Benchmarking Study of Three-Dimensional Subsurface Modelling Using Bayesian Compressive Sampling/Sensing

In recent years, three-dimensional (3D) subsurface models have attracted increasing attention for precise site characterization, driving the development of various methods for 3D subsurface modelling. However, limited standard tests (e.g., benchmarks) are available for fairly comparing the results from different 3D subsurface modelling methods. To address this challenge, a benchmarking study is presented in this paper. A series of benchmarking cases using real cone penetration test (CPT) data are developed to evaluate 3D subsurface modelling methods using sparse measurements as input. A suite of benchmarking metrics is proposed to quantify the performance of different methods in terms of accuracy, uncertainty, robustness, and computational efficiency. The presented benchmarking study is illustrated by an in-house software called Analytics of Sparse Spatial Data based on Bayesian compressive sampling/sensing (ASSD-BCS). The performance of ASSD-BCS is not only evaluated using proposed benchmarking cases and metrics, but also compared with GLasso, which is also a 3D subsurface modelling method. The results show that ASSD-BCS and GLasso have similar prediction accuracy, but ASSD-BCS has remarkably high computational efficiency. The computer runtime of ASSD-BCS is three orders of magnitude faster than that of GLasso. In addition, ASSD-BCS provides predicted results with quantified uncertainty, and performs robustly for different benchmarking cases.

Keywords: Benchmarking, Bayesian compressive sampling/sensing, 3D subsurface modelling, cone penetration test