| Abstract Scope |
Machine learning and artificial intelligence have demonstrated significant benefits to fields such as materials science. However, vision-based models for powder characterization are frequently constrained by the existence, quality, and quantity of labeled data sets. Creating training labels for microscopy images can be prohibitively expensive for tasks such as instance segmentation, and nearly impossible for tasks such as depth estimation, occluded region prediction, or 3D reconstruction. In this work, we explore the utilization of synthetic data for improving low-label or no-label powder characterization tasks. To generate useful synthetic data for these tasks, we implement a two-stage process that combines physics-guided simulations, capturing realistic powder-packing conditions, with image-domain translation, producing microscopy-style image artifacts. Finally, we evaluate how synthetic data can enable or improve vision model performance across multi-modal particle characterization tasks where real labels are limited, unavailable, or difficult to obtain. |