Machine Learning-guided Observational Method for Prediction of Preloading-induced Consolidation Settlement

Abstract

The Asaoka’s method, established in the late 1970 s and based on an assumption of one-dimensional (1D) consolidation, has been used worldwide for prediction of reclamation-induced consolidation settlement with the aid of field monitoring data. In the last several decades, the state of the practice in geotechnical engineering has advanced significantly. For example, numerical modeling (e.g., 2D finite element method, FEM) is now commonly used for geotechnical analysis and design of reclamations. The 1D consolidation assumption in the Asaoka’s method is not compatible with 2D FEM analysis for predicting consolidation settlement and its variation in a 2D spatial domain. To tackle this challenge, this study proposes a machine learning-guided observational method for improving prediction of consolidation settlement from 2D FEM models using field monitoring data. It uses random FEM to incorporate various uncertainties and generate many sets of FEM analysis outcomes (e.g., time-varying settlement in a 2D space). Then, these outcomes are adopted as basis functions, or dictionary atoms, under a sparse dictionary learning (SDL) framework and used together with the field monitoring data sequentially acquired to continuously improve predictions of consolidation settlement. A ground improvement project using combined vacuum and surcharge preloading is adopted to illustrate the efficacy of the proposed approach.