Deep seismic sounding maps the Earth's crust by setting off explosions and recording the waves at hundreds of seismometers laid out along a line. A geophysicist then marks, on each recorded trace, the moment a particular wave arrives. These marks are called "picks", and they are made by hand for weak, noisy phases such as the reflection off the crust–mantle boundary (the Moho). The arrival times are then inverted for a velocity-versus-depth model, and that model carries the geological conclusions. Everyone in the field knows picking is subjective. In practice, though, the subjectivity is reduced to one constant error bar per wave type, and the published velocity model is shown as if a single competent picker had produced it. Medical imaging has spent a decade learning that experts who disagree on the same image are a signal to be modelled, not noise to be averaged away.
We propose to bring that lesson into seismic traveltime inversion. First, we collect independent picks from many geophysicists on the same gathers. Second, we train a picker that reproduces the full distribution of human picks, including each picker's systematic style. Third, we push samples from that distribution through the inversion to obtain an ensemble of velocity models. The ensemble shows which interpreted structures survive realistic human disagreement. The central bet is that human pick disagreement is structured: it is correlated along neighbouring traces, it grows with distance from the source in a phase-dependent way, and it includes per-picker offsets. If so, the constant-error-bar convention misstates uncertainty exactly where conclusions are drawn, for example a thin low-velocity layer in the mid-crust or the depth of the Moho. If the bet is wrong, the experiments will show it cheaply, and that is itself a useful negative result for the field.
Core Claim
Modelling the joint distribution of human picks, rather than assigning independent Gaussian errors with a fixed spread, will change the inferred uncertainty of specific, published-style crustal features. The change will be large enough to flip some yes/no structural conclusions. The learned picker makes this tractable without re-recruiting a panel for every new survey.
The concrete test case is the class of result reported for GeoPRISMS Line 1. That 220 km active-source profile with 708 seismographs resolved crustal P-wave velocities of 5.7–6.9 km/s and a 2-km-thick low-velocity zone at 10–12 km depth beneath the eastern Piedmont. A feature that thin, detected from traveltime data, is precisely the kind whose existence could depend on whose picks were used.
Relation to Prior Work
The closest existing bridge between the two directions is Learned Full Waveform Inversion Incorporating Task Information for Ultrasound Computed Tomography. It trains a network to map acoustic recordings directly to sound-speed maps and adds a loss from a single pretrained segmentation network acting as an automated "observer". It motivates our work by showing that what the observer cares about should shape the inversion. It is different from our work in two ways. Its observer is one deterministic network, with no human disagreement anywhere in the pipeline. Its human-facing task sits after the inversion, whereas in deep seismic sounding the human judgement sits before the inversion, in the picks.
OpenFWI supplies synthetic benchmarks and reports that learned inversions are most uncertain at layer interfaces and faults. It treats simulator velocity maps as absolute truth and never asks how human interpreters vary. We reuse its simulation setting for a controlled study, and we ask whether network uncertainty and human uncertainty concentrate in the same places.
From medical imaging we reuse methods rather than conclusions. The Latent Doctor Model is a conditional variational autoencoder, a generative model that maps an input plus a random latent code to a label. In that model a set-encoder infers each rater's latent "style" from examples of their labels. It reproduced pathologist panels, with a KL divergence of 0.461 on Gleason2019 and a Spearman correlation of 0.728 between predicted and empirical disagreement on ISUP grading. We adapt its structure to sequences of arrival times along a gather.
QUBIQ both constrains and motivates the design. It showed that plain ensembles of deterministic networks beat variational models at reproducing multi-rater segmentations, so ensembles must be a baseline and not a strawman.
DiFF weights raters by how much their labels help a downstream diagnosis. It gained 2.04–2.73% AUC on fundus, thyroid and skin benchmarks. We deliberately do not import that principle. In geology the downstream "diagnosis" is the interpretation under test, so weighting pickers by how well they support it would be circular.
Research Program
The work has three parts.
The first part is a multi-picker dataset with two arms. In the field arm, 8–12 geophysicists with crustal-refraction experience independently pick first arrivals and the wide-angle Moho reflection on a fixed subset of shot gathers from a crustal profile. GeoPRISMS Line 1 is the target if its gathers can be obtained; otherwise we use any comparable open active-source profile. Pickers see the same display tools and the same instructions, and do not see each other's work.
In the controlled arm, we simulate crustal-scale gathers from two velocity models that match the GeoPRISMS velocity ranges: one with a 2-km low-velocity zone at 10–12 km and one without. We add field-recorded noise to both and present them blind, mixed among the field gathers. Because the true arrival times are known here, we can separate a picker's systematic offset from their scatter. That separation is impossible on field data. OpenFWI's shallower acoustic gathers serve as a secondary controlled set and connect our results to existing learned-inversion baselines.
The second part is a learned pick-distribution model. Let $D$ be a shot gather, $t$ the vector of arrival times along its traces, and $z$ a low-dimensional picker-style code inferred from that picker's other gathers. We train $q(t \mid D, z)$ as a conditional variational autoencoder whose decoder outputs arrival times jointly across traces, so that correlated errors along a gather can be represented. At test time we sample $z$ from its learned prior to imitate an unseen picker.
