Machine learning-based prediction of shear deformation in soil due to twin tunnelling

As urban tunnel construction becomes increasingly complex, accurately predicting ground shear deformation is crucial for ensuring structural stability. This study develops a machine learning-based model to predict shear deformation in soil during twin tunnel excavation. Numerical analysis using PLAXIS 2D, laboratory model tests, and machine learning models (XGBoost and LightGBM) were combined to assess deformation patterns. The results of inverse analysis closely matched the experimental data, confirming the validity of the numerical approach. The predictive models demonstrated high accuracy, with XGBoost outperforming LightGBM, achieving an error rate of 9.3% compared to 46.1% for LightGBM. Feature importance analysis revealed that pile length, tunnel spacing, and vertical offset significantly influenced deformation behavior. Additionally, hyperparameter tuning using Optuna enhanced the models' predictive performance. However, while the models effectively captured overall deformation trends, they exhibited limitations in accurately predicting localized deformation near the tunnel sidewalls and pile tip. This study highlights the potential of machine learning for geotechnical applications, particularly in underground construction. Future research should expand the dataset with diverse ground conditions and apply explainable AI techniques to enhance model interpretability. The findings contribute to improving the reliability and efficiency of tunnel design and construction in urban environments.