Abstract
Drilling‑induced fractures (DIFs) form when hoop stress around a borehole exceeds tensile strength during drilling. In caprocks, they can compromise integrity and elevate risk in geological carbon storage (GCS). To warrant inclusion in containment models, DIFs must be systemic rather than incidental. This requires regional assessment, which is often impeded by the lack of expensive cores and image logs-data types typically used to document DIFs. We address this knowledge gap through machine learning (ML). Using data from two Woodford Shale wells in the US Midcontinent, we train a simple tree-based classifier model, Extreme Gradient Boosting, to predict DIF from conventional wireline logs. The model, tuned on one well and blind-tested on another, achieved 91.48% training and 85.18% testing accuracy. Only six logs-bulk density, shear slowness, neutron porosity, toolface azimuth, spontaneous potential, and intermediate resistivity-were required for achieving reasonable predictions. We applied the frozen model to a third nearby well and identified several intervals with high DIF susceptibility. Although the dataset used limits regional inferences, our workflow enables routine, transferable basin-scale screening for other caprocks.