Bayesian supervised learning of 2D subsurface soil stratigraphy using limited cone penetration tests with consideration of uncertainty

Cone penetration tests (CPT) have been widely used for soil stratification in geotechnical site investigation for decades. However, due to time and budget limits, the layout of CPT sounding at a specific project site is often sparse, leading to significant interpolation uncertainty in the development of subsurface soil 2D cross-section, particularly at locations without CPT measurements. Such development is often com­bined with empirical classification criteria, which further introduce model uncertainty to soil stratification. These uncertainties may pose great risks to the geotechnical engineering practice. A Bayesian supervised learn­ing method is presented in this paper for probabilistic soil stratification in a 2D cross-section using limited CPT. The proposed method cannot only automatically stratify soils in a 2D cross-section from limited CPT soundings, but also can properly quantify the associated uncertainties. Complete 2D CPT data cross-section is firstly learned from limited number of 1D CPT profiles using Bayesian supervised learning. The associated interpolation uncertainty is modelled numerically using non-parametric random field simulation based on the results of Bayesian supervised learning. Parametric autocorrelation function of CPT data along either vertical or horizontal direction is not needed. A probabilistic model is also developed to account for the model uncer­tainty of an empirical soil behavior type classification chart. The interpolation uncertainty and soil classifica­tion model uncertainty are then evaluated simultaneously in a Monte Carlo simulation framework. A simulated data example is used for illustration. The results suggest that the proposed method performs well.