Pyramid-shaped grid for elastic wave propagation

Feng Chen, Sheng Xu2012 · DOI

TL;DR

This paper introduces a novel "Pyramid-shaped grid" (P-grid) for simulating elastic wave propagation, significantly reducing the number of grid points needed compared to traditional methods. This efficiency gain, achieved by varying grid density with depth, leads to substantial speedups in seismic imaging and modeling, making complex simulations more accessible.

What Problem Does This Paper Address?

Simulating elastic wave propagation is crucial for seismic imaging and modeling, forming the backbone of techniques like reverse time migration (RTM). Traditional methods often rely on regular, rectangular grids. While effective, these grids can be computationally expensive, especially in 3D, requiring a vast number of grid points to accurately capture wave behavior across varying geological structures. Previous attempts to improve efficiency, like using irregular grids in only one direction (e.g., vertically), offered some gains but were limited. The challenge is to create a grid system that efficiently adapts to varying velocities in the Earth's subsurface, reducing computational cost without sacrificing accuracy or introducing significant implementation complexity.

Methods

The core innovation is the "Pyramid-shaped grid" (P-grid). This is a specific implementation of a more general "z-dependent lateral sampling rate scheme" (Z-Grid). The P-grid is characterized by:

  • Non-rectangular shape: The grid cells are not uniform rectangles in the transformed coordinate system.
  • Vertically varying grid density, horizontally constant: The grid becomes coarser (larger cells) in deeper, faster-velocity regions and finer (smaller cells) in shallower, slower-velocity regions. This is achieved through a coordinate transformation that maps the physical subsurface to a regular computational grid.
  • Simple transform/interpolation: The P-grid system has a straightforward mathematical relationship to a regular rectangular grid, allowing for efficient high-order interpolation and parallelization.
  • Compatibility with existing methods: The P-grid is designed to work with finite difference and pseudospectral methods, which are standard in seismic wave modeling.

The paper also discusses the theoretical implications of these transformations on differential operators and addresses the critical issue of numerical stability, particularly concerning the Nyquist wave number, proposing a phase-shifting correction for improved accuracy.

Datasets

The paper uses "real 3D data" for evaluation. Specifically, it mentions:

  • Dataset 1 and Dataset 2: These are characterized by different velocity profiles, leading to varying compression ratios along the z-direction.
  • Example 2D migration images: These are stacked from 168 shots, demonstrating the visual quality of the P-grid results.

No new benchmarks are introduced; rather, existing datasets are used to evaluate the performance of the proposed P-grid against traditional methods.

Key Insights and Ideas

  • Efficiency through geometric adaptation: The most significant insight is that by intelligently adapting the grid's density to subsurface velocity (coarser in fast, deep layers; finer in slow, shallow layers), a substantial reduction in grid points is achievable. This is a direct consequence of the P-grid's pyramid-like structure.
  • Balancing grid reduction and computational overhead: The P-grid design aims to minimize the computational cost associated with the grid transformation itself, ensuring that the gains from fewer grid points aren't negated by complex calculations. The authors highlight that their scheme has a "reasonably-small overhead."
  • Nyquist wave number is critical for stability: The paper emphasizes the importance of addressing numerical artifacts related to the Nyquist wave number, especially in higher dimensions. Their proposed phase-shifting correction is crucial for obtaining artifact-free seismic images.
  • Acceptable trade-offs for RTM: For applications like common-shot RTM, where regions outside the primary focus might be of lower interest, the P-grid's ability to omit modeling in certain areas is a practical compromise that enables significant efficiency gains.

Key Results

  • Grid point reduction: The P-grid scheme uses approximately a quarter of the grid points compared to the previous irregular-z grid scheme.
  • Runtime reduction: Numerical benchmarks demonstrate a reduction in runtimes of up to 79% using real 3D data and sub-optimized code.
  • Theoretical cost reduction: A theoretical reduction in computational costs by up to 83% is estimated for high-order finite difference and pseudospectral methods.
  • Image quality: The P-grid migration images are nearly identical in quality to those produced by regular grids, even with significantly fewer grid points.
  • Nyquist artifact mitigation: The proposed Nyquist wave number compensation effectively reduces artifacts, leading to cleaner seismic images.

Limitations and Future Work

  • Wavefield omission: The P-grid scheme may not model wavefields in regions not covered by the transformed grid. While deemed an acceptable compromise for RTM, this could be a limitation for other applications.
  • Nyquist compensation in 3D: The authors note that over-compensation for Nyquist terms is more significant in 3D and are currently implementing a safer and more efficient compensation method.
  • Theoretical analysis of stability: Further theoretical analysis of stability conditions and the integration of Perfectly Matching Layers (PML) for absorbing boundaries are suggested as important future directions.
  • Optimization of transformations: The current implementation of the transformations is sub-optimized, suggesting further performance gains are possible with code optimization.

Summary updated

Abstract

Elastic wave propagation is elemental to wave-equation-based migration and modeling. Conventional simulation of wave propagation is done on a grid of regular rectangular shape, though other styles, like spherical or ray-family-based, do exist. One of the previously proposed rectangular grid schemes is an irregular vertical (z) grid size. As an extension of this irregular z grid, we design a new grid system, a Pyramid-shaped grid (P-grid), and develop a numerical scheme associated with the elastic wave propagation that can reduce the number of grids, thus improving the efficiency of elastic wave propagation. In our scheme, the grid shape is non-rectangular, and the grid size changes vertically while remaining constant horizontally. For 3D wave propagation, our proposal has a simple transform/interpolation relationship to a regular rectangular grid in all 3 dimensions. Therefore, it yields a very low cost for high order interpolation and allows easy parallelization. In comparison with our previous variable vertical grid size scheme, our proposed scheme uses only a quarter of number of grid points. With numerical benchmarks, we demonstrate a reduction in runtimes of up to ∼79% using real 3D data and sub-optimized code.
BibTeX
@misc{chen2012pyramid,
  author = {Feng Chen and Sheng Xu},
  title = {{Pyramid-shaped grid for elastic wave propagation}},
  year = {2012},
  doi = {10.1190/segam2012-0890.1},
}
Figure