Machine learning-enhanced site characterization for tunnel risk assessment

Abstract: 
Tunnelling plays a crucial role in modern urban development and infrastructure construction. However, the underground environment exhibits significant variability. Coupling amplification effects under adjacent disturbances further complicate the scenario. Understanding geological strata distribution and soil parameters is essential for tunnel risk assessment. Site characterization typically relies on limited field information, such as borehole and cone penetration test (CPT) data. Therefore, quantitatively characterizing stratum layering and soil spatial variability based on limited data is crucial. This report presents a machine learning-enhanced approach for site characterization using limited borehole and CPT data to assess tunnel risk. This research first introduces a deep learning-based soil classification model using CPT data. Then, a dynamically updating model to simulate geological uncertainty with tunnel excavation process is given. Subsequently, the intelligent characterization of spatial variability parameters using CPT data is presented, followed by an evaluation of the coupling effect and importance assessment of geological uncertainty and soil spatial variability on tunnel performance. Finally, a deep learning-based tunnel surrogate model is developed to improve computational efficiency, model uncertainty is also evaluated. These efforts contribute to the site investigation and design of tunnels, thereby facilitating decision making and efficient allocation of resources by consultants, operators, and stakeholders.