Prediction and Generalization Capability of Machine Learning Models for Shield TBM-Induced Settlement

Abstract

Ground settlement induced by shield tunnel boring machine (TBM) excavation is a major geotechnical concern in urban tunneling because it may affect the safety of adjacent structures and underground infrastructure. In this study, machine learning models were developed to predict the maximum settlement induced by shield TBM excavation using a three-dimensional numerical analysis database comprising 320 simulation cases generated from combinations of tunnel diameter (D), ground elastic modulus (E), face pressure (FP), and backfill pressure (BP). Random forest (RF) and extreme gradient boosting (XGBoost) models were developed and compared with an existing regression-based settlement prediction equation. Predictive performance and generalization capability were evaluated using random split and GroupKFold validation techniques. Under random split validation, RF achieved the highest predictive performance, with a coefficient of determination of 0.997 and a root mean square error of 0.438 mm, followed by XGBoost. Both machine learning models outperformed the existing settlement prediction equation. However, model performance decreased substantially under GroupKFold validation, indicating limited generalization capability under unseen D–E grouped conditions. The results demonstrate that the developed machine learning models provide accurate predictions within the range of tunnel–ground conditions represented by the adopted numerical analysis database. The findings highlight the importance of evaluating both predictive performance and generalization capability, particularly when machine learning models developed from numerical analysis databases are applied beyond the conditions represented in the training database.