Applying a wide and deep learning model to core-scan XRF data to reconstruct mineral assemblages for Pleistocene paleolake Olduvai, Tanzania

Abstract Paleolake coring initiatives result in large datasets from various proxies taken at different resolutions, ranging from continuous scans to samples collected at coarser intervals. Higher-resolution data (e.g., core-scan X-ray fluorescence XRF) can detect short-duration changes in the paleolake and help identify unit boundaries with precision; however, interpreting the causes of such changes may require sampling and more intensive laboratory analysis like X-ray diffraction (XRD). This study applies a published wide and deep learning model, developed for the Olduvai Gorge Coring Project (OGCP) 2014 cores from the Pleistocene Olduvai basin, Tanzania, to reconstruct the mineral assemblages from saline-alkaline paleolake Olduvai using core-scan XRF data and core lithology. A classification model (predicting mineral presence or absence) and a regression model (predicting relative abundances of minerals) yielded predictions for two OGCP cores (2A and 3A), which were compared with published XRD mineral data and detailed core sedimentological descriptions. The models were excellent at identifying dolomite-rich layers, carbonate-rich intervals, intervals of sandstone within claystone, and altered tuffs within claystone and at predicting whether illitic or smectitic clays dominate. The models struggled with less-altered tuffs and with zeolites in non-tuff sediments, especially when XRD identified chabazite and erionite (rather than phillipsite) as the dominant, non-analcime zeolite.