{"id":"8c143d0c-d69a-4f8d-9d67-8aaf2ac05335","arxiv_id":"2412.03299","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Gaussian-process fusion of two synthetic velocity models, propagated through acoustic wave simulations, produces a wider distribution of peak ground displacement than either model alone.","lead":"The paper fuses two synthetic 1-D seismic velocity models with a Gaussian process, then draws many possible velocity profiles from the fit and simulates earthquake shaking through each one. It shows that the spread of peak ground displacements from these samples is much wider than what either input model alone predicts, pointing to a missing uncertainty source in seismic hazard analysis.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GP fusion assumes m1 and m2 are conditionally independent observations of one latent velocity, but the synthetic construction makes them correlated; the reported PGD spread is therefore not validated as the true model-disagreement uncertainty.","rationale":"The reader's weakest assumption correctly identifies the conditional-independence assumption as the most load-bearing point, and the paper's own construction m2 = (2/3)s1 + (1/3)s2 makes the violation explicit. I agree that this is the key issue: the central claim is not merely that some distribution over velocity models yields a wider PGD histogram, but that this distribution represents the uncertainty arising from disagreement between velocity models. If the concatenated-data GP with a diagonal likelihood is misspecified for the two input profiles, then the predictive variance is not a reliable measure of that disagreement, and the PGD width in Fig. 2g has no clear probabilistic interpretation. I do not think this requires changing the reader's CONDITIONAL verdict: the paper is a proof-of-concept, the code is deposited, the limitations are partly acknowledged, and the qualitative direction of the claim is plausible. But the demonstration would have to be re-run against a correctly specified generative posterior or with independent-error synthetic tests before the quantitative width or the general method is validated. The proposed test settles exactly this by comparing the paper's PGD distribution to the one implied by the true generative process.","tokens_in":9617,"tokens_out":8302,"duration_ms":94227,"concrete_test":"Exploit the known generative process to form a validation distribution: condition the original independent GP priors on s1 and s2 on the observed data (m1 = s1 at X1 and m2 = (2/3)s1 + (1/3)s2 at X2), draw, say, 500 posterior samples of the latent profiles, propagate them through the same finite-difference acoustic solver, and compare the resulting PGD median and middle-70% interval to those in Fig. 2g. If the two distributions differ by more than the Monte Carlo standard error of the 200-sample estimate, the reported spread is an artifact of treating correlated models as independent observations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the identification of the PPGPR predictive distribution with 'uncertainty due to model disagreement.' In Section 3, m1 and m2 are concatenated and used to condition a single latent GP with a diagonal Gaussian likelihood. That likelihood says m1 and m2 are conditionally independent noisy evaluations of one underlying velocity profile. But the synthetic construction m2 = (2/3)s1 + (1/3)s2 with m1 = s1 makes the two profiles share the random function s1; no choice of latent f and independent noises can represent this (if f = s1, the error in m2 is -f/3 + s2/3 and is correlated with f; any other f leaves m1 with nonzero error). The mismatch is not a harmless idealization: it is exactly what controls how the GP converts the two input profiles into predictive variance. The paper's only quantitative check, the RMSE against N((m1+m2)/2, ((m1-m2)/2)^2), is a descriptive two-point heuristic and does not validate the variance as calibrated uncertainty, and the heuristic itself does not use the true generating process. The Limitations section flags missing input uncertainties but not this conditional-independence mismatch. Consequently, the PGD histogram in Fig. 2g is a functional of a misspecified posterior, and the headline claim that physics-based hazard predictions understate ground-motion spread is not established by this synthetic experiment. The qualitative conclusion may survive, but the 'probabilistic' width is not shown to be the right width.