A deep learning method for 3D geological modeling using ET4DD with offset-attention mechanism

Abstract

Deep learning-based methods for 3D geological modeling can automatically identify significant geological features, which is crucial for intelligent 3D geological modeling. We propose a 3D geological modeling method based on ET4DD (Enhanced Transformer for Drilling Data). This deep learning model accurately predicts the lithology categories of 3D points. The study area for our experiment is located in the Tianfu New Area of Chengdu, Sichuan Province, China. We conduct data pre-processing operations, including resampling and standardization, on the data collected from 719 boreholes in the study area. The dataset was split into training and test sets at a 4:1 ratio. To validate the effectiveness of the model, we train a standard stacking model that integrates FNN, RF, GBDT, and XGBoost using the same dataset. The comparison shows that ET4DD achieves the highest precision, recall, and F1 score among all models, with respective scores of 97.33 %, 97.33 %, and 97.29 %. We use MapGIS software to visualize the lithology grid cells predicted by ET4DD, and select three subregions from the geological model for detailed comparison with the stacking method, complemented by visualizations of uncertainty. The results demonstrate that our method effectively captures geological variability and reduces the informational complexity of the geological model. In addition, the geological model generated by our method reveals the geological regularities, including the topology of geological bodies, the geometrical form of strata, and the spatial distribution characteristics of lithological units, leading to more accurate and reasonable simulations of actual subsurface geological conditions.