Comparison of Data-Driven Site Characterization Methods through Benchmarking: Methodological and Application Aspects

Abstract

Site characterization is one of the most crucial steps for decision making in geotechnical engineering and to the fullest extent possible should be conducted based on objective data. The current reliance on engineering judgment to interpret data directly cannot exploit the rapid growth of data, machine learning, and other digital technologies. Data-driven site characterization (DDSC) has received much attention in an emerging field called data-centric geotechnics, because a knowledge of the ground is fundamental to geotechnical engineering. As a result, many DDSC methods have been developed recently. Differences and similarities between DDSC methods, however, have not been well studied in terms of methodological and application aspects. This paper proposes a comparison between three emerging DDSC methods from these methodological and application perspectives: (1) geotechnical lasso (Glasso), (2) geotechnical lasso with basis-functions (Glasso-BFs), and (3) Gaussian process regression (GPR). From a methodological perspective, this paper presents a unified Bayesian framework to derive these DDSC methods, in order to shed light on the methodological similarities and differences. From the application perspective, the prediction accuracy for the validation dataset and runtime cost of these three DDSC methods were compared through benchmarking. The differences in performance can be better understood within the unified framework. This paper further proposes a new benchmark involving complex intermixing of soil types, to test the three methods under more realistic and challenging field conditions, although the training and validation datasets remain synthetic.