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a proof-of-concept workflow for propagating uncertainty in seismic velocity models into probabilistic earthquake ground-motion predictions. Two synthetic 1-D velocity profiles, m1 and m2, are fused with a sparse variational Gaussian process (PPGPR) so that disagreements between the models appear as predictive variance. Two hundred velocity samples are drawn from the GP predictive distribution and propagated through a 1-D acoustic finite-difference wave solver; the resulting peak ground displacement (PGD) histogram is compared with simulations using only m1 or m2. The paper concludes that the probabilistic fusion yields a much wider PGD distribution than the two individual velocity models, motivating probabilistic treatment of velocity-model uncertainty in physics-based hazard analysis.","tokens_in":9931,"tokens_out":6168,"duration_ms":67804,"significance":"If the result is validated, the workflow addresses a real gap in physics-based seismic hazard analysis: the arbitrary choice of velocity model is currently not treated as an uncertainty source, and this paper shows a practical way to turn model disagreement into a distribution of ground-motion outcomes. The paper's strengths are its controlled synthetic setup, use of a standard wave-propagation solver, a clear two-model benchmark, and released code, which makes the experiment reproducible. The significance is at the proof-of-concept level: the method is not yet calibrated or compared with existing fusion approaches, and the synthetic example is too small to demonstrate scalability in practice.","major_comments":[{"comment":"The synthetic construction violates the conditional-independence assumption of the likelihood used in the fusion. With m1 = s1 and m2 = (2/3)s1 + (1/3)s2, no latent function f and independent noises ε1, ε2 can represent both profiles as y_i = f + ε_i: if f = s1, then ε2 = (s2 - s1)/3 is correlated with f; if f is any other function, then ε1 and ε2 are correlated through their shared dependence on s1 and s2. Since the diagonal Gaussian likelihood in Section 3 is exactly what converts the two input profiles into predictive variance, the width of the posterior, and hence of the PGD histogram in Fig. 2g, is not established as the correct model-disagreement uncertainty. This issue is not flagged in the Limitations section. Please re-run the synthetic experiment with profiles generated as independent noisy observations of one common latent profile, or explicitly model the correlation between the two models.","section":"Section 3"},{"comment":"The reported validation against N((m1+m2)/2, ((m1-m2)/2)^2) is a per-location two-point heuristic, not a test of probabilistic calibration. Low RMSE of the PPGPR mean and variance relative to this heuristic shows that the predictive distribution tracks the local difference between the two profiles, but it does not establish that the predictive intervals have correct coverage for the true velocity or for the resulting PGD. A coverage or probability-integral-transform check on the synthetic data would directly support the 'probabilistic' claim, which is central to the paper's message.","section":"Section 3"},{"comment":"The headline comparison in Fig. 2g contrasts a continuous predictive distribution with two deterministic endpoint simulations. Because the predictive distribution is constructed to have nonzero width, the conclusion that the PGD spread is 'much wider' is partly by construction; what is missing is a quantitative statement of how often the sampled PGD values fall outside the range defined by the two endpoint simulations, and whether the m1 and m2 PGD values are plausible under the predictive distribution. Reporting quantiles, the fraction of samples outside the endpoint range, or a proper scoring rule would make the central claim more informative and less visual.","section":"Section 4"}],"minor_comments":[{"comment":"The abbreviation is introduced as SVGP, but Fig. 1 and some text use 'SVGPR'; please unify the notation throughout.","section":"Section 2 and Fig. 1"},{"comment":"The observational noise variance is denoted σ2_y and then later σ2_obs; please use one symbol consistently.","section":"Section 2"},{"comment":"The paper states m = 20 inducing points with n = 25 data points per profile, so the total dataset n = 50 and the condition m ≪ n is not satisfied; the scalability argument should be stated as a motivation for future large-scale applications rather than a demonstrated property of this experiment.","section":"Section 3"},{"comment":"The Ricker wavelet parameters (central frequency and amplitude) are not specified in the text; since PGD scales with source amplitude, these values should be given in the paper for reproducibility, even if the code is available.","section":"Section 4"},{"comment":"There is a typo: 'libaries' should be 'libraries'.","section":"Software"}],"recommendation":"major_revision","confidential_remarks":"This is a short proof-of-concept paper, and the core workflow is sensible, but the main technical concern is that the synthetic data-generating process is incompatible with the conditional-independence assumption of the fusion model. This is fixable within the scope of the paper by re-running the experiment with an additive-noise construction (m1 = f + ε1, m2 = f + ε2) or by explicitly modeling the correlation. I would not reject, because the proposed pipeline is clearly presented and reproducible, but the revision should include the