Data-driven U-net machine learning (ML) models, including the pix2pix conditional generative adversarial network (cGAN), are shown to predict 3D printed voxel geometry in digital light processing (DLP) additive manufacturing. A confocal microscopy-based workflow allows for the high-throughput acquisition of data on thousands of voxel interactions arising from randomly gray-scaled digital photomasks. Validation between prints and predictions shows accurate predictions with sub-pixel scale resolution. The trained cGAN performs virtual DLP experiments such as feature size-dependent cure depth, anti-aliasing, and sub-pixel geometry control. The pix2pix model is also applicable to larger masks than it is trained on. To this end, the model can qualitatively inform layer-scale and voxel-scale print failures in real 3D-printed parts. Overall, machine learning models and the data-driven methodology, exemplified by U-nets and cGANs, show considerable promise for predicting and correcting photomasks to achieve increased precision in DLP additive manufacturing.
Keywords: 3D printing; digital light processing; machine learning; neural networks; pix2pix; stereolithography; vat photopolymerization.
Published 2023. This article is a U.S. Government work and is in the public domain in the USA.