A dataset of synthetic additive manufacturing lack-of-fusion pores generated using a three-dimensional Generative Adversarial Network based on X-ray computed tomography images

Synthetic additive manufacturing (AM) lack-of-fusion (LOF) pores are generated using a three-dimensional Generative Adversarial Network (3D-GAN) framework. The LOF pores, introduced by varying laser-based powder bed fusion (L-PBF) process parameters, were extracted from segmented X-ray computed tomography (XCT) scans previously published (doi:10.18434/M32162). The extracted pores served as training data for the 3D-GAN model, which was designed to learn and generate the complex morphologies characteristic of real AM pores. To reduce the generation of fragmented components rather than singular unified structures, a component aware loss module is integrated into the adversarial learning mechanism. The dataset contains 3503 synthetic and training pores, each represented as binary 3D volumes in .mat format and as corresponding two-dimensional (2D) .tiff image stacks. The synthetic pores are developed to generate realistic simulated XCT data sets with AM pore morphology for evaluation and benchmarking of different image segmentation and defect detection algorithms under controlled ground truth conditions. The pores may be useful for structural integrity and other digital twin applications.