Core principles for machine learning at European XFEL

Abstract Streamlining and automating the operation of complex large-scale X-ray light sources like the European XFEL can minimize inefficiencies, thereby boosting their scientific outcome. To that end, advanced mathematical methods such as machine learning can be employed. In this paper, we define the core concepts related to the exploitation of such methods at European XFEL. We aim to empower scientists and operators through strict information quality control and easily explainable metrics. As methods based on machine learning may be error-prone, interpretable procedures and validation metrics are essential to gauge the advantage of using them and identify possible difficulties. To illustrate these principles, we present two selected applications. The first aims at preventing damage to X-ray imagers caused by the interaction of X-rays with a crystallizing liquid jet used as a sample delivery system. The second optimizes the in-plane calibration of the position of modules composing an imager.