Bayesian Neural Network for Estimating Stress-Strain Behaviors of Frozen Sand

Abstract

Accurately estimating the mechanical behavior of frozen soil plays a central role in frozen ground engineering. Owing to the nonlinear and uncertain nature, modeling the stress-strain behaviors of frozen soil has been challenging the physics-based models. This study proposed a data-driven approach on the Bayesian neural network (BNN) framework that can precisely estimate the stress-strain behaviors of frozen sand with minimum input requirements. First, a series of triaxial tests were conducted to explore the mechanical behaviors of frozen sand under different conditions of confining stress and temperature. The acquired data were utilized for training the BNN to learn the stress-strain patterns under various conditions. Complicated coupled effects of confining stress and temperature on the variation of stressstrain behaviors of frozen sand were identified by experiment results. The low root-mean-squared error of 0.036 and statistical analysis of the absolute error distribution demonstrated the excellent performance of the BNN in providing a pseudo-continuous stress-strain relationship of frozen. Furthermore, hypothesis cases were presented to analyze the limitations and the applicability of the proposed approach in practices. Given the simplification and flexibility, the BNN based approach is expected to be a versatile means for estimating the mechanical behavior of frozen soil.