Soil property recovery from incomplete in-situ geotechnical test data using a hybrid deep generative framework

Abstract

Geotechnical testing serves to assess the strength and stiffness of in-situ soils, for purposes such as informing foundation design. Despite its importance, time constraints, financial considerations, and site-specific limitations often restrict testing to isolated locations with limited horizontal resolution. Therefore, this paper presents a novel hybrid generative deep learning model designed to approximate soil properties across sites based on sparsely sampled geotechnical data. The model uses geological subsurface samples derived from random field theory as ‘a priori’ data for a conditional variational auto-encoder (CVAE) model. By doing so, it attempts to map the relationship between in-situ data and the corresponding spatial coordinates, as well as the inherent link between in-situ data and spatial distribution. Then, in the post-processing phase, a Kriging model interpolates minor discrepancies between the measured and predicted values. To demonstrate its practical application, this paper focuses on cone penetration testing (CPT) as the geotechnical test method. The model's development is thoroughly discussed, followed by the validation using in-situ data and an analysis conducted with synthetic data. It is shown that the uncertainty associated with CVAE-Kriging depends upon both the distance from the sample point and the site's inherent complexity. The proposed methodology not only offers refined subsurface modeling but also expands the understanding of uncertainty in geotechnical testing. Practically, it can assist geotechnical engineers with insights during the survey phase.