Machine Learning for Ptychography

Abstract: Over the past couple of years, ptychography has become increasingly computationally intensive with larger detectors running at higher speeds. Today, it is routine to run Ptychography experiments which produce Gb/s data. This trend will continue, and traditional iterative reconstruction algorithms will struggle to keep pace with the speed of data generation even when GPU-accelerated and running on HPC resources. Recently, machine learning (ML) methods have proven to be successful – speeding up image reconstruction by orders of magnitude and allowing for advanced algorithms to improve image quality. Nevertheless, challenges and caveats exist with respect to the reliability of results predicted by ML methods.