Machine learning-aided selection of CPT-based transformation models using field monitoring data from a specific project

Transformation models have been widely used in geotechnical engineering to relate data from lab or field tests (e.g., cone penetration tests, CPT) to design parameters required in geotechnical analysis and design. Proper selection of transformation models is crucial but challenging for accurate prediction of geotechnical responses (e.g., reclamation-induced settlement) in practice. This study proposes a general machine learning framework that accommodates a wide variety of existing CPT-based transformation models and uses field monitoring data (e.g., settlement data observed from a specific project) to select suitable transformation models for improving prediction of spatiotemporally varying reclamation-induced settlement. The proposed approach takes advantage of sparse dictionary learning (SDL) and achieves prediction of settlement by a linear weighted sum of dictionary atoms that are constructed using outputs from finite element models (FEM) of reclamation-induced consolidation. Input parameters of the FEM models are determined using existing transformation models in literature. A transformation model database that relates multiple soil consolidation parameters with CPT data is also compiled for consolidation analysis and dictionary construction in SDL. The proposed approach is illustrated using a real reclamation project in Hong Kong. Results show that the proposed approach provides an effective and transparent vehicle to leverage existing abundant transformation models, identify appropriate transformation models using field monitoring data, and improve prediction of spatiotemporally varying reclamation-induced settlement, with greatly reduced prediction uncertainty. The transformation model selection and settlement prediction are also improved continuously as more field monitoring data are obtained.