Skip to content
Back to search results

Seismic Full-Waveform Inversion

Introduced by
Albert Tarantola
Year
1984
Type
Seismic imaging and model parameterization technique
Family
Geophysical inversion
Font size

What does it do?

Seismic Full-Waveform Inversion (FWI) is a computational technique that reconstructs detailed subsurface velocity and impedance models by iteratively fitting synthetic seismic waveforms to observed data.

Introduced by Albert Tarantola in 1984, FWI has become the leading method for high-resolution imaging in exploration geophysics, engineering seismology, and subsurface characterization.

The intuition behind the method

FWI works by forward-modeling synthetic seismograms from an initial velocity model, comparing them to real field recordings, and adjusting the model to minimize the difference. This process repeats iteratively, progressively refining the velocity structure. The technique captures all wave phenomena—reflections, refractions, diffractions, and multiples—making it more complete than conventional ray-based methods that use only arrival times.

History

Introduced by Tarantola in 1984 as a parameter estimation framework, FWI remained computationally prohibitive until the late 1990s when adjoint-state methods and GPU acceleration made it practical. Industrial adoption accelerated after 2010, with major applications in deepwater oil and gas exploration, geothermal prospecting, and CO2 storage monitoring. Recent variants include frequency-domain FWI, time-domain FWI, and machine-learning-augmented approaches.

When to use it

FWI is most effective for densely sampled seismic data, moderate frequency content (5–100 Hz), and complex geology where conventional methods fail. It requires known source signatures, good velocity field initialization, and computational resources. Prefer FWI when detailed impedance contrasts and layering are needed; use simpler ray-based methods for quick structural surveys.

Limitations and considerations

  • Computationally intensive, requiring large forward simulations and gradient calculations
  • Sensitive to velocity model initialization—poor starting models lead to non-unique minima
  • Assumes accurate source signature and receiver response; errors propagate through the model
  • High-frequency content improves resolution but requires denser sampling and higher signal-to-noise ratio

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Provides high-resolution velocity and impedance models that capture fine structural details
  • Uses complete waveform information, including multiples and scattered energy, for better illumination
  • Works in complex geological settings where ray methods break down
  • Integrates well and seismic constraints naturally in the inversion framework

Applications

  • Deepwater oil and gas exploration: creating detailed velocity models for accurate depth conversion and structural interpretation in complex salt geometries
  • Geothermal reservoir characterization: mapping fracture zones and thermal anomalies for optimal drilling locations
  • CO2 storage and monitoring: tracking temporal changes in subsurface rock properties to ensure containment integrity
  • Earthquake seismology: imaging crustal velocity structures near fault zones to understand rupture propagation

How it works

How it works

1

Seismic source and receiver geometries are recorded from field surveys, capturing the complete waveform response of the subsurface over time.

2

Raw field data are demultiplexed, gain-corrected, filtered, and de-noised to remove instrument artifacts and improve signal quality.

3

An initial velocity model is created from well logs, stacking velocities, or regional geology; differences between predicted and observed waveforms are computed as a misfit objective function.

4

Iterative optimization updates the velocity model using gradient-based methods (conjugate gradient, L-BFGS) to minimize waveform misfit; forward and adjoint simulations guide parameter adjustments.

5

The final inverted model is validated against well constraints, depth-migrated images, and interpreted geological structures to confirm physical plausibility.

Common pitfalls

  • Overfitting to noise or poor-quality data, introducing spurious velocity anomalies
  • Inadequate frequency content or bandwidth leading to incomplete illumination of the subsurface
  • Neglecting source and receiver signatures, which causes systematic model bias

Frequently asked questions

Why is FWI computationally expensive?

Each iteration requires at least two full 3D wave-equation simulations: a forward simulation to compute synthetic data and a backward (adjoint) simulation to calculate model gradients. With hundreds to thousands of iterations and millions of model parameters, the total cost can reach millions of CPU hours for 3D surveys.

How do I initialize the velocity model for FWI?

Start with well logs for direct velocity constraints, use stacking velocities from conventional processing, incorporate refraction or migration velocities if available, and smooth the result to avoid high-wavenumber artifacts. A well-initialized model reduces the risk of converging to local minima.

What frequency range should I use?

Begin inversion with low frequencies (2–8 Hz) to establish large-scale structure, then progressively add higher frequencies (up to 50–100 Hz) to refine details. This multi-scale strategy exploits the 'cycle-skipping immunity' of low frequencies and benefits from the resolution of high frequencies.

Can FWI handle anisotropy?

Yes, elastic and anisotropic FWI formulations exist but require more computational effort and greater computational complexity. Isotropic FWI is most common in practice due to cost constraints.

Related methods

Sources

  1. Tarantola, A. (1984). Inversion of seismic reflection data in the acoustic approximation. Geophysics, 49(8), 1259-1266.

    Open DOI reference
    Copied.
  2. Virieux, J., & Operto, S. (2009). An overview of full waveform inversion in exploration geophysics. Geophysics, 74(6), WCC1-WCC26.

    Open DOI reference
    Copied.