Abstract
Modern seismic exploration faces a significant "efficiency-cost trade-off," where the need for high-resolution subsurface mapping is often hindered by the high costs of field surveys and the resulting data quality issues. This work presents a suite of deep learning frameworks designed to address these challenges. By transitioning from standard supervised learning to self-supervised "Blind-Trace Networks" and physics-augmented models, we provide robust solutions for seismic deblending, interpolation, low-frequency extrapolation, and high-ratio data compression. These methods demonstrate the potential of deep learning to enhance seismic data processing while maintaining geophysical plausibility.
Introduction
Seismic imaging is the primary tool for mapping the Earth's subsurface, but it faces a massive data bottleneck. To reduce the high cost of field surveys, the industry has moved toward "blended acquisition" (firing multiple sources simultaneously) and sparse sampling. However, this creates "dirty" data characterized by interference, aliasing, and missing frequencies. Furthermore, the sheer scale of modern seismic datasets makes storage and transmission prohibitively expensive.
Traditional deep learning (DL) struggles in this domain because it usually requires "clean" ground-truth labels—which are impossible to obtain for the real Earth. This research addresses how to apply DL effectively when labels are scarce and physics is paramount, proposing specialized architectures that move away from "black-box" AI.
Methodology
Several specialized architectures are proposed to overcome the limitations of traditional supervised learning in seismic processing:
Blind-Trace Network (BTN)
The BTN is a novel self-supervised learning scheme. Instead of needing clean data for training, the BTN learns to predict a specific seismic trace based on its neighbors. This allows the model to perform deblending and interpolation using only the corrupted "field" data itself.
Progressive Transfer Learning (PTL)
Used for low-frequency extrapolation, PTL involves training a model on synthetic data and then progressively adapting it to field data. This bridges the "sim-to-real" gap, helping Full Waveform Inversion (FWI) algorithms avoid the "cycle-skipping" problem by providing necessary low-frequency components.
Dip-Informed Neural Network (DINN)
The DINN is a specialized architecture for anti-aliased interpolation. It incorporates the geological "dip" (the angle of subsurface layers) into the neural network to ensure that interpolated data follows realistic geological structures, significantly reducing artifacts that plague standard methods.
Variational Autoencoders (VAE) with Hyperpriors
This deep generative approach for seismic data compression optimizes the "rate-distortion trade-off." It allows for massive file size reduction while preserving the scientific integrity of the waveforms, outperforming traditional transform-based compression methods.
Physics-Augmented Deep Learning
By integrating the underlying physics of wave propagation into the loss functions, we ensure the network's outputs are geophysically plausible and consistent with established physical laws.
Results and Discussion
The methodologies were validated across a mix of synthetic and field datasets, demonstrating significant improvements over traditional algorithms.
- Deblending and Interpolation: The BTN successfully extracted clear signals from simultaneous source interference without requiring a pre-existing clean training library. The DINN approach preserved the continuity of seismic reflectors even in irregular and aliased data.
- Compression: The VAE with hyperpriors achieved high compression ratios while maintaining a lower mean squared error (MSE) compared to JPEG or standard wavelets.
- FWI Enhancement: DL-based low-frequency extrapolation provided a reliable starting model for FWI, successfully imaging complex structures in both synthetic and field cases.
However, limitations remain. Training specialized networks on massive 3D seismic volumes is computationally expensive. Additionally, the performance of anti-aliasing interpolation depends on the precision of initial dip estimation, and the "domain gap" between synthetic training data and variable field conditions remains a challenge for generalization.
Conclusion
This work introduces a comprehensive suite of deep learning frameworks that effectively address the efficiency-cost trade-off in modern seismic exploration. By leveraging self-supervision and physical constraints, these models provide robust solutions for data enhancement and representation, paving the way for more efficient and accurate subsurface imaging.
Summary updated