The paper presents a datadriven framework for parameter identification of welding models. Common identification procedures are based on iterative optimization algorithms which minimize the distance between experimental measures and simulations. The cost of repetitive evaluations of objective functions is prohibitive, especially in welding cases, due to the multiphysics, nonlinear and multiparametric aspects. This is why one proposes to use a novel datadriven approach to improve the efficiency of the inverse identification procedures. Based on a sparse sampling strategy, an a posteriori non-intrusive reduction method, i.e. HOPGD, is used in the offline training stage for constructing the computational vademecum. An online subspace learning method coupled with a global optimization algorithm is proposed for the online search. The efficiency of the proposed method for multiple parameters identification is demonstrated through examples based on a 3D welding model.