Multi-source data integration for soil parameter uncertainty quantification

Calibration of soil parameters is essential for reliable geotechnical analysis, yet remains challenging due to limited field measurements, model discrepancy, correlated outputs and the need to incorporate multiple-source data types. This paper proposes a sequential Bayesian framework that enables staged data assimilation and progressive refinement of a wide range of model parameters, thereby improving the reliability of predictions of engineering behaviour throughout the entire construction and operation life cycle. A reduced-order modelling approach, combining principal component analysis and polynomial chaos expansion, is employed to construct efficient surrogates that preserve output correlations and support global sensitivity analysis. A custom likelihood formulation is developed to integrate multiple observation types while explicitly accounting for model discrepancy and uncertainty. To focus calibration efforts on critical regions, a spatial weighting strategy is used, reflecting zones of engineering significance. The proposed methodology is demonstrated through a full-scale laterally loaded pile test. Results indicate that the surrogate model yields accurate and computationally efficient predictions, and that the Bayesian updating process systematically reduces parameter uncertainty while maintaining consistency with field observations.