Automated 3D Seismic Interpretation With Swin Transformer Pretraining
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current seismic interpretation methods lack computational efficiency and accuracy, particularly in 3D seismic imaging, due to the challenges of discontinuity in 3D neural networks, limited hardware memory, and the lack of large-scale, geologically-diverse datasets for pre-training, leading to performance degradation in field seismic data.
Innovation Solution
Pre-training a 3D Swin Transformer on a large unlabeled field seismic database and fine-tuning it with labeled synthetic data for tasks like fault detection, horizon prediction, and image denoising, using self-supervised algorithms and advanced decoders to extract multi-scale features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If 3D neural networks are used for seismic interpretation, then interpretation accuracy is improved, but computational efficiency deteriorates due to hardware memory limitations
Solution Approach 1:
The patent divides the 3D seismic volume into multiple 2D slices that can be processed independently or in small batches. This segmentation allows standard 2D CNNs to process each slice, avoiding the memory requirements of full 3D networks while maintaining interpretation accuracy through multi-scale feature extraction across slices.
Solution Approach 2:
The patent transitions from 3D spatial processing to 2D slice-based processing, effectively reducing the dimensionality of computation. By processing 2D slices with appropriate feature aggregation, the system achieves 3D interpretation capabilities without the computational burden of true 3D neural networks.
2Adaptability or versatility
If large-scale pre-training is performed on field seismic data, then model generalization is improved, but data requirements and computational resources increase
Solution Approach 1:
The patent performs pre-training on large-scale synthetic seismic data before fine-tuning on smaller field datasets. This preliminary action on synthetic data allows the model to learn general seismic features without requiring extensive field data, reducing the actual field data requirements while maintaining good generalization performance.
Solution Approach 2:
The patent uses synthetic seismic data as a copy or approximation of field data for pre-training. This synthetic data can be generated in large quantities without physical acquisition costs, providing sufficient training material for generalization while avoiding the need for large volumes of actual field data.
Data Source
AI summary
A method is described for 3D seismic interpretation using a pre-trained 3D transformer that is fine-tuned for desired types of interpretation. It may be used with field seismic or synthetic seismic. The method is executed by a computer system.