A teleseismic wave arriving at a layered, possibly anisotropic, possibly dipping crust produces a receiver function through a forward operator we can simulate accurately and cannot write a tractable likelihood for. The mapping from model to data runs through wave propagation, deconvolution and noise in ways that resist a clean probabilistic form.
The field's usual response is to substitute something hand-built, an H-kappa stack amplitude or a waveform L2, and then grid-search or run MCMC over two or three parameters. There is a different substitution available: keep the simulator, drop the likelihood, and learn the posterior directly.
Keep the simulator, drop the likelihood
The premise of simulation-based inference is that you can sample from a forward model even when you cannot evaluate its likelihood. For crustal receiver functions that is exactly the situation. Given a crustal model, layer thicknesses, velocities, Vp/Vs, dip and anisotropy, a reflectivity or neural-operator simulator produces a synthetic receiver function, and realistic noise can be drawn from the station's own statistics. What cannot be done is writing the likelihood in closed form to put into Bayes' theorem. Neural posterior estimation turns that from a blocker into the method: draw parameters from a geological prior, simulate the data, and train a conditional density estimator, typically a normalizing flow, to approximate the posterior [2,3,5].
A flow matters here for a specific reason rather than a fashionable one. Crustal posteriors are genuinely multimodal and correlated. The H-kappa trade-off is a correlation a flow represents natively, and real ambiguities produce multiple modes that a point plus a symmetric error bar cannot describe at all.
Amortization is the economic argument
The word carrying the weight is amortized. Simulating the training pairs and fitting the flow happen once, offline, at whatever the simulator costs. Every subsequent inversion is a single forward pass: milliseconds. Per-station MCMC re-pays the full sampling cost for every station and again for every reanalysis as new events arrive.
Across a network of thousands of stations that is the difference between an aggregate compute bill nobody approves and a fixed upfront cost with a negligible marginal one. It is what makes Bayesian crustal inversion at continental scale deployable rather than aspirational.
The prior is a modelling decision, and it has an edge
Simulation-based inference makes the prior explicit and central: it is the distribution the training parameters are drawn from, and it defines the model space the estimator will ever consider. That is a feature, because a geological prior is exactly the domain knowledge an asset team already holds, encoded where it belongs.
It is also a boundary. The estimator is only trained where the prior has mass, so a station whose true structure falls outside it is out of distribution and its posterior is an extrapolation. The exhibit below is mostly about what that costs and about why the obvious check does not catch it.
What simulation-based calibration can and cannot see
- Coverage delivered
- 51%
- Coverage claimed
- 90%
- SBC
- passes
- Outside the prior
- 16%
- Gated out
- 0%
The reported intervals claim 90% coverage and deliver 51%, a shortfall of 39 points. 16% of the network sits outside the prior, and the gate is off, so they are answered by extrapolation and reported like any other station. Simulation-based calibration passes at this setting, as it does at every setting on these controls, because it is computed inside the simulator.
| prior | gap 0 km | gap 0.5 km | gap 1.5 km | gap 3 km | gap 5 km | SBC |
|---|---|---|---|---|---|---|
| narrow | 72% | 66% | 45% | 16% | 5% | passes |
| standard (shown) | 81% | 74% | 51% | 19% | 5% | passes |
| broad | 91% | 84% | 56% | 17% | 0% | passes |
Simulation-based calibration is the right diagnostic and it is not the diagnostic most readers assume. The logic is sound: draw parameters from the prior, simulate data, infer the posterior, and check that the true value ranks uniformly among posterior samples. A systematically overconfident estimator shows up as ranks piling at the extremes, an underconfident one as ranks piling in the middle [4]. No inversion should ship without it.
Now notice where those test pairs come from. They are drawn from the same prior and pushed through the same simulator the estimator was trained on. SBC can therefore ask one question only: is the estimator faithful to the world it was trained on?
