Bayesian Back Analysis of Spatial Variability with Machine Learning Surrogates

Abstract

Geotechnical sensors provide the advantage of directly monitoring model responses that accurately reflect field conditions. Within these field monitoring data lies the latent potential to glean insights into soil parameters. Beyond relying solely on site-investigation data, the incorporation of field monitoring data serves as a valuable complementary strategy. It aids in evaluating soil spatial variability and addressing uncertainties related to field responses. In this study, a surrogate-based Bayesian back-analysis method is proposed to assess the spatial variability in ground profiles and the uncertainty of field responses. The surrogate models are constructed using machine learning algorithms. To validate the effectiveness of the proposed approach and select the optimal machine learning surrogates, a hypothetical example involving an unsaturated soil slope subjected to rainfall infiltration is first employed. The proposed method is further applied to a hydraulic monitoring project in Hong Kong. The results demonstrate the promising potential of Gaussian process regression with the Matern 5/2 kernel based on 100 training samples for training surrogate models. The saturated hydraulic conductivity obtained from the maximum a posterior (MAP) and borehole logs exhibit similarity, and the MAP estimate accurately captures the observed spatial variation in the dynamic probe test. The proposed method can effectively estimate the soil spatial variability and provides reasonable uncertainty predictions of pore pressure head.