Seismic Image Analysis Using Self-Attention for Long-Range Dependencies
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional backpropagation-enabled processes for seismic data analysis have a limited field of view, which restricts the capture of long-range dependencies in seismic images, leading to reduced accuracy and efficiency in identifying geological features and hydrocarbon occurrences.
Innovation Solution
A method involving dependency-training and label-training of backpropagation-enabled processes using self-attention weights to compute spatial relationships between elements of seismic data, allowing for the prediction of geologic features, geophysical properties, and hydrocarbon occurrences, thereby expanding the field of view and improving prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional backpropagation-enabled processes are used with limited field of view, then computational complexity is reduced, but the ability to capture long-range dependencies deteriorates
Solution Approach 1:
The patent transitions from local spatial relationships (3x3x3 pixel neighborhoods) to global spatial relationships by introducing self-attention mechanisms that operate across the entire seismic volume. This dimensional expansion allows the model to capture long-range dependencies without proportionally increasing computational complexity through hierarchical attention and efficient implementation strategies.
Solution Approach 2:
The patent introduces self-attention weights as an intermediary mechanism that computes spatial relationships between distant elements. These attention weights act as mediators that selectively capture relevant long-range dependencies while filtering out irrelevant information, enabling the model to overcome the limited field of view of conventional convolutional approaches.
2Loss of information
If larger filter sizes are used to expand field of view, then capture of long-range dependencies is improved, but computational time increases significantly
Solution Approach 1:
The patent segments the computation of spatial relationships into hierarchical levels, where self-attention mechanisms operate at multiple scales. This segmentation allows the model to capture long-range dependencies at coarse levels and refine details at finer levels, avoiding the need to process all pixel pairs at full resolution simultaneously, thus reducing computational time.
Solution Approach 2:
The patent implements partial attention mechanisms that focus computational resources on the most relevant spatial relationships rather than computing all possible pairwise interactions. By selectively attending to only the most important long-range dependencies, the model achieves effective field of view expansion without the full computational cost of exhaustive pairwise comparisons.
3Loss of information
If recursively applying small filters is used to expand field of view, then capture of long-range dependencies is improved, but device complexity increases
Solution Approach 1:
The patent merges multiple small filter operations into a single self-attention mechanism that directly computes global spatial relationships. By combining the functionality of multiple recursive convolutional layers into one attention-based operation, the model achieves equivalent long-range dependency capture with reduced architectural complexity and more direct computational paths.
Data Source
AI summary
A method for capturing long-range dependencies in seismic images involves dependency-training a backpropagation-enabled process, followed by label-training the dependency-trained backpropagation-enabled process. Dependency-training computes spatial relationships between elements of the training seismic data set. Label-training computes a prediction selected from an occurrence, a value of an attribute, and combinations thereof. The label-trained backpropagation-enabled process is used to capture long-range dependencies in a non-training seismic data set by computing a prediction selected from the group consisting of a geologic feature occurrence, a geophysical property occurrence, a hydrocarbon occurrence, an attribute of subsurface data, and combinations thereof.