This contribution presents a hybrid reduced-order modeling framework for aluminum resistance spot welding that combines physics-informed neural networks with transformer-based state extraction. The approach enables physics-informed learning using exclusively experimental process data and addresses key challenges related to noisy measurements and unstable PINN optimization. A novel training strategy is introduced to stabilize convergence without requiring prior material parameter calibration. Multivariate pre-process sensor data are exploited to infer latent material and contact states, which parametrize a low-dimensional thermal–mechanical surrogate model of the welding process. The resulting framework enables computationally efficient near real-time prediction of weld nugget evolution and dynamic electrode displacement from early process stages onward. This work provides a basis for future closed-loop process control and predictive digital twins in advanced manufacturing.