Conventional and Machine Learning-Driven Corrections for Beam Intensity Fluctuations in X-ray Fluorescence Imaging

Abstract X-ray fluorescence imaging (XFI) experiments are inherently vulnerable to beam-intensity instabilities, particularly during long duration measurements, where unanticipated instrumental failures can compromise data quality. Here we present a systematic comparison of conventional analytical and machine learning-based approaches for correcting time-dependent intensity fluctuations in XFI data. Using a representative high-resolution X-ray fluorescence imaging experiment, intensity drift was modeled using exponential decay functions, autoregressive integrated moving average analysis, support vector regression, and Gaussian process regression (GPR). While all approaches provided partial correction, GPR consistently yielded the most robust performance, particularly when applied to an experimentally derived surrogate intensity signal that is independent of sample composition. This probabilistic, nonparametric approach effectively captures nonlinear beam fluctuations without imposing predefined functional forms. The results demonstrate a generally applicable strategy for improving the quantitative reliability of photon-limited X-ray fluorescence imaging data affected by unforeseen beam instabilities.