| Abstract Scope |
Reliable qualification of metal additive manufacturing parts requires accurate prediction of thermal history, residual stress, and distortion. Physics-based finite element simulation captures this accurately but takes hours to days per build, making it incompatible with real-time control. This work develops a physics-aware two-stage graph neural network surrogate that converts finite element meshes into graphs with physics-informed features including laser proximity, boundary distance, and element birth timing. A DeeperGCN thermal model predicts full-field temperature across time steps and generalizes to unseen geometries without retraining. A recurrent graph neural network maps the predicted thermal history to full-field residual stress and displacement, preserving causal thermomechanical structure. These models enable rapid pre-build risk assessment orders of magnitude faster than finite element analysis. Looking ahead, the surrogate is designed to assimilate in-situ sensor data layer by layer, enabling subsurface inference, anomaly detection, and closed-loop process control as a physics-resolving digital twin for metal additive manufacturing. |