High-Fidelity Seismic Super-Resolution Using Prior-Informed Deep Learning With 3D Awareness.

Area of Science:

• Geophysics
• Artificial Intelligence
• Seismic Imaging

Background:

• Seismic vertical resolution is crucial for identifying thin beds, but current deep learning methods for seismic super-resolution face challenges with unrealistic outputs and poor generalization.
• Traditional 2D seismic approaches often suffer from stitching artifacts and lack spatial consistency, hindering accurate subsurface interpretation.

Purpose of the Study:

• To develop a novel deep learning framework for seismic super-resolution that enhances data fidelity and generalization.
• To address the limitations of existing methods in improving seismic vertical resolution for better thin bed identification.

Main Methods:

• Generation of realistic synthetic seismic training data mimicking field survey characteristics.
• Implementation of an enhanced 2D network with 3D awareness, combining 2D Swin-Transformer and 3D convolution for efficient spatial feature capture.
• Development of a prior-informed fine-tuning strategy using unlabeled field data, incorporating self-supervised data consistency and spectral matching losses.

Main Results:

• The proposed framework demonstrates improved fidelity and generalization in seismic super-resolution tasks.
• The enhanced 2D network effectively reduces stitching artifacts and improves spatial consistency compared to traditional 2D methods.
• Experiments on multiple field datasets confirm the robustness and practical applicability of the method for seismic resolution enhancement.

Conclusions:

• The novel framework offers a robust and generalizable solution for seismic super-resolution, significantly improving the identification of thin beds.
• The combination of realistic data generation, an aware 2D network, and prior-informed fine-tuning overcomes key challenges in deep learning-based seismic enhancement.
• This approach provides a practical tool for geoscientists to achieve higher seismic vertical resolution across diverse field datasets.