Prediction and homogenization of optical tomography images and microstructure during powder bed fusion of metals using a laser beam by means of a style-based generative adversarial network

Additive manufacturing enables the production of complex geometries with a high material efficiency, making it a key technology in modern manufacturing. However, in powder bed fusion of metals using a laser beam (PBF-LB/M), an inhomogeneous thermal energy input can lead to residual stresses and microstructural irregularities, resulting in inconsistent mechanical properties. Addressing these issues in advance requires an efficient predictive modeling and localized process parameter adaptations. While sole numerical simulations offer insights into the thermal field, their computational cost limits a scalability, making methods based on artificial intelligence (AI) a promising alternative. In this study, an AI-driven approach to predict and homogenize the thermal behavior and, therefore, the microstructure in PBF-LB/M was designed. The approach is based on a style-based generative adversarial network (sbGAN) considering process statistics while remaining consistent with the underlying process physics. Optical tomography (OT) data along with finite element simulations were used to train, to validate, and to test the sbGAN model for Inconel 718 in terms of OT image predictions. The predicted images were then utilized to apply a laser power modification approach to reduce two types of overheating. These are characterized by a geometry-induced overheating, caused by part shapes with a reduced heat flux towards the build platform, and a vector-induced overheating, arising from short scan paths and low laser beam return times within individual layers. The effectiveness of this approach was assessed through the degree of OT signal uniformity and the microstructural homogeneity. The findings of this study proved the operational performance of the sbGAN model in predicting OT data both qualitatively and quantitatively. Overheated regions were reliably predicted, and the results agreed well with the experimental observations. For various test build jobs, the predicted mean values deviated by a maximum of 1.64% from the experimental values, while the standard deviation values differed by a maximum of 15.72%. The subsequent homogenization approach was demonstrated to be useful in reducing overheating. This approach contributed to a homogenization of the thermal signals and, by this, of the microstructure in PBF-LB/M. These findings advance AI-driven thermal modeling and process optimization, improving the final part quality and enhancing the reliability of PBF-LB/M. • Suitable inputs for an AI model to predict OT images during PBF-LB/M were identified. • The sbGAN, separating content and stochastics, outperformed conventional GAN models. • Laser power adaptations in overheated regions resulted in a homogenized heat input. • By the adaptations, a finer and more homogeneous microstructure was achieved.