Can unsupervised machine learning boost the analysis of my synchrotron diffraction data (on-site)?

Kurzfassung

Time-resolved high energy synchrotron X-ray diffraction (HEXRD) experiments to study phase transformations can generate large amount of data owing the high acquisition rates which are possible nowadays. Moreover, the conventional data processing steps for revealing microstructure kinetics can be time consuming and represent a bottleneck in the research process. In our work, we explore the use of unsupervised machine learning to automatize the processing of HEXRD data. A machine learning algorithm is used to identify key features of analysis in HEXRD data sets. To this purpose, we trained an auto-encoder using five large data sets obtained during different in situ heat treatments of an additively manufactured Ti-6Al-4V alloy. The reconstruction error of the trained auto-encoder is correlated to the phase transformation kinetics. Furthermore, we show the evolution of the latent feature space of the auto-encoder during heat treatment and investigate its use for finding anomalies and characteristic features related to phase transformations in the diffraction data.