Lessons learned during on-site AI analysis of synchrotron diffraction data

Abstract

Materials scientists use time-resolved high energy synchrotron X-ray diffraction (HEXRD) experiments to study phase transformations in engineering materials. These experiments can nowadays generate large amounts of data owing the high acquisition rates possible. Usually, researchers analyse these data once they returned to their home labs. However, from previous studies, its expected that unsupervised machine learning can generate a first interpretation of synchrotron diffraction data, quickly, and possibly on-site directly at the beamline. This practice can, consequently, enable a more agile way of conducting experiments at the beamline. For our study, we wanted to find out how well we can apply on-site AI analysis for an agile beamtime and share our lessons learned. For this purpose, we prepare a generic auto-encoder neural network before starting our beamtime at DESY beamline P07. At the beginning of each experiment we fine-tune the model on diffraction data of the base material at room temperature. Then, we correlate the reconstruction error and single feature values of the trained auto-encoder to the phase transformation kinetics studied during heat treatments of different alloys. We will present the advantages of the on-site application of a fast AI analysis during our experiments but also point out its limitations and our lessons learned.