Use of deep learning and self-supervised methods to accelerate and improve the earth subsurface characterization

Area of Science:

• Geophysics
• Artificial Intelligence
• Machine Learning

Background:

• Well log data is crucial for understanding subsurface geology.
• Traditional methods for analyzing well logs can be time-consuming and limited in scope.
• Advancements in deep learning offer new possibilities for automated subsurface characterization.

Purpose of the Study:

• To develop a self-supervised deep learning model for accurate geological formation identification using well log data.
• To adapt the Convolutional Visual Transformer (CVT) architecture for 1D well log analysis.
• To demonstrate the model's effectiveness and generalizability across different geological basins.

Main Methods:

• A one-dimensional adaptation of the Convolutional Visual Transformer (CVT) model was employed.
• The foundation model was pre-trained using self-supervised learning on unlabeled multivariate sensor data from diverse basins.
• The model was fine-tuned for geological formation identification in the Williston Basin and validated in the Groningen gas field.

Main Results:

• The model achieved an average F1 score of 0.94 across six key formations in the Williston Basin.
• Demonstrated faster convergence, increased robustness to missing input data, and improved accuracy compared to baseline models (U-Net, XGBoost, SVM, KNN).
• Successfully generalized to geologically distinct basins, confirming its broad applicability.

Conclusions:

• The proposed self-supervised deep learning methodology offers a robust and accurate solution for subsurface characterization using well log data.
• This AI framework supports scalable and transferable solutions for hydrocarbon exploration, carbon storage, geothermal, and groundwater reservoir characterization.
• The findings contribute to advancing AI applications in the energy transition and resource management.