Interpretable machine learning for selection of site-specific soil constitutive models and consolidation settlement analysis

Machine learning (ML) has achieved great breakthroughs in many fields, such as computer vision and speech recognition, and it also gains an increasing attention in geotechnical engineering. However, it is often criticized for a lack of interpretability. This study proposes an interpretable sparse dictionary learning (ISDL) approach for selection of suitable soil constitutive models (SCMs) and geotechnical consolidation analysis in a specific site. ISDL is inherently interpretable since it expresses prediction of geotechnical responses (e.g., consolidation settlement) as a weighted summation of many elementary datasets (i.e., dictionary atoms in ISDL). The dictionary atoms are constructed using numerical analyses (e.g., finite element model) with different candidate SCMs and parameters. Interpretability of ISDL is improved by a small number of non-trivial atoms (or SCMs) selected for model prediction using site-specific monitoring data and further enhanced by contribution ranking of the selected SCMs using a game theory-based approach. The proposed method is illustrated using a real reclamation project in Hong Kong, and it is shown to effectively select the most suitable constitutive models for the given site, quantify contributions of the selected models, and significantly improve settlement predictions with rigorously quantified prediction uncertainty, particularly at locations without monitoring or future time steps.