Bead geometry prediction in wire arc directed energy deposition using physics-informed machine learning and low-fidelity data
Asif Rashid, Farzad Vatandoust, Akshar Kota, Shreyes N. Melkote
Abstract
Wire Arc Directed Energy Deposition (Wire Arc DED) is a promising metal additive manufacturing technique, yet accurate bead geometry prediction remains a challenge due to the complex thermal and geometric interactions in the process. In this study, we present a coupled Physics-Informed Neural Network (PINN) framework to predict the bead geometry by integrating the governing process physics and experimental data, thereby addressing the limitations of both computationally expensive numerical models and purely data-driven approaches. The model employs a sequential two-step workflow, where a thermal model first predicts temperature evolution, which subsequently informs a geometry model for predicting the bead geometry. Results indicate that a high-fidelity PINN model with high spatiotemporal resolution captures the intricately coupled thermal and geometric variations inherent to bead deposition with good predictive accuracy albeit at a higher computational cost, while a low-fidelity PINN model with lower spatiotemporal resolution offers a computationally efficient alternative with marginally higher errors. The incorporation of measured bead geometry data significantly enhances prediction accuracy, with a minimal amount of low-fidelity data sufficing to refine predictions effectively. Moreover, the model generalizes well across different bead locations along the deposition length, demonstrating reliable performance. The high-fidelity PINN model, using a temporal step size of 0.2 s, achieves an average height prediction error of 8.38% and width error of 1.09% after approximately 12.7 hours of training on four H100 GPUs. In contrast, the low-fidelity model, with a coarser temporal step size of 0.5 s, reaches nearly the same accuracy (8.33% height error, 1.56% width error) with just 2.7 hours of training on a single H100 GPU. This corresponds to a 79% reduction in training time and substantially lower hardware requirements, highlighting the scalability and efficiency of the proposed hybrid modeling approach.