Novel approach to predicting the spatial distribution of the hydraulic conductivity of a rock mass using convolutional neural networks

Abstract

Characterizing the spatial distributions of the hydraulic conductivity of a rock mass is important in geoscience and engineering disciplines. In this paper, the architecture of convolutional neural networks (CNNs) is proposed to predict the spatial distributions of hydraulic conductivity based on limited geological factors. The performance of the CNN model is evaluated using the new data of hydraulic conductivity. A comparative study with the empirical method is performed to validate the reliability of the CNN model. The effect of weathering and unloading on the spatial distributions of hydraulic conductivity is studied using the CNN model. The results show that the hydraulic conductivity predicted by the CNN model is within an error range of 5% compared to Lugeon borehole tests. The predictive accuracy of the CNN method is higher than the estimations made using empirical relationships. The spatial distributions of hydraulic conductivity v. depth can be divided into three stages. In the first stage, the hydraulic conductivity is slightly reduced with increasing depth. Increasing in the depth range of 300–600 m (second stage), the hydraulic conductivity is slightly reduced as a function of the lower degree of weathering. In the final stage, the hydraulic conductivity is not changed by weathering and values gradually converge to a constant as depth increases.