Coherence-Based X-ray Microscopy: Automated Tuning and Uncertainty Quantification

X-ray microscopy at large-scale facilities enables studying a wide range of specimens, from single cells to small animals or from single crystals to bulk materials in working conditions. To realize this and to make these methods available to the growing user community of synchrotron radiation sources, our methods must also grow. The imaging methods employed to assess these samples are based on coherent lens-less microscopy. Therefore, the image of the sample is not formed during the measurement process itself but using a reconstruction algorithm. The algorithm solves a so-called ill-posed inverse problem whereby a complex wavefield is recovered from an intensity-only measurement.
The project aims to investigate methods for tuning the parameters of the reconstruction algorithm using machine learning (ML) techniques to eliminate the need for manual fine-tuning. Another area that requires further investigation is the accuracy of the reconstruction results. To provide more accurate estimates of the uncertainty associated with the reconstructed values, we intend to apply uncertainty quantification in combination with ML techniques, thereby generating error bars for our reconstructed values.