Bayesian Subsurface Mapping Using CPT Data

Data-Driven Site Characterization aids geotechnical engineering and decision-making by producing a 3D soil parameter map of the subsurface. Unfortunately, limited site investigation data availability renders most traditional Machine Learning methods inadequate. In this paper, a framework for Bayesian Site Characterization (BaySiC) is applied on an artificial site investigation dataset. The framework aims to evaluate the statistics of the CPT measurements (cone resistance and sleeve friction) and establish their relationship, but also identify the soil heterogeneity patterns and classify the soil type at each point of the subsurface. The Bayesian framework can deal with small site investigation datasets and quantifies the prediction uncertainty with runtimes that are feasible for practical application. Essentially, the Bayesian framework maps the parameters over the subsurface on a probabilistic level. Therefore, the potency of the framework is not only judged based on traditional metrics, such as the adjacency of the mean prediction to the data, but also the likelihood attributed to the data, which is intuitively expressed by the visualization of the prediction credible intervals over the subsurface. This is shown by means of a case from a benchmark exercise.

Keywords: Site Characterization, Bayesian Inference, Subsurface Mapping, CPT measurements, Machine Learning