Predicting Melt-pool Morphology in powder-based Directed Energy Deposition using Deep Learning-based Generative Adversarial Networks

In powder-based Directed Energy Deposition (DED) additive manufacturing, melt pool stability is impactful for achieving quality depositions. Monitoring and controlling the melt pool in real-time are challenging due to its dynamic response to varying process parameters and the prevalence of abnormal patterns. A robust deep learning model could enable effective classification and identification of discrepancies in melt pool morphology. However, the limited availability of large, high-quality datasets for visual inspection poses substantial obstacles, making data acquisition time-intensive, costly, and complex. To resolve this issue, the study proposes Generative Adversarial Networks (GANs) to analyze high-speed DED process images, providing critical insights into melt pool behavior and generating synthetic images to augment existing datasets. By leveraging GANs for realistic data augmentation, this approach improves the dataset’s balance and supports subsequent classification models aimed at accurate melt pool morphology prediction. Enhanced by high-fidelity synthetic data, these models can forecast melt pool characteristics before printing, facilitating preemptive adjustments in process parameters for improved quality control. This research demonstrates the potential of GAN-based augmentation to enhance data-driven predictive modeling in DED, offering a pathway toward more adaptive and precise additive manufacturing systems.