<jats:p>Detailed knowledge of the lithostratigraphy of unconsolidated sediments is essential for many scientific and industrial applications. Although drill cores provide lithological and petrophysical information (e.g., lithofacies, permeability, consolidation), core recovery is time- and resource-intensive. In contrast, flush drillings are faster and less expensive but lack detailed geological context. Therefore, combining the advantages of both methods can significantly enhance subsurface investigations while reducing reliance on costly core drillings. This study explores the use of unsupervised machine learning to build data-driven stratigraphic models from standard petrophysical and geochemical wireline logging data. Specifically, it applies dimensionality reduction (UMAP) and hierarchical clustering to identify distinct lithofacies types. The method was tested on datasets from the former Rhine Glacier area in Germany, including one core-controlled well and a nearby flush-drilled borehole. The developed workflow predicted the lithofacies of the core-controlled well with ~76% accuracy, capturing both major stratigraphic units and finer internal features, directly linking them with the geology identified at the drilled site. This allows linking the reconstructed lithology of the core-controlled well with the flush-drilled well. The results demonstrate that unsupervised machine learning can significantly improve stratigraphic models of unconsolidated Quaternary sediments using wireline logs, with minimal dependence on core data.</jats:p>
