Search

• A novel Bayesian inference method was proposed to study deep excavations. • Machine learning with Bayesian Network could predict soil-structure interactions. • The Bayesian Network model could predict ground/wall responses accurately. • The sequential updating process could improve model performance effectively. • The Bayesian framework could be applied to various geological settings reasonably. The demand for deep excavations to construct underground infrastructure has been growing in metropolitan cities due to rapid urban development and limited land space. In this original research, a machine learning technique based on Bayesian Network is proposed to evaluate and predict soil-structure interactions caused by braced excavations. A novel framework based on Bayesian Network incorporating the Bayesian updating process was proposed to study the ground and wall responses induced by deep excavations. Three established case studies collected from the literature were used to implement and showcase the novel concept. Besides, original two-dimensional finite element models were developed to simulate the first two case studies involving deep excavations in clay in order to generate sufficient ground and wall response data for the training and updating phases of the proposed Bayesian model. The third deep excavation case study carried out in a totally different geology (predominantly sand) was then used to assess and verify the robustness of the Bayesian framework in predicting the soil-structure responses. It was eventually established that the developed novel Bayesian Network model managed to successfully predict the wall deflections and bending moments as well as the ground surface settlements of the third case study with good accuracy through using only two case studies as priors via the versatile sequential updating processes, albeit having totally different geological settings. [ABSTRACT FROM AUTHOR]

This research developed a robust Bayesian framework for evaluating geomaterials and soil-structure interactions involving deep excavation, utilizing Bayesian Regression and Bayesian Network methods that combined data training, validation, and updating into a cohesive framework. A methodology based on Bayesian Generalized Linear Model (a variant of Bayesian Regression) was applied to a laboratory-based case study on sustainable cementitious blends. The framework effectively used past research data to estimate design parameters even though the studies involved different mix design requirements. Then, a framework based on Gaussian Bayesian Network (a variant of Bayesian Network) incorporating Bayesian updating was applied to an established case history for deep excavation in clay. Two-dimensional finite element analysis was used to generate data for model training and updating. The GBN models accurately estimated responses such as wall deflections, wall bending moments, and ground surface settlements through a sequential updating process. The research was conducted using R programming language, with a custom R script designed to follow the process flow of the proposed framework. The successful application and validation of this framework demonstrated its potential for characterizing geomaterials and evaluating complex soil-structure interactions in deep excavation, leveraging on the availability of existing data and a reliable sequential updating process. [ABSTRACT FROM AUTHOR]