CPT-based two-dimensional soil stratification using non-stationary mixture of Gaussian processes

Abstract

In geotechnical practice, the cone penetration test (CPT) is widely used for soil stratification; however, the large spacing between soundings makes it difficult to construct accurate two-dimensional (2D) subsurface cross-sections. Conventional Gaussian Process (GP) models are effective for spatial interpolation under stationary assumptions, but real CPT data often exhibit strong non-stationarity due to geological layering, limiting the applicability of GP-based models in practice. To address this limitation, this paper develops a Mixture of Gaussian Processes (MGP) framework for explicitly modeling the distinct spatial features of different soil layers. The MGP integrates layer-specific GPs with a probabilistic layer identification mechanism, enabling joint representation of stratigraphic boundaries and within-layer variability. The framework is validated using synthetic ground profiles of varying stratigraphic complexity and applied to real CPT data from the Groningen region of the Netherlands. Results show that the MGP generally improves prediction accuracy and uncertainty characterization relative to the GP, particularly in inclined or more complex stratigraphic settings where layer-dependent variability and boundary effects are important.