Absorbing Boundary Conditions

Absorbing boundary conditions (ABCs) are mathematical treatments applied at the edges of a finite computational domain in seismic wave simulation, designed to prevent artificial reflections from the boundaries of the model returning back into the simulation area and contaminating the wavefield. In finite-difference (FD) or finite-element seismic modelling, the Earth is represented as a finite grid of cells that must have edges. Without special treatment, a seismic wave that reaches the grid edge would bounce back as a strong numerical reflection, appearing in the synthetic seismogram as a spurious event that looks like a reflection from a deep geological interface. Absorbing boundary conditions solve this by making the grid edges effectively transparent: the wave energy is gradually attenuated or mathematically absorbed as it approaches the edge, simulating the wave propagating outward into an infinite medium. The Perfectly Matched Layer (PML) is the most widely used modern implementation of absorbing boundary conditions in seismic modelling.

Key Takeaways

The Perfectly Matched Layer (PML), introduced by Jean-Pierre Berenger in 1994 for electromagnetic wave modelling and adapted for seismic applications shortly after, is the current standard absorbing boundary condition for seismic finite-difference modelling. The PML surrounds the computational domain with a layer of artificial attenuating material that is impedance-matched to the interior medium: a wave entering the PML layer passes through the interface without reflection, then decays exponentially as it propagates through the PML layer. The attenuation coefficient increases toward the outer edge of the PML. PML is more effective than older techniques (sponge boundary, tapered damping, Clayton-Engquist paraxial approximations) particularly for waves arriving at oblique angles to the boundary.
The width of the absorbing layer (in number of grid cells) controls its effectiveness. A wider PML absorbs the wave more gradually and more completely. A typical PML layer for a 30-Hz dominant frequency wavelet requires 20 to 40 cells to achieve less than 1 percent reflected energy from the boundary. If the PML is too thin (fewer than 10 cells for this frequency), the gradient of attenuation is too steep and causes numerical instability or imperfect absorption. The cell size must also be fine enough to sample the shortest wavelength in the simulation: at least 5 to 8 cells per wavelength is the standard criterion for second-order finite-difference operators.
In acoustic wave simulation (pressure-only, ignoring S-waves), PML boundaries work effectively for most exploration seismic applications. For elastic wave simulation (including both P-waves and S-waves), the PML formulation requires additional complexity because S-waves have shorter wavelengths (lower velocity) than P-waves at the same frequency, requiring either a finer grid or a modified PML that handles both wave types simultaneously. Anisotropic PML (for VTI or TTI media modelling) further complicates the implementation, as the attenuation must be matched to the directional-dependent wave velocities.
Sponge boundaries (an older alternative to PML) work by gradually increasing the damping coefficient of the medium near the boundary over a transition zone, rather than using impedance-matched attenuation. Sponge boundaries are simpler to implement but require wider transition zones (typically 50 to 100 cells for effective absorption) and perform less well for oblique-incidence waves. In early seismic full-waveform inversion (FWI) codes from the 2000s, sponge boundaries were common because of their implementation simplicity. Most modern FWI and RTM (reverse time migration) algorithms now use PML because of its better performance and efficiency.
Full-waveform inversion (FWI) and reverse time migration (RTM) — two computationally intensive seismic processing algorithms used for high-resolution velocity model building and imaging — both require forward and backward wavefield propagation in 3-D computational domains. The absorbing boundary conditions at the edges of the FWI/RTM domain must be accurate and computationally efficient because the forward-backward propagation is repeated millions of times during an FWI iteration loop. PML conditions with optimized corner treatment (where two PML layers meet at a computational domain corner) reduce the total number of extra grid cells needed while maintaining absorption quality.

Why Absorbing Boundaries Matter for Seismic Modelling

Seismic modelling is used throughout the exploration workflow: to create synthetic seismograms that calibrate the well-to-seismic tie, to test processing algorithms before applying them to field data, and in the inner loop of full-waveform inversion where the modelled wavefield is compared to field data thousands of times to update the velocity model. In all these applications, the synthetic wavefield must faithfully represent how seismic energy would behave in the real Earth, where waves propagate outward without boundary effects.

If boundary reflections contaminate the synthetic data, the synthetic seismogram will show fictitious arrivals that the interpreter might mistake for real geological events. In full-waveform inversion, boundary reflections can bias the velocity model update by appearing as signals that the algorithm tries to match by updating the velocity field in the wrong direction. The consequence is a velocity model with artefacts near the model edges that leak into the interior and reduce imaging quality.

High-quality absorbing boundary conditions are therefore not just a numerical technicality but a practical necessity for reliable seismic interpretation and inversion results.

