TUNSET: A Machine Learning Tool for Predicting Tunnel Settlement

Resumo

Urban expansion and increasing population density underscore the need for efficient underground transportation systems such as subways and tunnels. These infrastructure projects frequently face significant engineering challenges, particularly regarding surface structure damages caused by ground settlements and the high costs associated with extensive soil characterization through traditional physical testing methods. This study aims to overcome these challenges by proposing a machine learning-based metamodel for accurately predicting ground displacement induced by tunnel excavation.The methodology integrates empirical correlations between Standard Penetration Test (SPT) data and soil properties into numerical simulations using the finite element software ADONIS. Subsequently, a Random Forest regression model, implemented with the Python scikit-learn library, was developed to replicate the numerical model's predictions. Monte Carlo simulations were employed to evaluate the reliability of the metamodel, significantly enhancing computational efficiency and facilitating a high volume of simulations.Results demonstrate that the Random Forest algorithm achieved high predictive accuracy with a coefficient of determination (R²) of 0.994 for displacements and 0.992 for axial forces, with moderate accuracy (R² = 0.857) for bending moments. The metamodel notably reduced computational resources and analysis time compared to conventional finite element methods, highlighting its practical applicability and efficiency in real-time geotechnical engineering scenarios.The successful integration of robust geotechnical correlations, numerical modeling, and advanced machine learning underscores the transformative potential of this approach, offering a reliable tool for proactive risk management and optimized decision-making in tunnel construction projects.