Abstract
Seismic data acquisition is a major capital expense in geophysics. To mitigate these costs, researchers have turned to compressive sensing and matrix completion to reconstruct full wavefields from sparse, randomized subsampling. However, designing an optimal sampling mask typically requires computationally expensive wavefield simulations or detailed prior knowledge of the subsurface. This paper proposes a "simulation-free" optimization strategy that treats survey design as a graph connectivity problem. By using a simulated annealing algorithm to maximize the spectral gap of the sampling mask, we demonstrate significant improvements in reconstruction quality on realistic synthetic datasets without the need for physics-based modeling.
Introduction
The challenge of seismic survey design lies in determining which specific locations to measure to ensure the most accurate wavefield reconstruction. While uniform random or jittered sampling are common, they are often suboptimal. Traditional optimization methods involve "simulation-based" loops—repeatedly running expensive wavefield models to evaluate mask performance—which becomes computationally prohibitive for large-scale 3D surveys.
Our approach shifts the focus from physics to the mathematical properties of the sampling mask itself. By optimizing a proxy for reconstruction quality known as the spectral gap, we can design superior surveys efficiently and independently of the underlying geology.
Methodology
The core of our method is the link between the "connectivity" of the sampling mask and the performance of matrix completion algorithms. This connectivity is quantified by the spectral gap, defined as the distance between the first two singular values of the mask.
The Spectral Ratio Proxy
Instead of maximizing the gap directly, we minimize the Spectral Ratio (SR): $$\text{SR} = \frac{\sigma_2}{\sigma_1}$$ where $\sigma_1$ and $\sigma_2$ are the largest and second-largest singular values of the sampling mask. A smaller SR indicates a larger spectral gap, which theoretically correlates with better matrix completion results. The optimization is performed in the midpoint-offset domain, where seismic data naturally exhibits the low-rank structure required for reconstruction.
Simulated Annealing (SA)
To find an optimal mask, we employ a simulated annealing algorithm. Starting with a standard jittered mask, the algorithm iteratively "nudges" a subset of source positions.
As shown in Figure 1, the process involves fixing a majority (e.g., 80%) of the subsampled positions and allowing a smaller fraction (e.g., 20%) to move within a defined range $R$. If a move reduces the SR, it is accepted. If it increases the SR, it may still be accepted with a probability that decreases over time, allowing the search to escape local minima and converge toward a global optimum.
Experimental Results
We evaluated the proposed method using a 2D marine synthetic dataset based on the Compass model, a standard benchmark in the geophysics community. The survey consisted of 300 source and 150 receiver locations with a 20% sampling rate (80% missing data).
Performance Gains
The simulated annealing approach consistently improved the mathematical properties of the mask and the resulting data reconstruction. In our primary experiment, the SR was reduced by approximately 30% compared to the initial jittered sampling.
As illustrated in Figure 6, the optimization process significantly alters the distribution of samples in the midpoint-offset domain. This mathematical improvement translated directly into better reconstruction: the optimized mask yielded a Signal-to-Noise Ratio (SNR) of 14.91 dB for the full shot record, compared to 14.6 dB for the standard jittered mask. Robustness tests across five independent trials showed a consistent SR reduction of at least 11%.
Discussion and Conclusion
The primary advantage of this method is its "physics-free" nature. By optimizing the sampling mask's spectral properties, we bypass the need for expensive wavefield simulations, making the process computationally efficient and scalable. Furthermore, the algorithm respects practical constraints, such as maintaining a jittered distribution to ensure broad coverage across the survey area.
While this study demonstrates success in 2D synthetic environments, future work will focus on scaling the algorithm to 3D field data, where the combinatorial search space is significantly larger. Additionally, we aim to investigate how these optimized masks perform with other reconstruction philosophies, such as sparsity-based curvelet transforms, and under more complex real-world noise conditions.
Summary updated