Dual Generative Adversarial Networks for Neighborhood Learning and Prediction of Melt-Pool Dynamics in Additive Manufacturing

Abstract Laser powder bed fusion (LPBF) is one of the metal additive manufacturing technologies that uses a laser beam to selectively fuse the material into the desired shape layer by layer. At each fusing point, the laser heats and melts fine-grained powders to form a melt pool. The characteristics of melt pool play a critical role in determining the microstructure and mechanical characteristics in the final part. Melt-pool analysis is essential for applications such as process optimization and real-time defect prevention. However, LPBF involves time-varying process parameters and the need to incorporate neighborhood-based melting history to capture spatiotemporal evolution of melt pools. In this article, we propose a novel dual generative adversarial network (GAN) modeling framework to explicitly embed neighborhood-based melting history and predict the evolving melt-pool dynamics in accordance with process parameters. First, we design a novel encoder to characterize the spatiotemporal heterogeneity of the target melt pool and its neighbors. Then, we introduce a dual GAN architecture that simultaneously predicts the new melt pool and its spatiotemporal variations from the most recent neighbor. Experimental results show that the proposed framework effectively learns and predicts spatiotemporal dynamics from the melting history in comparison with traditional baseline models. This framework is generally extensible to other applications involving spatiotemporal modeling and prediction of neighborhood-structured data.