Toward prediction and insight of porosity formation in laser welding: A physics-informed deep learning framework

Abstract

The laser welding process is an important manufacturing technology for metallic materials. However, its application is often hindered by the occurrence of porosity defects. By far, an accurate prediction of the porosity defects and an insight into its formation mechanism are still challenging due to the highly nonlinear physics involved. In this paper, we propose a physics-informed deep learning (PIDL) framework by utilizing mechanistic modeling and experimental data to predict the porosity level during laser beam welding of aluminum alloys. With a proper selection of the physical variables (features) concerning the solidification, liquid metal flow, keyhole stability, and weld pool geometry, the PIDL model shows great superiority in predicting the porosity ratio, with a reduction of mean square error by 41 %, in comparison with the conventional DL model trained with welding parameters. Furthermore, the selected variables are fused into dimensionless features with explicit physical meanings to improve the interpretability and extendibility of the PIDL model. Based on a well-trained PIDL model, the hierarchical importance of the physical variables/procedures on the porosity formation is for the first time revealed with the help of the Shapley Additive Explanations analysis. The keyhole ratio is identified as the most influential factor in the porosity formation, followed by the downward flow-driven drag force, which offers a valuable guideline for process optimization and porosity minimization.