Researchers have developed a transfer learning-enhanced physics-informed neural network (TLE-PINN) for predicting melt pool morphology in selective laser melting (SLM). This novel approach combines physics-informed constraints with deep learning techniques, achieving superior accuracy, faster training times, and reduced computational demands. Published in Advanced Manufacturing, this breakthrough has significant potential to improve the efficiency of SLM processes, enable intelligent real-time process control, and enhance manufacturing quality.
Selective Laser Melting (SLM) has emerged as a transformative technology in additive manufacturing, enabling the production of high-precision metal components for industries such as aerospace, automotive, and healthcare. However, accurately predicting melt pool morphology—a critical factor influencing material properties and process quality—remains a significant challenge. Traditional numerical simulations are computationally intensive and time-consuming, while purely data-driven models often lack physical consistency to capture the complex multi-physics nature of SLM.
To address this issue, researchers from Wuhan University have developed a transfer learning-enhanced physics informed neural network (TLE-PINN) method that combines enhanced EPINN with deep learning models through a transfer learning framework. This novel approach drastically reduces computational costs while achieving high prediction accuracy.
"This method represents a significant advancement in additive manufacturing," explains Professor Yaowu Hu. "By integrating physics-informed modeling with transfer learning, TLE-PINN bridges the gap between traditional numerical simulations and artificial intelligence, offering precise and efficient solutions for predicting melt pool morphology."
The EPINN component of the TLE-PINN framework enforces strong physical constraints during training by incorporating heat transfer equations and boundary conditions directly into the neural network's loss function. This ensures that the model accurately represents the melt pool morphology, even in complex scenarios. Meanwhile, the transfer learning framework fine-tunes the model using high-fidelity data, updating only the final layers while freezing earlier network parameters. This process significantly enhances training efficiency and computational scalability.
A critical challenge in SLM is balancing computational efficiency with prediction accuracy. Traditional models require significant computational resources, making them impractical for real-time applications. TLE-PINN overcomes this limitation by leveraging physics laws and transfer learning to streamline training and inference processes, achieving superior performance while minimizing resource requirements. "This design offers a unique combination of speed and accuracy, making it particularly suitable for industrial applications," adds Professor Hu.
To validate the framework, the research team conducted extensive simulations and experiments. Using 42CrMo steel samples, the model predicted melt pool morphology across a range of laser scanning speeds (1–9 mm/s). Experimental results confirmed the TLE-PINN framework's superior accuracy, with predictions closely matching high-fidelity simulation data and experimental measurements. Compared to traditional PINN and data-driven methods such as Random Forest and XGBoost, TLE-PINN demonstrated significantly lower temperature deviations and more consistent results, highlighting its robustness and reliability.
The framework's computational efficiency is another key advantage. While traditional models require extensive training time and computational power, TLE-PINN achieves faster convergence with reduced computational demand, making it a cost-effective solution for large-scale manufacturing applications. Its scalability and adaptability also ensure compatibility with a wide range of SLM parameters and material types, further broadening its potential applications.
Looking ahead, this method holds great promise for broad application in SLM online process control and manufacturing optimization, providing intelligent solutions for real-time adjustments and enhanced efficiency. The researchers are also exploring ways to expand the framework's capabilities to handle more complex material systems and larger parameter ranges, which could enable even greater adaptability in industrial scenarios.
While further refinements are needed to fully capture the complexities of melt pool behavior under diverse manufacturing conditions, this study represents a critical step toward integrating artificial intelligence with physics-based modeling for smarter and more efficient manufacturing.
The paper, titled "Transfer Learning-Enhanced Physics-Informed Neural Network for Accurate Melt Pool Prediction in Laser Melting," was published in Advanced Manufacturing.
Zhu Q, Lu Z, Hu Y. Transfer learning-enhanced physics informed neural network for accurate melt pool prediction in laser melting. Adv. Manuf. 2025(1):0001, https://doi.org/10.55092/am20250001.
