Simulating Only What Changes: How to Recompute Elastic Wavefields Without Recomputing the Whole World
Wave simulations show up everywhere: earthquake science, seismic imaging, non-destructive testing, and medical ultrasound. The goal is simple: you specify a source (an earthquake, a pulse, a transducer) and a model of the material, and you compute how waves travel, scatter, and arrive at sensors.
The hard part is that realistic models are big, and realistic wave physics is expensive. Even worse, many workflows need not one simulation, but hundreds or thousands of very similar simulations. In imaging and inversion, for example, you keep updating the model—often changing only a small target zone each time. Running a full-domain simulation every time is like re-rendering an entire movie scene because you moved one prop.
This blog explains the idea behind immersive wavefield modelling: a way to recompute wavefields after a local model change by simulating only a local region, while still reproducing the same answer you would get from a full-domain simulation (including long-range interactions).
Two key references:
- Acoustic (pressure-wave) immersion, numerically exact via finite-difference injection: van Manen, Li, Vasmel, Broggini & Robertsson (GJI, 2020), DOI: 10.1093/gji/ggaa317.
- Elastic (P–S wave) immersive modelling with efficient injection: Li, Koene, van Manen, Robertsson & Curtis (JCP, 2022), DOI: 10.1016/j.jcp.2021.110826.
The promise is:
- Speed for repeated simulations when only a small target zone changes.
- Accuracy that still includes wave interactions with the full exterior model (including waves that leave and later return).
We only simulate waves inside the local box (red dashed line). A second surface (blue dashed line) is where we record what the outside world would do. Using precomputed responses (think: “how waves travel from blue to red”), the boundary can inject the correct incoming waves and absorb outgoing ones—so the small simulation behaves like it is still embedded in the full model.
1) Why repeated wave simulations become painful
Many wave-modelling tasks are iterative by nature. Full waveform inversion (FWI), survey design, uncertainty studies, and time-lapse analysis all demand many wave simulations. Even if each update is small, rerunning a global simulation recomputes propagation everywhere—again and again.
- Imaging and inversion: update the model slightly, simulate again, compare with data, and repeat.
- Monitoring: test many scenarios where only a small region changes (fluid movement, damage growth, stiffness changes).
- Design and sensitivity studies: try many “what if” versions of a target zone.
If most of the model stays unchanged, it is natural to ask: can we simulate only the part that changes—without losing the influence of the unchanged surroundings?
2) Conventional local replay: what it gets right—and what it can miss
A standard local strategy is:
- Run a global simulation once.
- Record wavefields on a closed surface around a target zone.
- Change the model inside that surface.
- Re-inject the recorded wavefield to reproduce the “incoming” energy inside the target zone.
This is useful, but it can miss an important effect: waves scattered by the modified target can travel into the exterior, reflect/scatter from distant structures, and then come back—sometimes multiple times. One-shot replay does not fully reproduce these higher-order, long-range interactions.
3) The core idea: Immersive Boundary Conditions (IBCs)
Immersive Boundary Conditions make a small local simulation behave as if it were still sitting inside the full model. The boundary is active: it absorbs waves leaving the local box (to avoid artificial reflections) and it emits the correct waves coming in from the unchanged exterior. Because this happens continuously in time, the local region remains two-way coupled to the outside world.
In practice, the boundary updates rely on a library of precomputed wave responses in the unchanged background model (often described using Green’s functions). You compute that library once, then reuse it across many local reruns.
4) Two implementation routes: FD injection (exact) vs MPS (efficient)
Once you accept the IBC concept, the practical question becomes: how do we record and inject wavefields on a finite-difference grid without adding numerical artefacts?
4.1 Finite-difference injection (FD injection): “exact by construction”
FD injection makes the boundary consistent with the discrete finite-difference solver itself. That is why, for acoustic waves, it can match a full-domain finite-difference run down to machine precision.
Strength: extremely accurate (often machine-precision agreement).
Trade-off: more bookkeeping and overhead, especially for higher-order finite-difference stencils.
Reference: van Manen et al. (2020), GJI, DOI: 10.1093/gji/ggaa317.
These are simple test worlds used to check the boundary method. The patterns (dots/objects/lines) act like strong scatterers, so waves bounce and scatter in many directions. This is a tough test: a good immersive boundary must still handle all that scattering without creating fake reflections at the edge of the local box.
The snapshots compare an immersive simulation (local box + IBCs) against a full-domain reference run. The “difference” panels are scaled up heavily; without that exaggeration, the mismatch is essentially invisible. This is the key point: the local simulation can reproduce what the full simulation would have done, while only computing a small region.
4.2 Method of Multiple Point Sources (MPS): “physics-like and cheaper” (elastic case)
For elastic waves (P-waves and S-waves), simulations track several coupled wave components. Many elastic solvers use a staggered grid (different components live at slightly different grid locations), which makes clean recording/injection harder—especially near corners.
MPS represents the boundary effect using many point sources distributed around the boundary. In practice it can be much lighter than FD injection, especially as finite-difference stencils become higher order.
Reference: Li et al. (2022), JCP, DOI: 10.1016/j.jcp.2021.110826.
The full model contains scatterers (green) that reflect and redirect waves. The immersive simulation only runs inside the local box, but it is connected to the outside world via two surfaces: a recording surface (where we measure what the exterior would do) and an emitting boundary (where we inject boundary sources so the local run stays fully coupled to the exterior).
The first column is the full-domain reference wavefield. The next columns show immersive simulations using FD injection and MPS. The difference panels are scaled up strongly to make any mismatch visible. The main takeaway is that both approaches can reproduce the full-domain behaviour inside the local box—while MPS can be cheaper to run and store.
5) What the workflow looks like in practice
Immersive modelling is most valuable when you need many reruns where only a small region changes. A typical workflow is:
- Pick a target zone: draw a box around the part you expect to modify.
- Precompute exterior responses (once): in the unchanged background model, compute the information needed to update the boundary through time.
- Run many local simulations: for each update, simulate only the local box while the boundary supplies the “outside world.”
- Extract outputs: inside the local region (and at receivers), you obtain the same result as a full-domain run.
6) When is immersive modelling worth it?
Immersive modelling tends to pay off when:
- Local changes: only a small subdomain changes between runs.
- Many repeats: you need many recomputations (e.g., inversion iterations or many scenarios).
- Exterior matters: reflections/scattering outside the target zone feed energy back into it.
It is less attractive when:
- changes are widespread (not local anymore),
- you only need one or two runs (setup cost may not pay back),
- memory for stored responses is severely limited (though compression/reduced-order ideas can help).
7) Why this matters (beyond one method)
The broader point is not just a new boundary condition. It’s a way of thinking about wave simulation that matches how many problems are structured: a large environment with a small region of interest that changes.
- Seismic imaging and inversion: efficient updates as local properties evolve.
- Time-lapse monitoring: local changes such as fluid movement or damage growth.
- Non-destructive testing: compute near a suspected defect while respecting the whole structure.
- Medical acoustics: local organs/lesions inside a larger body context.
- Laboratory wave experiments: link a physical experiment to a virtual exterior (“immersive experimentation”).
References
- van Manen, D.-J., Li, X., Vasmel, M., Broggini, F., & Robertsson, J. (2020). Exact extrapolation and immersive modelling with finite-difference injection, Geophysical Journal International. https://doi.org/10.1093/gji/ggaa317
- Li, X., Koene, E., van Manen, D.-J., Robertsson, J., & Curtis, A. (2022). Elastic immersive wavefield modelling, Journal of Computational Physics, 451, 110826. https://doi.org/10.1016/j.jcp.2021.110826