Rapid analysis of 2D X-ray diffraction data from high repetition rate shock compression experiments using neural networks

Using high repetition rate (HRR) 10 Hz lasers for laser-driven dynamic compression experiments at X-ray free electron laser (XFEL) facilities poses a unique data analysis challenge. Reducing analysis time from minutes to fractions of seconds is critical to the outcome of an experiment performed at HRR. This analysis assesses shot success and the level of compression achieved. Rapid analysis can halt further use of valuable targets should problems develop, and simultaneously provide information regarding the sampled experimental parameter space, as well as the emergence of unexpected physics.

Neural network based machine-learning models are currently being developed to directly interpret the 2D X-ray diffraction (XRD) patterns produced during XFEL experiments, primarily from the VAREX detectors at EuXFEL. Model outputs can be produced in ~0.1 s with pressure estimates within ~0.1-10% error, dependent on material and degree of compression, as well as classification accuracies of ~99%. It is envisaged that such a system will guide an HRR experiment towards even more favourable regimes, aiding the exploration of any given parameter space. A method for generating synthetic 2D XRD patterns, along with a procedure for training models on these patterns and then integrating them into the EuXFEL's data systems is presented. Example data is used to demonstrate the model’s ability to classify different crystal structures and determine their lattice constants.