Detection of geotechnically critical zones from borehole geophysical logs using Machine Learning

Transport infrastructure in the Pegnitztal (Northern Franconian Alb, Bavaria) traverses karstified Upper Jurassic carbonates, where abrupt strength contrasts, fracture corridors, cavities, and clay-filled voids pose localized geotechnical hazards. This thesis develops and evaluates a reproducible workflow that uses routine borehole log parameters, combined with machine learning, to obtain a geotechnical classification tailored to these conditions. The workflow derives four complementary parameters: dynamic Poisson’s ratio from full-wave sonic data, gamma ray-based shale volume Vsh using the Larionov pre-Tertiary formula, televiewer-based Rock Quality Designation (RQD), and acoustic reflectivity from ABI amplitude, probing stiffness, clay content, jointing, and impedance contrasts, respectively. A dataset of 4083 depth-indexed records from BLM/Deutsche Bahn boreholes is labeled into six classes. The study contrasts unsupervised structure (t-SNE; k-means/agglomerative clustering) with supervised models ranging from linear baselines to tree ensemble methods. Unsupervised clustering shows only weak agreement with the labels, motivating supervised, non-linear decision boundaries. CatBoost is selected as the final classifier, achieving an accuracy of 0.85 and a macro-F1 of 0.81 on the testing set. The resulting workflow offers a practical, data-driven screening tool for weak zones in karstified carbonates, while acknowledging limitations due to class imbalance, overlapping carbonate responses, missing value artifacts, which presumably necessitate recalibration before transfer to other settings.