Geochemical variations of sedimentary records contain vital information for understanding paleoenvironment and paleoclimate. However, to obtain quantitative data in the laboratory is laborious, which ultimately restricts the temporal and spatial resolution. Quantification based on fast-acquisition and high-resolution provides a potential solution but is restricted to qualitative X-ray fluorescence (XRF) core scanning data. Here, we apply machine learning (ML) to advance the quantification progress and target calcium carbonate (CaCO3) and total organic carbon (TOC) for quantification to test the potential of such an XRF-ML approach. Raw XRF spectra are used as input data instead of software-based extraction of elemental intensities to avoid bias and increase information. Our dataset comprises Pacific and Southern Ocean marine sediment cores from high- to mid-latitudes to extend the applicability of quantification models from a site-specific to a multi-regional scale. ML-built models are carefully evaluated with a training set, a test set and a case study. The acquired ML-models provide better results with R2 of 0.96 for CaCO3 and 0.78 for TOC than conventional methods. In our case study, the ML-performance for TOC is comparably lower but still provides potential for future optimization. Altogether, this study allows to conveniently generate high-resolution bulk chemistry records without losing accuracy.