Diffraction image selector

About the Project

The quality of images used during serial crystallography data processing often weighs heavily on the accuracy of results. It is routine for experimenters to visually inspect and tune parameters to select the best images to process. By using state-of-the-art machine learning technologies, coupled with realistic forward modeling of crystallographic data, one can build a regression model which accepts images as inputs, and outputs experimentally interesting properties. This paves the way towards a machine intelligence capable of auto-selecting the highest quality images during diffraction collection.

For this project, students will use resmos2 data comprising about 10000 simulated diffraction images with corresponding labels, as well as additional simulated datasets, to build a regression model that will work in real-time to label images during data collection at SSRL beamlines. This will greatly enhance the user experience at the SSRL/SMB beamlines.