Opportunity
The creation of accurate 3D geological models of subsurface volumes is critical for geotechnical engineering, urban planning, and infrastructure development. However, existing methods face significant challenges due to limited site-specific borehole data and the lack of reliable tools to infer missing subsurface stratigraphy from sparse measurements. Traditional approaches often rely heavily on manual interpretation, which is time-consuming, subjective, and computationally inefficient. Additionally, the absence of qualified 3D training images in practical applications further complicates the modelling process. These limitations hinder the ability to construct high-resolution, data-driven 3D geological models that can serve as digital twins for real-world geographic sites, necessitating an innovative solution to integrate prior geological knowledge with machine-learning techniques for efficient and accurate modelling.
Technology
The patent introduces a computer-implemented method that leverages a machine-learning-based model (e.g., gradient boosting algorithms like XGBoost) to generate 3D geological models from limited borehole data and trusted geological data. The innovation lies in its sequential simulation framework, which processes 2D geological cross-sections along predefined slices and reassembles them into a 3D model. Key advancements include:
1. Use of 2D Training Images: The method bypasses the need for 3D training images by utilizing one or two 2D images representing stratigraphic patterns. These images capture prior geological knowledge (e.g., from nearby sites) and are aligned with borehole data in a 3D coordinate system.
2. Sequential Slice Processing: The subsurface volume is divided into slices (e.g., parallel to training images), and the model processes them in an order prioritized by the amount of available borehole data. Extrapolation and interpolation operations are performed to infer stratigraphy at unsampled locations, with results from earlier slices used as additional data for subsequent slices.
3. IC-XGBoost3D Algorithm: An enhanced version of the 2D IC-XGBoost algorithm enables boundary extrapolation and integrates pre-training for computational efficiency. Multi-scale templates extract stratigraphic relationships from training images, and Monte Carlo simulations quantify uncertainty.
Advantages
- Efficiency: Reduces computational time (e.g., ~14 minutes for 0.5 million voxels) compared to traditional methods.
- Accuracy: Achieves high prediction accuracy (e.g., 95.4% in tests) by combining borehole data with prior geological knowledge.
- Flexibility: Works with isotropic or anisotropic geological conditions using one or two 2D training images.
- Uncertainty Quantification: Provides confidence levels for each voxel via dispersion metrics derived from multiple realizations.
- Practical Applicability: Integrates with GIS platforms and supports dynamic updates as new borehole data becomes available.
Applications
- Smart City Development: Building digital twins for underground infrastructure (e.g., tunnels, foundations).
- Geotechnical Engineering: Site investigation, slope stability analysis, and excavation planning.
- Resource Exploration: Oil/gas reservoir modelling or groundwater mapping.
- Risk Assessment: Quantifying stratigraphic uncertainty for construction projects.
