Finding patterns in subsurface using Bayesian machine learning approach

Abstract Stochastic simulation approaches and uncertainty quantification are usually adopted for gaining insight into variability in soil stratigraphy configurations. Previous investigations at geotechnical site characterization and interpretation can be broadly categorized into geostatistics- and process-based methods. On the other hand, modern site exploration techniques provide high-quality, dense datasets in physical spaces with high resolution, either directly from sensors (for example, cone penetration testing data) or indirectly from geophysical inversion (such as seismic inversion, electromagnetic induction inversion, and ground penetrating radar). In this work, anisotropy and heterogeneity are considered as possible patterns that inherently exist in the observations, and these are inferred and described in a Bayesian manner. To this end, a Bayesian machine learning approach is employed to extract these patterns from the original or interpreted data. The patterns are divided into two parts: spatial and statistical patterns. These patterns are considered as the “hidden link” among multiple spatial datasets. The proposed modeling method is demonstrated using a real-world, one-dimensional example as well as two two-dimensional numerical examples. It is revealed that the proposed clustering approach is a promising tool for subsurface modeling and pattern extraction.