Machine Learning-Enabled Real Time Analysis in Ultrafast X-Ray Scatterings

The X-ray free-electron laser (XFEL) has opened up numerous unique scientific opportunities for ultrafast dynamics and particle imaging studies. However, obtaining meaningful experimental signals for complex samples remains a significant challenge. In this talk, we present two recent works to help address this issue by integrating machine learning techniques into data collection and analysis. Firstly, we combine neural network (NN) models with Bayesian experimental design algorithms for real-time experiment steering. The NN facilitates uncertainty quantifications in experimental settings and provides immediate parameter estimations, leading to physics-informed experimental decisions and more meaningful experimental measurements, as demonstrated through simulated X-ray Photon Fluctuation Spectroscopy (XPFS) measurements on magnetic excitations. In the second topic, we present an NN-based reconstruction algorithm for X-ray single particle imaging (SPI), capable of simultaneously estimating particle orientations and recovering the complete reciprocal space information. Our method remains robust under challenging experimental conditions, including strong shot-to-shot photon count fluctuations. It enables successful reconstructions from datasets with limited and highly-noisy diffraction patterns, surpassing limitations of the conventional algorithm.