Fast Facts

Early seismic finite-difference modelling in the 1970s and 1980s used simple velocity-tapering or amplitude-damping zones at the grid edges — the "sponge" approach — which required large numbers of extra grid cells to achieve reasonable absorption and still performed poorly for steep-angle waves. Clayton and Engquist introduced the first practical paraxial absorbing boundary conditions in 1977, which approximated the one-way wave equation at the boundary, achieving much better performance for waves near the boundary's normal direction. The Berenger Perfectly Matched Layer (PML), published in IEEE Transactions on Antennas and Propagation in 1994, was rapidly adopted by the seismic computing community after its superior performance was demonstrated. Most commercial seismic modelling codes (Tesseral, OpendTect, Devito, Madagascar) implement PML or multi-axial PML (M-PML) boundaries as their default edge treatment.

Practical Implementation in Seismic Codes

In a typical seismic forward modelling code, the computational grid represents a rectangle (in 2-D) or a rectangular prism (in 3-D) of Earth model volume. The PML layer is added as a border around the geological model, with the PML cells not representing geological medium but instead functioning as the absorber. The total grid size is the geological model plus 2 × PML_thickness in each dimension (one PML layer on each side).

For a geological model that is 10 km by 10 km by 5 km depth, at a cell size of 10 metres and a PML thickness of 30 cells (300 metres), the total computational grid is 10.6 km × 10.6 km × 5.3 km (adding 300 metres on each side). The total number of grid cells increases from 10,000 × 10,000 × 5,000 = 500 billion cells for the geological model alone to 10,600 × 10,600 × 5,300 = approximately 595 billion cells for the full domain including PML. This 19% overhead in grid cells corresponds to a 19% increase in memory and computation time — a modest cost for the benefits of artifact-free boundaries.

Absorbing boundary conditions are also called non-reflecting boundary conditions, transparent boundary conditions, or radiation boundary conditions. Related terms include Perfectly Matched Layer (PML, the most widely used absorbing boundary condition in seismic finite-difference modelling; uses impedance-matched attenuation to absorb outgoing waves without reflection at any angle of incidence), finite-difference modelling (a numerical method for solving the wave equation on a discrete spatial grid and time steps; requires absorbing boundary conditions at the grid edges to prevent artificial reflections from the numerical domain boundaries), full-waveform inversion (FWI, an iterative seismic processing algorithm that minimizes the difference between modelled and recorded wavefields to update the velocity model; requires absorbing boundaries in the inner-loop forward modelling to avoid model contamination from boundary reflections), reverse time migration (RTM, a seismic imaging algorithm that propagates the source wavefield forward in time and the receiver wavefield backward in time, applying absorbing boundaries to prevent boundary reflections from contaminating the cross-correlation imaging condition), and synthetic seismogram (a modelled seismic trace computed from geological velocity and density models; requires absorbing boundary conditions in the finite-difference simulation to avoid boundary reflection artefacts that would obscure the geological reflections of interest).

How Inadequate Boundary Conditions Created False Amplitude Anomalies in a North Sea Model

A geophysics team at an operator was performing finite-difference forward modelling to test their amplitude versus offset (AVO) analysis workflow on a prospective gas sand in the Brent Group of the Norwegian North Sea. The geological model represented a 15 km × 15 km × 6 km block with a gas sand horizon at 3.2 km depth. The modelling code used a sponge boundary with a 15-cell transition zone at each edge of the domain.

Synthetic shot records from the model showed a series of high-amplitude events at late times (beyond 4.5 seconds TWT) that the team initially misidentified as multiples from the water-bottom and Brent reflectors. When these events were carried into the AVO analysis workflow, they produced apparent Class IIp (gas-sand-like) AVO anomalies at depths below the Brent horizon that corresponded to locations with no geological justification for gas in the velocity model.

A senior geophysicist recognized the events as boundary reflections from the base of the computational domain: the sponge boundary with only 15 cells was providing insufficient absorption for the steep-angle P-waves generated by the modelling at offsets beyond 4 km, and energy was bouncing back from the lower boundary and arriving at the receivers as late-time events. The 15-cell sponge was too thin to absorb these high-angle arrivals effectively.

The team switched to a 40-cell PML boundary, which added 12% to the model domain size and proportional computation time but eliminated the boundary reflections entirely. The corrected synthetic records showed no late-time anomalies, and the AVO analysis of the Brent gas sand target produced a clean Class IIp response consistent with the modelled gas saturation. The false deep anomalies disappeared. The original modelling run had taken 3 weeks on the team's computing cluster; the corrected run with PML added 4 days. The lesson was that the boundary condition specification should always be verified by running a homogeneous half-space model and checking that no energy returns from the boundaries before using the code on a geological model.