Laser-based directed energy deposition (DED-LB) with multiple powder feeders enables additive manufacturing of functionally graded materials (FGMs) through controlled variation of powder compositions during deposition. However, changes in powder composition affect spatial powder stream distributions, influencing focal morphology and position that can lead to laser-powder focus misalignment. Such a misalignment can induce process anomalies that result in part defects such as porosity or dimensional inaccuracies. While numerical methods (e. g. , computational fluid dynamics) have provided insights into powder flow physics, they remain computationally expensive and rely on idealized conditions without capturing machine-specific behaviors. Therefore, a data-driven approach is needed that is computationally efficient while capturing variations in process parameters and varying powder compositions. This work presents a deep learning approach using a convolutional neural network (CNN) for predicting powder stream distributions for a specific machine configuration under varying carrier gas flow rate, shield gas flow rate, total powder mass flow rate, and powder composition mixtures of 316L and 17-4PH stainless steels. The model is evaluated using experimental validation on the test dataset, demonstrating strong predictive performance with R^2 scores averaging 0. 939 and structural similarity of 0. 97 across meaningful signal regions. To assess the practical utility of the CNN predictions for downstream applications, the predicted powder streams are integrated into coupled thermal-geometric process simulations of DED-LB. These simulations yield part geometries with mean dimensional deviations below 1. 5% (max below 2. 5%) of those obtained using ground-truth powder stream measurements (PowderSpy). This close agreement between simulations using predicted versus experimental powder distributions validates the model’s suitability for process modeling of manufacturing FGMs with implications on in-situ process optimization and control systems given the computational efficiency.
Schürmann et al. (Mon,) studied this question.