3D temperature field prediction in direct energy deposition of metals using physics informed neural network

Predicting the temperature field during the direct energy deposition (DED) process is vital for the microstructure control and property tuning of fabricated metals. The widely used data-driven machine learning method for accurate temperature prediction, however, is impractical and computation-intensive due to its sole reliance on large datasets; also being a black-box model in nature, it lacks interpretability. We propose a physics informed neural network (PINN) model, which adopts a novel physics-data hybrid method by embedding the heat transfer law into the loss function of the neural network, to model the temperature field in both single-layer and multi-layer DED. The results show that the PINN-based models with additional extrapolation ability can accurately predict temperatures with a mean relative error of 4.83%, and achieve identical prediction accuracy with only 20% of the labeled data required for training the data-driven deep neural network. The proposed model is more explainable in terms of the physics of the DED process and is also applicable for the DED of different metals.