The third part is propagation through the inversion. Let $T(m)$ be the ray-traced traveltimes for a velocity model $m$, and $R(m)$ the usual smoothness penalty. For each sampled pick vector $t_k$ we solve
$$ m_k = \arg\min_m \; \lVert T(m) - t_k \rVert^2 + R(m). $$
The spread of $\{m_k\}$ is the interpretive uncertainty. Traveltime inversion is cheap compared with waveform inversion, so a few hundred members per profile is affordable. We summarise each ensemble by the quantities geologists actually argue about: the fraction of members that need a velocity decrease at 10–12 km, the spread of Moho depth along the profile, and the spread of lower-crustal velocity. We also run the same ensemble over the real human panel's picks, which gives an empirical target for the learned model to match.
Baselines
The naive reference point is current practice. We take one picker's picks, or the median across the panel, assign a fixed uncertainty per phase, and perturb each pick independently with Gaussian noise of that size. We then invert the same number of samples with the same inversion code and regularisation. This baseline is what the claim must beat on field and controlled data.
The strongest simple competitor is a calibrated version of that baseline. The independent noise is given a per-phase, offset-dependent spread fitted to the empirical scatter of the human panel. It therefore matches the marginal variance of human disagreement exactly and differs from our model only in correlation structure and picker style. We give it the same panel data, the same number of inversion runs, and a grid search over its spread parameters on held-out pickers.
The strongest learned competitor is a deep ensemble of deterministic pickers. Each network is trained on a different picker's labels, mirroring the ensembles that won QUBIQ, where top teams reached Dice 0.80–0.85. It gets the same training data and GPU budget as our model. For the pick-distribution task itself on seismic data there is no reported result found on Lacuna for multi-picker arrival-time modelling. The nearest reported figures for distribution matching are the Latent Doctor Model's KL of 0.461 on Gleason2019 and Spearman of 0.728 on ISUP. We will report the same two statistics on held-out pickers.
For the secondary comparison with network uncertainty, the reference is InversionNet on OpenFWI. Its reported SSIM is above 0.99 on FlatVel-A and around 0.60 on CurveFault-B (OpenFWI).
Two ablations isolate the key ingredients. The first removes the style code $z$, leaving a single conditional distribution. The second replaces the decoder's joint output with a per-trace independent output of matched marginal spread.
Our model counts as beating the baselines only if both of the following hold. First, it achieves lower KL and higher disagreement correlation than the ensemble on held-out pickers. Second, in the controlled arm, its velocity ensemble covers the true model at nominal rates where the calibrated independent baseline does not. Credible intervals for the low-velocity-zone indicator and for Moho depth are the specific test.
Why not just…?
Why not just take the median pick across experts and move on? A single consensus pick discards exactly the information the study is about. The Latent Doctor Model showed that consensus training makes models overconfident on ambiguous cases. In our setting the median also hides per-picker offsets that may shift a whole phase branch coherently. The controlled arm settles this directly. If median picks already recover the true model, including the presence or absence of the low-velocity zone, within the baseline error bars, the elaborate model is unnecessary.
Why not just fit a better constant error bar per phase? This is the real threat. The calibrated independent baseline above is that alternative, and it receives the same human data. If disagreement turns out to be close to independent across traces with no stable picker styles, it will perform as well as our model. The ablation removing joint structure will show this, and we will report that outcome as the finding. Our hypothesis predicts specific measurable structure: positive correlation of pick deviations over tens of neighbouring traces, and picker offsets that persist across gathers. Both are checked on the raw panel data before any model is trained.
Why not just use an automatic picker and remove humans entirely? An automatic picker trained on one analyst's labels inherits that analyst's convention while presenting it as objective. Its variance comes from model parameters, not from the plausible range of expert readings. The deep-ensemble baseline tests the strongest version of this idea. Our Spearman comparison against empirical panel disagreement shows whether its spread tracks where humans actually disagree.
Why not skip picking and run full waveform inversion? Waveform inversion gets trapped in wrong solutions unless it starts from a good low-frequency model (OpenFWI). For crustal profiles that starting model usually comes from traveltime picks, so picking subjectivity moves upstream rather than disappearing. Our ensemble is a natural set of starting models for such inversion, and it could later be used to test that dependence.
Decisive Checks
The first check is whether panel disagreement is structured: trace-to-trace correlation of deviations, and picker offsets that are significant against the known truth in the controlled arm. If neither appears, we stop at a measurement paper.
The second check is the blind low-velocity-zone test. On the two controlled crustal models, we ask whether our ensemble assigns high probability to the zone where it exists and low probability where it does not. Separately, we check whether the independent-error baseline is overconfident in either case.
The third check is the field profile. We report the ensemble probability that the 10–12 km low-velocity zone is required, together with the Moho-depth spread, under each noise model. A material disagreement between the noise models on that probability is the result that would change how such profiles are reported.
As a final, secondary analysis, we correlate the spatial spread of the velocity ensemble induced by human picks with InversionNet ensemble variance on OpenFWI-style models. This asks whether machine and human uncertainty concentrate on the same interfaces, a question that OpenFWI raised but did not answer.
References
- Crustal structure of the eastern Piedmont and Atlantic coastal plain in North Carolina and Virginia, eastern North American margin
- Learned Full Waveform Inversion Incorporating Task Information for Ultrasound Computed Tomography
- OpenFWI: Large-scale Multi-structural Benchmark Datasets for Full Waveform Inversion
- The Latent Doctor Model for Modeling Inter-Observer Variability
- QUBIQ: Uncertainty Quantification for Biomedical Image Segmentation Challenge
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle