Bayesian evidential learning of soil-rock interface identification using boreholes

Identification of the soil-rock interface of geological profiles has been a challenging task for underground construction because of lack of sufficient borehole data. Traditional spatial prediction methods for geostatistics require subjective assumptions in the functional form of the variograms or covariance models. This study aims to evaluate the uncertainty of the soil-rock interface using the Bayesian evidential learning (BEL) framework without the subjective assumptions. A borehole-intensive site is selected to investigate the impact of borehole number and layout on the estimation of the soil-rock interface. The BEL is further applied to predict the soil-rock interface for metro tunnelling, and the results are validated through geophysical interpretations. The study has shown that BEL can effectively learn the covariance features of the priors. The results underscore the importance of borehole planning in obtaining an optimal reduction in geological uncertainty. Sequential estimation of soil-rock interface can significantly reduce uncertainty in elevations across the site, particularly in areas near the boreholes. To mitigate biases in geophysical interpretation of the soil-rock interface, the utilization of BEL prediction could be beneficial.