%0 Thesis
%A Sohail, Zain Muhammad
%T Denoising Methods for Multi-Dimensional Photoemission Spectroscopy
%I RWTH Aachen
%V Masterarbeit
%C Aachen
%M PUBDB-2025-00428
%P 92
%D 2024
%Z ProposalID: 20010115 and 20010001 additionally
%Z Masterarbeit, RWTH Aachen, 2025
%X The probabilistic nature of photoemission, combined with the exploration of large multidimensional parameter spaces–including momentum, energy, time and spin polarization–necessitates time-intensive data acquisition to ensure statistical robustness. These measurements are especially important for capturing ultrafast phenomena, where pulsed light sources, such as free-electron lasers (FELs), become indispensable due to their ability to deliver high-brightness, ultrashort X-ray pulses. However, the low repetition rates of current FEL sources significantly extend acquisition times, impeding the real-time decision-making that could otherwise enhance experimental results. Hence, to optimize experimental outcomes for the valuable beamtimes, techniques that can harness the structures and correlations within the multidimensional space are necessary to accelerate data acquisition without compromising data fidelity.To address these challenges, we present an investigation into advanced denoising methodologies for multidimensional photoemission spectroscopy (MPES) data acquired with time-of-flight momentum microscopes. We focus on two key approaches: (1) employing BM3D with variance stabilization through the Anscombe transform in moderately noisy datasets and (2) leveraging a deep learning-based 3D UNET architecture, based on the Noise2Noise paradigm, excelling in low-count regimes where classical methods fail.We further establish that the photoemitted electron distributions measured with SASE FELs deviate from traditional Poissonian statistics, instead following negative binomial statistics, an outcome that has implications for denoising strategies in the MPES data.Our results demonstrate that BM3D delivers robust denoising performance for datasets with moderate average-counts (on the order of 1 ×10^−2 counts per voxel). However, in extreme low-count regimes (on the order of 1 ×10^−3 counts per voxel), where most conventional denoising techniques fail, the deep learning-based approach achieves exceptional denoising performance. Remarkably, we show that MPES datasets acquired in just 10 minutes using an FEL light source can, when processed with our deep learning model, reveal key features that remain indistinguishable even after hours of conventional measurement. The findings presented have therefore the potential to streamline data acquisition at both laboratory-scale table-top setups and large-scale facilities such as FEL FLASH. By optimizing acquisition parameters, researchers can conserve valuable beamtime or extend the scope oftheir studies to broader parameter spaces, results that hold broader implications for related experimental techniques.
%F PUB:(DE-HGF)19
%9 Master Thesis
%U https://bib-pubdb1.desy.de/record/622633