Abstract: With the advancement of artificial intelligence technology, certain machine learning models have gradually been applied in the field of geotechnical engineering. These models aim to accurately predict the surface settlement caused by shield tunneling and facilitate the timely implementation of appropriate measures. Leveraging data from the construction of the Shanghai Beiheng Tunnel’s super large diameter shield tunneling, three types of machine learning models were employed: random forest (RF), support vector machines (SVM), and extreme gradient boosting (XGBoost). Geological, geometric, and advancement parameters were utilized as inputs to forecast the final surface settlement resulting from the excavation of a super large diameter shield tunnel. A genetic algorithm (GA) was used to obtain the optimal hyperparameters for each machine learning model. Using the initial 70 data points within a specific section as the foundational training set, dynamic predictions were made for 27 subsequent data points. This allowed for the assessment of the models’ accuracy in predicting the final surface settlement. The results indicated that the SVM’s predictive performance was relatively poor (with a mean squared error (MSE) of 5.06 and a coefficient of determination (R²) of 0.13). In contrast, the two decision-tree-based ensemble algorithms (RF and XGBoost) exhibited better predictive results for surface settlement. Specifically, RF demonstrated an MSE of 1.35 and an R² of 0.70, whereas XGBoost had an MSE of 2.2 and an R² of 0.57. These ensemble algorithms can accurately predict the final surface settlement and thereby provide a novel approach for controlling construction and predicting surface settlement in the context of super large diameter slurry shield tunneling in soft soil regions.

Key words: super large diameter shield, machine learning, surface subsidence prediction, dynamic prediction