Bayesian neural network-based uncertainty modelling: application to soil compressibility and undrained shear strength prediction

This study adopts the Bayesian neural network (BNN) integrated with a strong non-linear fitting capability and uncertainty, which has not previously been used in geotechnical engineering, to propose a modelling strategy in developing prediction models for soil properties. The compression index C_c and undrained shear strength s_u of clays are selected as examples. Variational inference (VI) and Monte Carlo dropout (MCD), two theoretical frameworks for solving and approximating BNN, respectively, are employed and compared. The results indicate that the BNN focused on identifying patterns in datasets, and the predicted C_c and s_u show excellent agreement with the actual values. The reliability of the predicted results using BNN is high in the area of dense datasets. In contrast, the BNN demonstrates low reliability in the predicted result in the area of sparse datasets. Additionally, a novel parametric analysis method in combination with the cumulative distribution function is proposed. The analysis results indicate that the BNN-based models are capable of capturing the relationships of input parameters to the C_c and s_u . BNN, with its strong prediction capability and reliable evaluation, therefore, shows great potential to be applied in geotechnical design.