Seismic Geofeature Discovery Using Self-Supervised Vision Transformers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning models for geofeature discovery in seismic data lack efficiency and effectiveness in unsupervised learning, requiring extensive labeled data and expert intervention for feature extraction.
Innovation Solution
A self-supervised learning approach using a Vision Transformer (ViT) model with self-distillation techniques trains a pre-trained model on unlabeled seismic data to generate rich feature representations, enabling automatic geofeature discovery and rapid seismic interpretation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional supervised learning methods are used for geofeature discovery, then model training requires extensive labeled data and expert intervention, but this increases time consumption and reduces productivity
Solution Approach 1:
The patent implements self-supervised learning where the model learns to predict rotated versions of seismic data patches itself, without requiring external labeled data or expert annotations. The model generates its own training targets by applying rotations (0°, 90°, 180°, 270°) to input patches and learning to predict these transformations, enabling autonomous feature extraction that maintains accuracy while eliminating dependency on labeled datasets
Solution Approach 2:
The patent performs preliminary data augmentation by pre-computing rotated versions of training patches and organizing them into structured datasets before model training. This preliminary preparation of training targets (rotated patches paired with their rotation angles) enables the self-supervised learning process to proceed efficiently without requiring expert intervention during the actual training phase
2Reliability
If complex deep learning models are used to process large volumes of seismic data, then feature representation completeness improves, but computational complexity and training time increase
Solution Approach 1:
The patent divides seismic data into smaller patches (e.g., 64x64 or 128x128 pixel segments) and processes them independently through the Vision Transformer model. This segmentation allows the complex model to focus on local features within manageable units while the collective set of patches covers the entire seismic dataset, making the computational task tractable without sacrificing overall feature representation completeness
Solution Approach 2:
The patent applies data augmentation by generating multiple rotated versions (0°, 90°, 180°, 270°) of each training patch, effectively multiplying the training dataset size without collecting additional real seismic data. This partial action of creating synthetic variations provides sufficient training diversity for the model to learn robust features without requiring proportionally larger computational resources
3Ease of manufacture
If self-supervised learning is implemented on unlabeled seismic data, then the need for labeled data and expert intervention is reduced, but the model requires sophisticated training techniques to achieve comparable performance
Solution Approach 1:
The patent changes the training objective from traditional classification or regression tasks to a rotation prediction task. By modifying the parameter being optimized (from geofeature classification accuracy to rotation angle prediction accuracy), the model can learn meaningful seismic features from unlabeled data through a simpler self-supervised framework, reducing the need for complex labeled datasets while maintaining training effectiveness
Data Source
AI summary
A method for discovering geofeatures in a subterranean formation includes receiving first input data including training seismic surveys. The method also includes training a model based on the training seismic surveys to produce a pre-trained model using a self-supervised learning (SSL) technique. The method further includes receiving second input data. The second input data includes one or more target seismic surveys associated with the subterranean formation. The method also includes generating geological representations of the subterranean formation using the pre-trained model and based on the one or more target seismic surveys. The method also includes generating a database based on the geological representations. The method also includes generating an output based on the database.