Data-driven hydrogeological mapping and its impact on the tunnel infrastructure
Loading...
Files
Full Text E-thesis
Date
2025
Authors
Hajighasemi, Zahra
Journal Title
Journal ISSN
Volume Title
Publisher
University College Cork
Published Version
Abstract
Tunnel maintenance and predicting failures during operation have been heavily investigated. This investigation has led to new methods to improve prediction accuracy and extend the service life of tunnels. Among the many factors affecting tunnel performance, the hydrogeological profile of the tunnel route is widely considered one of the most important factors. Over time, it can cause various types of deterioration, such as deformation, material degradation, water leakage, cracks in the lining, and other subsurface issues. For these reasons, creating subsurface geological maps and studying tunnel infrastructure hydrology is essential.
Several traditional methods already exist for building geological maps along tunnels, including spatial interpolation, kriging, and Inverse Distance Weighting (IDW). However, these traditional approaches have notable limitations; they struggle to manage sparsely distributed data, scale poorly with large datasets, and have limited ability to capture complex and nonlinear relationships. As an alternative, Machine Learning (ML) methods have gained popularity because of their precise, fast performance and better handling of the aforementioned drawbacks. Supervised ML techniques can learn the correlations between the input features and predict the appropriate label for the output regarding the training set.
In this study, borehole spatial features and lithological information were manually extracted from paper-based reports from multiple sources, then structured, preprocessed, and integrated into a unified dataset. These data sets were used to train different ML algorithms, including Random Forest (RF), Support Vector Machine (SVM), Categorical Boosting (CatBoost), and eXtreme Gradient Boosting (XGBoost); the performance of these models was compared. Each data point has the following attributes: latitude, longitude, altitude, ground level, borehole name, and type of soil/rock. In addition, the neighborhood aggregation was used to add new features to improve the models' accuracy. Among the implemented algorithms, the RF and CatBoost performed better than the others; both algorithms were used to develop a geological map for the Large Hadron Collider (LHC) tunnel in the European Organization for Nuclear Research (CERN) from point 3 to point 4, where leakage data was available. To complete the hydrogeological profile, the hydrological characteristics of the study area were assessed.
Comparing the hydrogeological profile and the tunnel’s leakage data, no obvious relationship was noted between the geological characteristics and leakage. Nevertheless, a river near SPM9 (about DCUM 8700) contributes to the high leakage percentage in that section, whilst the terrain map is related to the leakage distribution in the longitudinal direction. Darcy's Law states that groundwater flow is directly proportional to the hydraulic conductivity, the cross-sectional area, and the hydraulic gradient. As the difference in hydraulic head (i.e., the elevation difference between the ground surface and the tunnel level) increases, the discharge per unit of groundwater into the tunnel also increases. Therefore, the greater the tunnel depth, that is, the larger the difference between the altitude of the tunnel and ground level, the higher the groundwater table, and thus, more leakages are likely to occur.
Description
Keywords
Geological mapping , Machine learning , Data-driven methods , Borehole log
Citation
Hajighasemi, Z. 2025. Data-driven hydrogeological mapping and its impact on the tunnel infrastructure. MSc Thesis, University College Cork.