A Three-Dimensional Geological Modeling and Uncertainty Analysis Framework Integrating Semi-Supervised Deep Learning and Multiple-Point Geostatistics

To address the problems of low accuracy, poor spatial continuity and high uncertainty in 3D geological modeling caused by sparse borehole data, unbalanced categories and complex geological structures, a phased joint modeling method integrating semi-supervised deep learning and multiple-point geostatistics is proposed in this paper. Firstly, a semi-supervised Spatial-Aware Attention Net (SAAN) based on neighborhood spatial context features is constructed, and the pseudo-label mechanism is utilized to improve the classification accuracy of strata and to generate the initial model; Subsequently, the multiple-point geostatistics Random Neighborhood Search (RNS) strategy is combined to iteratively optimize the areas with low confidence and geological semantic conflicts; Finally, the collaborative consistency entropy is proposed to quantify the two-stage prediction consistency and joint uncertainty. Taking the borehole data in the Chengdu Plain area as an example, the proposed method outperforms baseline methods such as Inverse Distance Weighting (IDW), Support Vector Machine (SVM), Random Forest (RF), Multi-layer Perceptron (MLP), and Kolmogorov-Arnold Networks (KAN) in terms of reproduction of stratum spatial distribution, boundary characterization, and structural continuity. In addition, the collaborative consistency entropy can effectively identify structural uncertainties and avoid misjudgments of high entropy values but consistent predictions. The proposed phased joint modeling framework can achieve high-precision, structurally reasonable and uncertainty controllable three-dimensional geological model reconstruction under the conditions of limited and unbalanced borehole data, providing effective technical support for the characterization and engineering application of underground structures in complex geological environments.