Data-driven modeling of a laser-plasma accelerator-based x-ray source

Laser-plasma accelerators (LPAs) enable compact, bright x-ray sources, but their practical application demands a significant reduction of beam instabilities that originate from drive laser fluctuations. Moreover, the complexity of the laser-plasma interaction makes it difficult to disentangle and quantify the impact of individual parameters on machine performance. To address this challenge, we develop a data-driven modeling strategy and apply it to an extensive dataset collected during a daylong operation of the LPA-based x-ray source LUX. By making use of an orthogonal-distance-based training objective, our approach reduces bias originating from measurement errors, which allows us to estimate the functional dependencies between laser, accelerated electrons, and generated undulator radiation. In addition, we demonstrate accurate prediction of the x-ray spectrum based on noninvasive measurements, showcasing the potential of our approach for building virtual diagnostics.