Machine Learning-Based Identification of Fault Structures in Seismic Refraction and Resistivity Tomography

Abstract

This study identifies subsurface geological structures by integrating geophysical tomography and machine learning techniques. Borehole drilling methods for subsurface detection are both costly and spatially limited. Geophysical imaging methods, including seismic refraction tomography (SRT) and electrical resistivity tomography (ERT), are typically employed to address these challenges. In this study, geophysical imaging was utilized, and their inversion results were processed with the k-means clustering machine learning approach to address geological challenges along the proposed Porsgrunn Highway in Norway. A total of 39 SRT and 5 ERT profiles were conducted and comprehensively evaluated using advanced post-inversion processing with the k-means clustering algorithm from the scikit-learn open-source resource in Python. The study findings show six exemplary SRT and four ERT profiles. We employed both the Elbow and Silhouette evaluation methods to ascertain the optimal cluster numbers for unsupervised k-means clustering. These methods consistently identified the optimum cluster number, with only a few exceptions. The use of both methods of evaluation is beneficial since they can complement each other and improve cluster validation. The results show that the clustered models adequately correspond to distinct lithologic units, effectively revealing fault zones and bedrock depths. Our findings suggest that combining geophysical imaging with machine learning provides valuable insights for improving subsurface geological interpretations, allowing for informed decisions in geological hazard assessment and engineering construction site selection.