Bayesian Neural Network Prediction and Uncertainty Analysis of Bio-Cemented Soil Strength

Abstract

Microbially induced carbonate precipitation (MICP) has emerged as a promising ground improvement technique, with MICP-treated soils exhibiting substantial enhancements in strength. However, experimental results revealed significant variability in strength outcomes of MICP-treated soils, even under identical treatment conditions and soil properties. This uncertainty in strength is challenging to capture using traditional predictive approaches such as conventional constitutive models. The present study leverages artificial intelligence to address the challenge by developing a Bayesian neural network (BNN) model for predicting the strength of bio-cemented soils while considering uncertainty. A dataset comprising 480 experimental samples was used to develop the model. The results indicate that carbonate content and confining pressure emerge as the most influential factors governing the strength of bio-cemented soils. The BNN model exhibits lower uncertainty when predicting bio-cemented soils with relatively low strength, while demonstrating higher uncertainty for soils with strength exceeding 2 MPa. Moreover, micromechanical investigations using the discrete element method (DEM) reveal that multiscale factors, including crystal distribution patterns, fabric and spatial heterogeneity of precipitates, contribute significantly to the strength uncertainty of bio-cemented soils. The developed BNN model provides an alternative tool for predicting bio-cemented soil strength with quantified reliability, facilitating the design of MICP treatment and its application in geotechnical engineering.