Abstract: Residual stress induced by additive manufacturing is associated with deformation, cracking, and various structural defects in the workpieces, thereby limiting geometry control and performance stability in industrial applications. A residual stress prediction method based on direct energy deposition is proposed, employing a graph neural network framework. The approach begins with the computation of the 3D surface temperature field of the workpieces, captured using an infrared thermal imager, which is subsequently transformed into the global temperature field through finite element analysis. A mapping is then established between the transient temperature field at the end of the printing process and the residual stress field after cooling by means of a graph neural network. This enables the rapid prediction of the 3D global residual stress field following the cooling stage. Experimental results indicate that the 3D residual stress field of workpieces with varying geometries and boundary conditions can be predicted within approximately 2 seconds, about 7 200 times faster than traditional finite element simulations, while maintaining an average relative error of 13.72%, satisfying both real-time and accuracy requirements. Furthermore, comparative analyses reveal that the use of the temperature gradient field for prediction yields higher accuracy than the direct temperature field, achieving an overall improvement of 28.61%. This method offers a practical data-driven solution for the dynamic adjustment of process parameters in additive manufacturing.

Key words: additive manufacturing, residual stress, graph neural networks, 3D global residual stress field, temperature gradient field