Multivariate Gaussian Process Regression for 3D site characterization from CPT and categorical borehole data
Orestis Zinas, Iason Papaioannou, Ronald Schneider, Pablo Cuéllar
Abstract
Accurate prediction of subsurface stratigraphy and geotechnical properties, along with quantification of associated uncertainties, is essential for improving the design and assessment of geotechnical structures. Several studies have utilized indirect data from Cone Penetration Tests (CPTs) and employed statistical and Machine Learning methods to quantify the geological and geotechnical uncertainty. Incorporating direct borehole data can reduce uncertainties. This study proposes a computationally efficient multivariate Gaussian Process model that utilizes site-specific data and: (i) jointly models multiple categorical (USCS labels) and continuous CPT variables, (ii) learns a non-separable covariance structure leveraging the Linear Model of Coregionalization, and (iii) predicts a USCS based stratigraphy and CPT parameters at any location within the 3D domain. The results demonstrate that integrating geotechnical and geological data into a unified model yields more reliable predictions of subsurface stratification, enabling the parallel interpretation of both USCS classification and CPT profiles. Importantly, the model demonstrates its potential to integrate multiple variables from different sources and data types, contributing to the advancement of methodologies for the joint modeling of geotechnical, geological, and geophysical data. • A methodology for joint modeling of geotechnical and geological data is proposed. • The proposed approach leverages the Linear Model of Coregionalization. • The model parameters are learned through Variational Inference. • The approach allows joint probabilistic prediction of stratigraphy and CPT parameters.