Efficient Bayesian model updating for settlement prediction of the immersed tunnel of HZMB

Abstract

To predict the settlement of the Hong Kong-Zhuhai-Macau Bridge (HZMB) tunnel, a physics-informed machine learning (PIML) algorithm was proposed. This method is effective in predicting total settlement with limited training data. However, its performance is poor when applied to the prediction of joint differential settlement, indicating that the commonly used physical model in the algorithm needs further updating. To ensure the efficiency of model updating, a synthetic case study is employed to demonstrate that prediction performance is inconsequential to the simplification of joint shear stiffness but rather to the insufficient number of foundation moduli in the physical model. Subsequently, five criteria are proposed to guide the model design process and guarantee the efficiency of model class selection. The PIML algorithm and the Bayesian probabilistic approach are then employed to select the most suitable model for predicting settlement. The results of model class selection indicate that the criterion related to tube differential settlement is the optimal choice for the HZMB tunnel, with an optimal number of 57 unknown foundation moduli in the updated model. The analysis with field data proves that the updated model class effectively improves predictions for both total settlement and joint differential settlement.