A foundation model is a large AI model pretrained once on a broad pool of data, then adapted cheaply to many downstream jobs. The pattern transformed language and vision, and it arrived in seismic interpretation with a reflex attached: more pretraining data is always better. For seismic, that reflex misfires. The pretraining corpus that makes a model trustworthy on your acreage is not the largest one available. It is the one that shares your basin's geology and your surveys' acquisition physics. In the illustrative comparison this briefing walks through, a region-targeted model leads a global seismic model by roughly 15 accuracy points on the tasks where amplitude carries the answer, leads an off-domain natural-image baseline by nearly 30, and reaches its target accuracy with about half the labeled data. The buying rule fits on one line: buy match before you buy size.
Why the scaling reflex misfires in seismic
Scaling laws, the empirical result that model quality improves predictably as data and compute grow, made "more data" the default acquisition strategy for AI programs [1]. The result holds when the extra data resembles the deployment domain. Reflection seismic breaks that assumption. Every survey encodes basin-specific rock physics, wavelet character, noise conditions, and acquisition geometry. Pour a hundred mismatched surveys into pretraining and the model spends its capacity learning structures your licence area never contains, while diluting the statistics of the one region you are paying to understand.
A matched corpus does the opposite. It concentrates the model's capacity on the patterns your interpreters will actually meet, so the model starts closer to the answer and climbs faster with each labeled example. Size still helps, but only after match. A smaller corpus that shares your region's physics is a better foundation than a larger one that does not.
The advantage lives where amplitude carries the signal
Seismic tasks split into two families, and the split decides where corpus match pays. Structural tasks, picking horizons and tracing faults, depend on the geometry of reflectors: where events sit in time and how they connect. Geometry generalizes reasonably well across regions, which is why a global model stays competitive on structural work.
Amplitude-driven tasks are different. Direct hydrocarbon indicators, fluid and lithology discrimination, and AVO-style inference (amplitude versus offset, reading how reflection strength changes with source-receiver distance) all depend on the strength of reflections, not just their shape [4]. Reflection strength is governed by local rock physics, and local rock physics is exactly what a mismatched corpus averages away. A model pretrained on someone else's basins reproduces your shapes and misreads your intensities. That is the quiet failure mode: the section looks right, and the fluid call underneath it is wrong.
The instrument below makes this concrete. Drag the amber marker to set the task mix, from purely structural work on the left to purely amplitude-driven work on the right, and watch the three pretraining strategies re-score. At the structural end the three sit within a few points of each other, and an executive could be forgiven for concluding the corpus does not matter. As amplitude takes over, the region-targeted lead grows from about 3 points to about 15, and the natural-image baseline falls furthest.
The pattern to internalize is the widening margin, not any single bar height. Structural benchmarks alone would hide the gap; amplitude-weighted evaluation exposes it. If your portfolio's value concentrates in fluid and lithology calls, and in most exploration and appraisal settings it does, then the corpus-match advantage compounds precisely on your highest-stakes decisions.
Half the labels, and why that is a schedule item
Labeling seismic means paying scarce interpreter time to annotate sections, and it is routinely the binding constraint on an interpretation AI program. In the illustrative figures here, the region-matched model reaches target accuracy with roughly half the labeled sections of the global model, because it is not spending its first thousand labels relearning the basin from scratch. The off-domain baseline needs nearly twice the reference amount. This is standard transfer-learning behavior: features transfer best between domains that resemble each other, and the benefit of a pretrained start shrinks as the source and target domains diverge [3, 5].
Halving the label bill does more than cut cost. It pulls the deployment date forward and releases interpreters from bulk annotation back to the edge cases where their judgment is irreplaceable.
Fewer unexplainable failures, faster sign-off
Decision-makers do not deploy accuracy; they deploy confidence. The failures that stall an AI rollout are rarely the average-case errors a benchmark measures. They are the out-of-distribution surprises, the confident nonsense on a survey vintage or noise regime the model never saw, that a review committee cannot explain and therefore cannot approve. A region-matched model has, by construction, seen data that resembles what production will feed it, so it produces fewer of these surprises. Fewer unexplainable failures means fewer review cycles, and fewer review cycles is the difference between a model that ships this quarter and one that ships next year. Trust is a schedule item, and corpus match is the cheapest way to buy it.
