Optimized extreme gradient boosting machine learning for estimating diaphragm wall deflection of 3D deep braced excavation in sand

Abstract

In a deep braced excavation, wall deflection plays an important role in its safety and also adjacent buildings. However, most of previous studies focused on predicting wall deflection caused by excavations in clay. Therefore, this paper intends to develop a machine learning (ML) method to predict the maximum wall deflections of deep braced excavations in sand, which has not yet received much attention. To this end, an advanced ML model of extreme gradient boosting (XGBoost) is employed. The performance of XGBoost model is optimized by using Bayesian optimization (BO) algorithm and fivefold cross-validation technique. For the development of the ML model, an extensive dataset is generated from large numbers of three-dimensional (3D) numerical simulations. The validation of the numerical model is conducted by using a centrifuge test. The wall deflections predicted by the optimized ML model indicate good agreement with the numerical results. The feature importance and feature interactions of the model are clearly assessed. Parametric study is also implemented to quantitatively investigate the effects of the input parameters on the maximum wall deflection. A web application for predicting the maximum wall deflection induced by excavations in sand is deployed and made available to the public.