It is tempting to believe the crown jewel is a model: a particular network, a particular checkpoint, a particular benchmark score. Models are perishable. A rival licenses a bigger base model next quarter, and today's state of the art is a commodity within a year.
What does not commoditise is the loop that produces the model. We run a stack of physics forward models, RAYSUM for anisotropic receiver functions, finite-difference solvers for full waveforms, LitMod3D and ABAQUS for lithospheric structure and mantle-flow-driven deformation, and each one manufactures labelled synthetic data. A synthetic waveform arrives already tagged with the exact crustal structure that produced it: Moho depth, LAB depth, anisotropy, interface dip. That data trains a better model, which improves the products, which sharpen the priors and surface real usage, which tells the forward models what to generate next. The model is an output. The loop is the asset.
Why this particular data is hard to copy
Anyone can buy compute and download weights. Almost nobody can cheaply produce labelled subsurface data, because in the real world those labels barely exist. You cannot dig down to the Moho to check its depth. Ground truth comes from sparse, expensive experiments and is itself an interpretation, which is the quiet problem under every supervised model in this field.
A calibrated forward-modelling stack sidesteps it. The simulator prints ten thousand crustal models spanning every plausible Moho depth and anisotropy geometry overnight, each with its noise-free and its realistically noisy waveform, in volumes and configurations fieldwork will never hand you. That is the point at which advantage stops being about talent or funding and becomes a process a rival cannot simply purchase. The forward models are the mint; the labelled waveforms are what they print.
It also explains why the loop sharpens rather than merely enlarges. As real station data reveals which crustal structures actually occur and where the model is weak, the forward models are steered to synthesise more of exactly those cases: harder anisotropy, ambiguous Moho picks, dipping interfaces. Each generation of training data is more useful than the last, not just more plentiful.
The arithmetic nobody likes
- Parity
- never
- Pace needed from here
- 4x
- Level start needs
- 1x
- Behind at horizon
- 1.2x
- Never reach parity
- 35%
Starting 6 turns late and working at 3 times the leader's opening pace, this follower does not reach parity inside 24 turns. At the horizon it is still 1.2 times behind. 35% of the settings on this board are in the same position at a gain of 8% a turn.
| start delay, turns | effort multiple needed | reaches parity |
|---|---|---|
| 0 | 1x | yes |
| 2 | 2x | yes |
| 4 | 3x | yes |
| 6 (shown) | 4x | yes |
| 8 | 4x | yes |
| 10 | 6x | yes |
| 12 | 6x | yes |
Here is the part worth sitting with, and the exhibit is explicit that every figure in it is illustrative. No growth rate here is measured, for us or for anyone else. What it computes honestly is the shape the arithmetic forces once one side compounds and the other does not.
Left to right is how late a follower starts. Back to front is how much harder it works, as a multiple of the leader's opening pace. Height is how long parity takes.
The intuition most people bring is that these trade off smoothly: start twice as late, work twice as hard. They do not. At a gain of 8% a turn, a follower starting level needs 1 times the pace. One starting 6 turns later, which is 18 months, needs 4 times it. That is the difference between a plan that reads as ambitious and a plan that reads as a hiring impossibility, and it arrives from nothing more exotic than one curve bending while the other does not.
Then raise the gain and watch the back of the block. An amber wall appears: settings where no effort multiple on the control reaches parity inside the horizon at all. At 12% a turn that is 63% of the board. Not slower. Not expensive. Unreachable at any effort the control offers, which is the practical meaning of a compounding advantage and the reason the cliff rather than the slope is the thing to plan against.
What follows for three different readers
For an executive evaluating us, the flywheel reframes the risk question. You are not betting on one model surviving the next research cycle. You are betting on the party positioned to ride every cycle, the one whose forward models can generate training data for tomorrow's problem before anyone has fieldwork for it.
For a would-be competitor, the same arithmetic is a warning worth reading precisely. Catching up is not a sprint against the current product. It is a chase against a moving target, and the required pace is set by how long you wait to start rather than by how hard you are willing to work.
For a partner, it argues for integrating early, while the priors your data helps shape are still being written. The loop absorbs what it is fed, and what it is fed early has disproportionate influence over what it generates later.
The zoom-out
Step back from crustal imaging and this is the pattern that has decided most technology markets. Durable advantage rarely lives in a single artefact, because artefacts get copied, expire, or are leapfrogged. It lives in a compounding process a competitor cannot replicate without paying the same accumulated cost, turn by turn. Search, mapping and recommendation all won on data loops rather than on any one algorithm.
A synthetic-data flywheel is that structure applied to a domain where real labels are close to unobtainable. You cannot sample the Moho directly, which is exactly the setting where a self-feeding, physics-labelled data engine compounds fastest, and exactly why the loop rather than any checkpoint is the thing worth defending.
Key takeaways
- Checkpoints commoditise within about a year; the loop that produces them does not, so the loop is the asset and the model is an output.
- Physics forward models manufacture perfectly labelled synthetic data in a domain where real labels barely exist, because the Moho cannot be sampled directly.
- The loop sharpens rather than merely enlarges: real station data steers the simulators toward the cases the current model handles worst.
- Catching up is not proportional. At an illustrative 8% compounding gain, starting level needs 1 times the leader's opening pace and starting 18 months late needs 4 times it.
- At a 12% gain, 63% of the delay-and-effort settings never reach parity inside the horizon at all. The cliff, not the slope, is what a plan has to survive.
- Every figure in the exhibit is illustrative. What it computes honestly is the shape compounding forces, not any measured rate.
References
- Our own account of what a simulator buys that a sensor network does not: 32 synthetic models beat 1,000 stations.
- RAYSUM, for synthetic receiver functions from dipping and anisotropic layered media: A. Frederiksen, M. Bostock. Modelling teleseismic waves in dipping anisotropic structures. Geophysical Journal International, 141, 401 to 412, 2000. doi:10.1046/j.1365-246x.2000.00090.x
- LitMod, for integrated geophysical and petrological modelling of the lithosphere: J. C. Afonso, M. Fernàndez, G. Ranalli, W. L. Griffin, J. A. D. Connolly. Integrated geophysical-petrological modeling of the lithosphere and sublithospheric upper mantle. Geochemistry, Geophysics, Geosystems, 9, 2008. doi:10.1029/2007GC001834
