Crack density estimation in rock structures using machine learning techniques

Crack density is a key quantitative indicator for assessing the fracture state and stability of rock structures. However, direct field measurement is often impractical due to time, cost, and accessibility constraints, necessitating alternative predictive approaches. This study aims to estimate rock crack density using machine learning techniques with input features based on physical properties consistent with the Biot theory. Rock samples from a tunnel site were categorized into five mineralogical groups – albite, quartz, biotite, calcite, and chlorite – via X-ray diffraction (XRD). To simulate varying fracture states, samples were artificially weathered through cycles of chemical treatment with saline water and slake durability testing. ML models were trained to predict crack density using measured physical properties, yielding R 2 values from 0.03 to 0.98 depending on the mineral group. To enhance performance under data-sparse conditions, an oversampling algorithm was applied, resulting in improved R 2 values exceeding 0.9 across all groups. In addition, feature importance analysis was conducted to identify practical input parameters. Results indicate that compressional and shear wave velocities are among the most influential predictors, enabling accurate and efficient crack density estimation. This study demonstrates the potential for using minimal, measurable parameters in conjunction with ML algorithms to assess rock fracture conditions reliably, offering a practical tool for stability evaluation in construction environments.