Where the alternatives fit
A global seismic corpus is a strong second, not a winner. It beats off-domain baselines everywhere, holds up on structural tasks, and is the right choice when you genuinely have no regional data to curate. Its ceiling appears on amplitude-sensitive work, where averaging across many regions blurs the local signal you are paying to recover.
Natural-image pretraining, borrowing weights from models trained on photographs, is convenient and cheap, and it is the weakest of the three where it counts. Photographs share edges and textures with seismic sections and nothing of the underlying physics. It is reasonable scaffolding for a cold start with almost no seismic data, and it should be treated as scaffolding to be replaced, not a destination [2].
The economics, in one table
Match lowers three costs at once: fewer labels, fewer failed reviews, and less compute spent pretraining on data that does not transfer. All numbers below are illustrative teaching values from the schematic comparison, not measured benchmarks.
| Metric (illustrative) | Region-targeted | Global seismic | Natural-image |
|---|---|---|---|
| Structural accuracy (geometry) | 0.93 | 0.90 | 0.79 |
| Amplitude-task accuracy | 0.94 | 0.79 | 0.66 |
| Labeled sections to target | ~0.5x | 1.0x (reference) | ~1.8x |
| Out-of-distribution failures | Low | Medium | High |
| Time to sign-off | Fastest | Medium | Slowest |
| Relative pretraining cost | Low | High | Low |
Framed as spend, the governing question is not "how much data can we gather" but "how much of our data actually resembles the decisions we need to get right."
Limitations
The scores, label multipliers, and the 3-to-15-point lead in the instrument are schematic teaching values, interpolated linearly between the structural and amplitude columns of the illustrative table above; they are not benchmarked results from a named model or basin, and no real deployment is reported here. The two-family split of tasks is itself a simplification: real interpretation work mixes geometry and amplitude within a single task, and the clean linear sweep in the exhibit will not hold exactly in practice. The framing derives from lithospheric-modelling work in the source thesis, not from a controlled pretraining ablation, and the argument assumes a curated regional corpus is available to pretrain on; where regional data is genuinely scarce, the global-corpus strategy this briefing ranks second becomes the practical first choice. Before committing budget, ask any vendor for the comparison in the table above measured on your own held-out surveys, with amplitude-driven tasks weighted to match your portfolio.
What this briefing argues
- For seismic foundation models, matching the pretraining corpus to your region's geology and acquisition physics beats maximizing raw corpus size.
- The region-targeted advantage is largest on amplitude-driven tasks, where local rock physics governs the signal; structural benchmarks alone will hide it.
- A matched model reaches deployment accuracy with roughly half the labeled data in the illustrative figures, which pulls the deployment date forward.
- Match reduces out-of-distribution failures, and fewer unexplainable failures is what actually shortens the path to sign-off.
- Global seismic pretraining is a strong second choice; natural-image pretraining is a cold-start scaffold, not a destination.
- Decision rule: buy corpus match before corpus size, and demand amplitude-weighted evaluation on your own held-out surveys.
References
[1] Kaplan, J., McCandlish, S., Henighan, T., et al. (2020). Scaling Laws for Neural Language Models. arXiv:2001.08361
[2] Bommasani, R., Hudson, D. A., Adeli, E., et al. (2021). On the Opportunities and Risks of Foundation Models. arXiv:2108.07258
[3] Yosinski, J., Clune, J., Bengio, Y., and Lipson, H. (2014). How transferable are features in deep neural networks? Advances in Neural Information Processing Systems, 27. arXiv:1411.1792
[4] Amplitude versus offset (AVO), overview: Wikipedia
[5] Transfer learning, overview: Wikipedia
[6] Source: 2018 PhD thesis, Fig 1.3 (LitMod3D lithology slices). The podium scores and label multipliers in the exhibit are schematic illustration for teaching, not benchmarked numbers.
EarthScan scopes region-targeted pretraining corpora for basins where amplitude drives the decision. If your interpretation AI is being evaluated on structural benchmarks alone, we should talk.
