TL;DR
This paper introduces a novel method for seismic inversion that significantly speeds up the process of creating detailed subsurface images. By combining a "frozen Gaussian approximation" (FGA) with a multi-level particle swarm optimization (MLPSO) algorithm, the authors achieve accurate results much faster than traditional methods, making high-resolution seismic imaging more accessible.
What Problem Does This Paper Address?
Seismic inversion is a crucial technique for understanding the Earth's subsurface, used in everything from oil and gas exploration to earthquake hazard assessment. The goal is to reconstruct a detailed map of underground rock properties (like velocity) by analyzing how seismic waves travel through the Earth. The most powerful method for this is Full Waveform Inversion (FWI), which aims to match simulated seismic wave data with real-world observations.
However, FWI faces two major challenges:
- Non-convexity: The mathematical problem of FWI is highly "non-convex," meaning there are many "local minima" or incorrect solutions that can trap standard optimization algorithms. This makes it difficult to find the true, global best solution without a very good starting guess, which is often unavailable.
- Computational Cost: To achieve high resolution, FWI requires simulating seismic waves at very high frequencies. These high frequencies have short wavelengths, necessitating extremely fine grids for accurate simulations. This leads to an enormous computational demand, making FWI practically prohibitive for many applications.
This paper tackles these issues by developing a faster and more robust FWI method.
Methods
The authors propose a two-pronged approach:
- Frozen Gaussian Approximation (FGA): Instead of using computationally expensive direct simulations of wave propagation, FGA approximates the wavefield using a sum of Gaussian functions. This method, based on asymptotic analysis, is significantly faster, especially for high-frequency waves. However, FGA's accuracy depends on the smoothness of the subsurface model and can degrade in certain areas.
- Multi-Level Particle Swarm Optimization (MLPSO): To address the non-convexity and leverage the speed of FGA, the authors employ a modified Particle Swarm Optimization (PSO) algorithm.
- Constrained Optimization: To mitigate FGA's accuracy limitations, they introduce a constraint that keeps the optimization search within regions where FGA is reliable. This effectively smooths the resulting velocity profile, making it consistent with the original problem when solved exactly.
- Multi-Fidelity: The MLPSO algorithm uses FGA solvers with different levels of fidelity (i.e., different numbers of Gaussian functions). At the beginning of the optimization, when particles are far from the solution, a less accurate but faster FGA is used. As the swarm converges, the algorithm gradually switches to a more accurate, albeit slower, FGA. This "multi-level" approach balances computational cost and accuracy.
Datasets
The paper evaluates its method using a synthetic dataset based on the smoothed Marmousi model. This is a standard benchmark in seismic imaging. The authors simulate seismic wave data using a highly accurate method and then use their FGA-based MLPSO to invert for the velocity model, comparing the results to the known "true" model. They do not introduce new benchmark datasets.
Key Insights and Ideas
- FGA as a Game Changer for High-Frequency Inversion: FGA offers a substantial speedup for high-frequency seismic wave propagation, which is a major bottleneck in FWI. This opens the door to more detailed subsurface imaging.
- The Power of Multi-Fidelity Optimization: The MLPSO approach intelligently uses computational resources by employing low-fidelity (fast) models when accuracy is less critical and high-fidelity (accurate) models when it matters most. This is a key strategy for tackling computationally intensive problems.
- Constraining for Robustness: Introducing a constraint based on the FGA's validity region is crucial. It prevents the optimization from getting stuck in regions where the FGA approximation is poor, leading to more reliable inversion results.
- PSO's Adaptability: The paper demonstrates that PSO, a well-established global optimization technique, can be effectively adapted to handle constrained problems and multi-fidelity models.
Key Results
- The proposed FGA-based MLPSO method successfully reconstructs the smoothed Marmousi velocity model.
- The FGA with 3000 Gaussians provides a sufficiently good approximation of the wavefield, while 300 Gaussians show larger errors.
- The MLPSO method, by strategically using different FGA fidelities, achieves comparable or better results than standard PSO with only a high-fidelity model, but with significantly reduced computational cost.
- Table 1 shows that the MLPSO with a carefully chosen percentage function (p2) achieved comparable results to the high-fidelity model (p=1) while using fewer high-fidelity model evaluations and less computation time. For instance, p2 required only 6 hours of computation compared to 56 hours for p=1, while achieving a similar best function value.
Limitations and Future Work
- FGA Accuracy Dependence: The accuracy of FGA is still dependent on the smoothness of the velocity profile. While the proposed constraint helps, it doesn't entirely eliminate this limitation.
- Constraint Parameter Tuning: The choice of the constraint parameter ($\delta$) in the optimization problem is problem-dependent and can be difficult to determine a priori. The authors suggest trying different values and selecting the one that yields the best results.
- Future Work:
- Combining FGA with Direct Solvers: The authors suggest combining FGA with direct simulation methods to achieve even higher accuracy, perhaps by using FGA to build a surrogate model that is then refined by direct simulations.
- Quantifying Uncertainty: They propose extending their work to Bayesian inversion methods, which are crucial for quantifying uncertainty in the reconstructed subsurface models. This would involve using FGA to efficiently perform the large number of simulations required for Bayesian analysis.
Summary updated