Micro‐Mechanism Informed Neural Networks for Process‐Property Prediction in Laser Powder Bed Fusion

ABSTRACT Data‐driven machine learning for additive manufacturing suffers from data scarcity and poor extrapolation beyond training distributions. Physics‐informed neural networks have addressed thermal physics but neglect the metallurgical micro‐mechanisms governing material properties. This work presents micro‐mechanism informed neural networks (MMIANN), embedding six metallurgical mechanisms as differentiable constraints: Eagar–Tsai thermal physics, Koistinen–Marburger martensite kinetics, Hall–Petch strengthening, Ashby–Orowan particle bypass, JMAK precipitation and SiC decomposition. The framework is demonstrated on SiC‐reinforced SKD61 tool steel—a material system with no published LPBF studies. Four approaches are compared: random forest, baseline neural network, MMIANN with soft constraints and physics‐guided models with hard physics embedding. All achieve on test data. However, extrapolation performance differs dramatically: Random forest degrades to , whereas soft‐constrained models (baseline NN and MMIANN) achieve . The physics‐guided model maintains on extrapolation data. These results demonstrate that physics integration architecture determines extrapolation capability more than constraint quantity—hard embedding substantially outperforms soft constraints for generalisation beyond training conditions. This finding carries implications for physics‐informed machine learning design in engineering applications.