Accurate thermal prediction is a critical step toward achieving high-quality metal additive manufacturing (AM) components, where temperature evolution is tightly coupled with the complex laser scan path, process parameters, and laser profiles. Achieving accurate and near real-time thermal predictions is essential for process map optimization prior to printing, enabling rapid evaluations within optimization loops without the prohibitive cost of simulations. However, such fast predictions have been limited by conventional modeling approaches, which are either based on time-consuming numerical simulations or require large volumes of data to train machine learning models. In this work, a Physics-Informed Neural Network (PINN) framework is introduced, through which near real-time, data-free thermal prediction is enabled. Power-velocity-position maps for a given scan layer within the Gcode along with laser profiles are directly embedded into the neural network, and the underlying thermal physics is enforced without the use of external training data. The method is verified against numerical simulations, with a maximum relative error of only 3.36 % at peak temperatures.
By leveraging the transfer learning capability, the model achieves a 60 % reduction in training time, allowing adaptation across various path planning strategies, process maps, and beam profiles. Furthermore, immediate thermal field estimations of a given path across various process maps are enabled by a quick-shot prediction approach, offering a practical solution for near real-time predictions in AM process design optimization workflows. Finally, the study provides key insights into training PINNs and optimizing architecture, establishing a foundation for more accurate, real-time thermal predictions in metal AM.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号