure repeatable training and validation data. Weld coupons are digitized pre-/post-process, and distortion fields are extracted from 3D scans. A transient thermo-mechanical FE model is then constructed and calibrated against measured distortions, providing a numerically consistent, high-resolution reference for subsequent sampling of the process parameter space. Particular emphasis is placed on practical numerical aspects that govern surrogate-ready datasets: stable representation of the moving heat source, mesh/time-step sensitivity, thermo-elasto-plastic coupling choices, and robustness of automated boundary-condition handling across design points. On this basis, two surrogate strategies will be investigated: First, a full-field replacement model inspired by the work of Yi et al. [1] learns the mapping from process/design descriptors to spatial distortion fields using a convolutional encoder-decoder (U-Net) trained in an adversarial setting to preserve sharp gradients and global modes. Second, a hybrid surrogate-solver approach retains the mechanical FE solve while replacing costly thermo-mechanical history by predicting equivalent shrinkage/inherent-strain descriptors via Gaussian processes or neural networks, creating a machine learning assisted inherent-strain model, similar to the approach of Zhang et al. [2]. These surrogate models should allow fast distortion evaluation with physics-consistent constraints. The resulting models are assessed with respect to accuracy, generalization across the parameter domain, and speed-up relative to the baseline FE analysis, aiming at a practical route toward data-efficient, simulation-informed distortion compensation for automotive sheet-metal assemblies.