Universal physics-informed machine learning framework for the prediction of porosity defects in high-power laser beam welding with different metallic materials

Abstract

High-power laser beam welding (LBW) is a widely used joining technique for metallic materials. However, porosity defects frequently arise during the process, leading to significant degradation of weld properties. Accurate prediction of porosity remains challenging due to the highly nonlinear and material-dependent underlying physics. In this study, a universal physics-informed machine learning framework is proposed to predict porosity levels in the LBW of aluminum and steel. Systematic LBW experiments were conducted to quantify porosity ratios across a broad parametric space. In parallel, a well-validated multiphysics simulation model is employed to characterize molten pool behavior and keyhole dynamics under corresponding conditions. By selecting relevant physical variables and incorporating them into dimensionless features with explicit physical significance, describing keyhole stability and bubble capture, the proposed model demonstrates strong predictive performance of porosity ratio across different metallic systems. A general root mean square error of 1.35 is achieved. Furthermore, the potential universal mechanisms governing porosity formation for different metallic materials are identified and hierarchically evaluated. It is found that the Stokes number and the keyhole aspect ratio are the two most dominant physical factors governing porosity formation.