Practice of Machine Learning in Geotechnical Engineering

This work aims to elaborate data-efficient modelling of soil properties through effective uncertainty quantification, data fusion and auto-acquisition. Uncertainty quantification is first tailored into various deterministic ML algorithms to enhance their reliability and interpretability. Multi-fidelity learning is then used to fuse low- and high-fidelity data from multiple sources to reduce data demand for estimating soil properties from the specific sites of interest. Given these contributions, an uncertainty-based active learning strategy is further incorporated with multi-fidelity learning to guide data acquisition and reduce data demand, compared with those without data fusion and auto-acquisition for validation. Finally, a generic ML-based modelling platform with a user-friendly graphical user interface is ultimately developed to facilitate their application, such that click is all users need for data import, preprocessing, algorithm selection, hyper-parameters optimization, model development, evaluation, storage, application to new data and illustration of captured correlations.