A deep learning-driven three-dimensional geological modeling method using sparse borehole sampling data

Abstract

Constructing a three-dimensional (3D) geological model based on limited borehole sampling data is highly important for resource exploration and utilization. However, the sparsity of borehole sampling data is a key factor restricting the accuracy of geological modeling. Traditional explicit modeling still overly relies on expert experience, and the commonly used implicit modeling methods have high requirements for the modeling process, both of which limit the effect of geological modeling under sparse borehole sampling data. Therefore, this paper proposes a deep learning-driven 3D geological modeling method. A data self-organization method based on implicit modeling theory was innovatively developed to solve the problem of dataset construction. Moreover, an autoencoder was employed to eliminate the noise in the self-organized dataset and extract deep features. A ResCapsNet, which combines the advantages of the residual network and the capsule network, was proposed to predict the lithology of discrete units in the target area. We tested the proposed method via simulation experiments and field tests. In the simulation experiments, the method achieved an accuracy of 96.7 % in 3D geological modeling, outperforming 94.5 % of AE-ResNet, 93.6 % of ResCapsNet, 58.1 % of support vector machines, and 35 % of random forest. In the field tests, the accuracy of the constructed 3D geological model reached 99.1 %. Additionally, the influences of borehole sampling intervals and dataset volume on modeling accuracy were analyzed. The results show that the deep learning-driven 3D geological modeling method proposed in this paper can fully utilize sparse borehole sampling data to construct high-precision 3D geological models.