High-Frequency Seismic Data Reconstruction With Physical and Structure-Guided Constraints

Reconstructing high-frequency seismic data is essential for detailed subsurface imaging and accurate reservoir characterization, particularly when dealing with complex geological formations. Traditional machine learning approaches typically depend on large labeled datasets, which are often challenging to acquire and may fail to adequately represent the diverse conditions found in subsurface environments. To overcome these limitations, we propose a novel physcis-guided, self-supervise learning methodology that eliminates the requirement for labeled data. This approach integrates physical constraints to ensure that the reconstructed seismic data retains its fidelity. A significant innovation in our method is the application of structure-guided constraints, enhanced by the use of cosine similarity. This allows for the adaptive calculation and minimization of angle errors, effectively preserving the orientation and continuity of geological features. Comprehensive testing on both synthetic and field datasets has demonstrated the robustness and accuracy of our method, showing its capability to maintain high-frequency details and faithfully reconstruct seismic signals in the presence of complex subsurface structures. The results indicate that this method significantly improves the resolution and reliability of seismic data, providing high-quality inputs that can enhance subsequent interpretation in reservoir management.