Visiting Student - XSD - Koda, Ryota - 10.21.26

This project proposes a real-time workflow for 3D X-ray ptychography that uses a sample-specific surrogate model to reduce reconstruction time. In 3D ptychography, a sample is illuminated from many angles, and a 2D image is recovered for each angle with iterative phase retrieval (e.g., ePIE) before the images are combined into a 3D volume. Because the phase retrieval is repeated for every direction, it dominates the total runtime. Earlier surrogate-model studies sped up this step but did not evaluate the whole workflow, and they typically required a training dataset prepared in advance. The proposed method instead builds the surrogate model on-the-fly. First, directional subsampling selects a subset of the projection angles. ePIE is run only on those angles, and the resulting diffraction pattern and object image pairs form a small training set. A surrogate model is trained on this set and predicts initial object estimates for the remaining angles. After preprocessing, these predictions initialize a second ePIE pass. The workflow will run on GPUs. Additional work will include the development of a foundation surrogate model for diverse samples and testing other surrogate architectures.