TL;DR
MetaPOCS is a novel deep unfolding framework that brings "budget-awareness" to seismic data interpolation. By using meta-learning to coordinate multiple "exits" within a single neural network, it allows geophysicists to trade off reconstruction accuracy for computational speed at inference time without needing to retrain separate models.
What Problem Does This Paper Address?
Seismic data is the backbone of subsurface imaging, but it is rarely complete. Terrain obstacles, equipment failures, and high costs often result in "missing traces" in the raw data. While traditional iterative algorithms like Projection Onto Convex Sets (POCS) can fill these gaps, they are computationally expensive and slow.
Recent "Deep Unfolding Networks" (DUNs) have sped this up by turning iterative steps into neural network layers. However, these models are typically "fixed-depth"—they are optimized to give a final answer only at the very last layer. If you stop the network early to save time (early exit), the intermediate results are often poor because those layers were trained to produce "features" for the next stage, not a finished seismic record. Training a separate model for every possible time budget is a storage and management nightmare.
Methods
The authors introduce MetaPOCS, which treats seismic interpolation as a multi-exit problem.
1. The Backbone: SSI-POCS
The model builds on a spatial-spectral backbone. Each stage (or "unrolled iteration") contains:
- Spatial Branch: Uses Restormer blocks to capture local waveform continuity.
- Spectral Branch: Uses Fast Fourier Convolutions (FFC) to fill large gaps by looking at global frequency patterns.
- Data Reinsertion: A "hard" constraint that ensures the original, observed data points are preserved exactly in the output.
2. Meta-Learned Fusion
The core innovation is how the network is trained. Instead of simply summing the losses from every exit (which causes "gradient conflicts" where different exits want to move parameters in opposite directions), MetaPOCS uses a meta-learning strategy based on Meta-GF.
As shown in the architecture (see Figure 2 and 3 below), the framework treats each exit as a separate "task." It learns tensor-wise fusion weights to combine the parameter updates from different exits. This allows the model to prioritize updates that benefit all exits or resolve conflicts where a shallow exit and a deep exit disagree on how a specific layer should behave.
Datasets
The authors evaluated MetaPOCS on several challenging seismic datasets:
- SEG C3: A structurally complex synthetic dataset.
- MAVO: A field dataset from the North Sea providing real-world acquisition characteristics.
- Model94 & F3: Used for transfer learning and generalization tests to see how the model performs on surveys it hasn't seen before.
Key Insights and Ideas
- Beyond Scalar Weighting: Most multi-task models just reweight the loss functions (a single number per task). MetaPOCS reweights the updates for every individual parameter tensor. This "fine-grained" control is what allows it to stabilize training across many exits.
- Rapid Convergence: MetaPOCS achieves high-fidelity reconstructions much earlier in the network than standard baselines. In Figure 5, you can see the "steep ascent" in PSNR compared to fixed-depth or simple multi-exit strategies.
- Physics-Consistent Supervision: Every exit is supervised in both the time-space domain (amplitude) and the frequency domain (focal frequency loss), ensuring the recovered wavefields are geophysically plausible.
Key Results
MetaPOCS demonstrates a superior "quality-latency" Pareto frontier (see Figure 11).
- Accuracy: On the SEG C3 dataset with 50% random missing traces, MetaPOCS achieved a Signal-to-Noise Ratio (SNR) of 28.30 dB, significantly outperforming other deep learning methods like D3UNet (24.30 dB) and POCS-Net (22.48 dB).
- Efficiency: A single MetaPOCS checkpoint can provide multiple operating points. For example, an early exit (T3) in MetaPOCS can sometimes outperform the final output of a standard fixed-depth network, providing a massive speedup for time-critical applications like real-time quality control during data acquisition.
- Visual Fidelity: Qualitative results (Figure 4) show that MetaPOCS resolves complex "event entanglements" and conflicting dips much more clearly than traditional schemes.
Limitations and Future Work
- Training Overhead: While inference is fast and flexible, the meta-learning training process is more computationally expensive than standard training because it requires calculating multiple "inner" trajectories.
- Noise Handling: The current "hard reinsertion" of data assumes the observed traces are noise-free. In very noisy environments, this could preserve unwanted artifacts.
- Downstream Scope: The study focused on Reverse Time Migration (RTM) for validation. Future work could extend this to evaluate fidelity for Amplitude Variation with Offset (AVO) analysis or elastic wave imaging.
Summary updated