Prediction of 3D temperature field through single 2D temperature data based on transfer learning-based PINN model in laser-based directed energy deposition

Abstract

Temperature field is critical in laser additive manufacturing and significantly influences component deformation, microstructure, and mechanical properties. Accurate monitoring and control of temperature evolution are essential for achieving optimal fabrication quality. However, current monitoring technologies struggle to capture the 3D temperature field comprehensively, necessitating reliance on numerical simulations that are cost-prohibitive and time-inefficient. In this regard, we integrate physics-informed neural networks (PINNs) with transfer learning, leveraging these promising techniques to solve computational physics problems and data scarcity issues, respectively. We proposed a transfer learning-based PINN framework for efficiently and accurately predicting the 3D temperature field during the blue laser deposition of aluminum‑magnesium alloy powder. The PINN architecture incorporates a lightweight attention block, ResNet block, and fully connected layers, with a customized loss function that includes residual terms of physical rules, particularly the often-overlooked thermal convection process. The PINN is pre-trained using numerical simulation data on additive processes and fine-tuned using a 5-dimensional tensor dataset derived from infrared images of the single-track single layer blue laser deposition experiment. The blue laser deposition experiment demonstrated the effectiveness and superiority of the proposed method. Results indicated that the average temperature prediction error is <1.3 %, and the training time on the target task is reduced to one-third of the pre-training time. Our model also showed higher prediction accuracy than other PINN derivatives. This framework is highly adaptable and can be extended to other metal AM processes under the proposed architecture, enhancing the real-time temperature field prediction for metal additive manufacturing.