Abstract The emergence of multi-laser additive manufacturing (AM) signifies a pivotal advancement in 3D printing, enabling concurrent operations that substantially shorten production times and increase efficiency. In multi-laser AM, however, managing the thermal history of printed components becomes critical, as temperature distributions are influenced by factors such as laser paths, laser powers, and scan speeds. Consequently, accurate prediction of temperature distributions is essential for minimizing thermally induced defects and ensuring high-quality products. Numerical simulations offer valuable insights into thermal distributions in AM but can be computationally intensive and lack transferability and ground-truth data integration. While traditional machine learning shows promise, its practical use is limited by scarce AM-specific data. Combining physics-based formulations with machine learning presents an exciting opportunity to develop efficient, accurate predictive models for multi-laser AM. This study proposes a path-aware Physics-Informed Neural Network (PINN) framework for predicting temperature distributions in multi-laser, multi-track AM scenarios. The residuals of the governing thermal equations and boundary conditions are embedded into the model as a physics-based loss function to be minimized. Additionally, complex scanning paths with various laser powers and scan speeds maps for each laser are directly incorporated into the loss function. Unlike conventional neural networks that require large datasets, PINNs act as solvers that can be trained without any data. They achieve this by minimizing the physical loss at randomly generated collocation points. In addition to independently solving the multi-laser, multi-track AM scenarios by performing random initialization of trainable network parameters, transfer learning is used to reduce the training time by initializing the network for multi-laser, multi-track AM configurations with the trainable parameters obtained from single-laser, single-track scenarios. Utilizing transfer learning eliminates the need to retrain models with random initialization for each new configuration, improving computational efficiency. By transferring the solution from single-track to multi-laser, multi-track scenarios, the required training epochs are reduced by 65% compared to the case without transfer learning. Moreover, by incorporating an additional data term into the loss function, the model can capture the temperature field more quickly, reducing the maximum temperature deviation between converged and ground-truth value by 96.9% compared to the case without the data term. Therefore, the proposed approach combines the rigor of physics-based modeling with the flexibility of real-world data integration, enhancing predictive accuracy and adaptability for practical multi-laser AM applications.
Faegh et al. (Mon,) studied this question.