Open a simulation-to-reality gap in the exhibit, the kind a simulator creates by omitting attenuation, fine layering or 3-D structure, and watch the two readouts separate. The plain of stations lifts, coverage against real structure falls, and SBC does not move. It cannot: training and test share the shift, so it is never sampled. Narrow the prior until a quarter of the network falls outside it and SBC still does not move, because those stations are never drawn in the first place.
The numbers make it concrete. Well specified, with a broad prior and no gap, the network delivers 91% coverage against the 90% an honest interval claims, which is what correct looks like. At a standard prior with a 1.5 km gap it delivers 51%. The intervals still say 90. SBC still passes.
Gating out-of-distribution stations is the only control here that recovers anything, and it is worth being precise about what it does: it removes stations from the reported set rather than fixing them. With a narrow prior it buys back most of the loss by declining to answer for a quarter of the network. That is the honest trade and it should be stated in those terms, not presented as improved accuracy.
Three failure modes, named plainly
An asset team should hold us to all three.
The simulation-to-reality gap is the first, and the exhibit is a picture of it: the estimator learns the posterior for the simulator's world, and SBC on simulated data will not catch the bias. Closing it requires validating against independent data, and embedding more real physics in the forward model rather than more capacity in the flow.
Prior misspecification is the second. Too narrow and the posteriors are falsely tight, which the exhibit shows as a coverage loss with no warning attached. Too broad and capacity is spread across structures the region does not have, diluting resolution.
Out-of-distribution input is the third. A real observation unlike anything in training gets an extrapolated, confident and meaningless posterior unless a detector flags it. None of these disqualifies the method. They are the price of amortization, and they are manageable with broad priors, embedded physics, gating and validation against whatever ground truth exists.
Where it sits
Amortized inference is the engine that ties the rest of the stack together. It consumes a forward operator, ideally a neural-operator surrogate so that simulating millions of training pairs is cheap. It produces calibrated H-kappa posteriors as a special case, and its per-station posteriors feed continental-scale mapping as calibrated inputs rather than as picks with habit-formed error bars.
What it does not do is remove the obligation to check against the world. It relocates that obligation, from the inner loop where SBC lives to the outer one where independent data does.
Key takeaways
- Crustal receiver-function inversion has an intractable likelihood and a tractable simulator, which is the regime simulation-based inference is built for.
- A normalizing flow represents the multimodal, correlated posteriors crustal inversion actually has, which a point plus a symmetric error bar cannot.
- Amortization converts a per-station compute bill into a fixed offline cost plus one forward pass per station, which is what makes network-scale Bayesian inversion deployable.
- Simulation-based calibration draws its test pairs from the same prior and simulator the estimator was trained on, so it can only check faithfulness to that world.
- In the exhibit, coverage falls from 91% when well specified to 51% at a standard prior with a 1.5 km simulation-to-reality gap, while SBC passes at every setting.
- An out-of-distribution gate is the only control that recovers coverage, and it does so by declining to answer for some stations rather than by improving any posterior.
References
- K. Cranmer, J. Brehmer, G. Louppe. The frontier of simulation-based inference. PNAS, 117, 30055 to 30062, 2020. doi:10.1073/pnas.1912789117
- G. Papamakarios, I. Murray. Fast epsilon-free inference of simulation models with Bayesian conditional density estimation. NeurIPS 2016. arXiv:1605.06376
- D. Greenberg, M. Nonnenmacher, J. Macke. Automatic posterior transformation for likelihood-free inference. ICML 2019. arXiv:1905.07488
- S. Talts, M. Betancourt, D. Simpson, A. Vehtari, A. Gelman. Validating Bayesian inference algorithms with simulation-based calibration. 2018. arXiv:1804.06788
- G. Papamakarios, E. Nalisnick, D. J. Rezende, S. Mohamed, B. Lakshminarayanan. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research, 22, 1 to 64, 2021. jmlr.org
- The companion piece on what a posterior's shape shows that a stack peak cannot: Simulation-based inference for crustal structure.