corrected experiment or an explicit argument that the misspecified posterior is being used only as a sensitivity analysis rather than as a calibrated uncertainty estimate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a well-written proof-of-concept that does exactly what it promises: it fuses two 1-D velocity models with a scalable GP (PPGPR), draws samples, and propagates them through an acoustic wave solver to get a distribution of peak ground displacement. The novelty is in the workflow, not the machinery—PPGPR already exists, and the wave solver is standard. What's genuinely new is treating velocity-model disagreement as an input-dependent variance and carrying that through to ground-motion predictions. The code is deposited, so the numbers are reproducible.\n\nThe paper earns credit for being honest about its limits. The limitations section lists several extensions. But it misses the softest spot: the synthetic construction makes the two input models correlated. m2 is a weighted sum of the same random function s1 that defines m1. The GP likelihood then treats m1 and m2 as conditionally independent noisy observations of one latent velocity. That's exactly what controls the predictive variance, and no choice of latent function and independent noise can represent the true generating process. So the width of the GP posterior—and hence the width of the PGD histogram—is not actually a validated measure of model disagreement. The RMSE check against N((m1+m2)/2, ((m1-m2)/2)^2) is a descriptive two-point heuristic; it doesn't test calibration. And the 200-sample histogram has no sampling uncertainty, so we don't know how stable that spread is.\n\nI don't think this kills the paper. The qualitative point—that simulating only m1 or m2 understates the possible spread of ground motions—probably survives. But the quantitative claim about how wide the distribution should be is not established by this experiment. That's a fixable flaw: use independent synthetic profiles, or better, use real velocity models with known discrepancies and see if the GP spread matches the spread of wavefield solutions. Even a misspecified test with a known truth would help.\n\nWho should read this: anyone working on uncertainty quantification in seismic hazard, or on GP fusion of geophysical models. It's a useful demonstration that the workflow is computationally feasible. It deserves a serious referee—a workshop paper, but one with a real idea and reproducible code. I'd want the conditional-independence issue addressed before it's used as a basis for hazard practice.\n\nRecommendation: send to peer review, with heavy revision expected.","headline":"Useful proof-of-concept for propagating velocity-model disagreement through wave simulations, but the synthetic setup doesn't validate the quantitative uncertainty width.","tokens_in":10508,"tokens_out":2503,"would_cite":false,"duration_ms":24242,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fusing two seismic velocity models with Gaussian processes and simulating waves through the sampled profiles shows that peak ground displacement spreads much wider than any single model predicts.","keywords":["Gaussian process regression","seismic velocity models","earthquake ground motion","peak ground displacement","probabilistic fusion","uncertainty quantification","acoustic wave equation","physics-based seismic hazard"],"falsifier":"Replace the second synthetic model with a copy of the first; if the workflow is calibrated, the predictive variance should collapse and the PGD distribution should concentrate on a single value. If substantial predictive variance or a wide PGD spread remains, the uncertainty attributed to model disagreement is not actually measuring disagreement.","tokens_in":9410,"feed_emoji":"🌍","tokens_out":6940,"duration_ms":59245,"temperature":0.7,"pith_summary":"Seismic hazard estimates depend on which velocity model is chosen, and current physics-based workflows ignore disagreement among models. This paper proposes treating two or more velocity models as noisy observations of one true profile, fusing them with scalable Gaussian-process regression, and then sampling velocity profiles to drive acoustic wave-propagation simulations. The central finding is that the distribution of peak ground displacement across these samples is much wider than the spread implied by the two original models alone. The paper argues that probabilistic fusion is therefore necessary to capture the uncertainty that model disagreement adds to physics-based hazard analysis.","feed_headline":"GP fusion widens earthquake ground-motion uncertainty","feed_subtitle":"A 1-D proof-of-concept finds peak ground displacement spreads far wider than any single velocity model predicts.","key_machinery":"The central object is the parametric predictive Gaussian process regressor (PPGPR), a sparse variational GP variant whose objective directly targets the predictive distribution $q(y_\\ast)$ rather than the marginal likelihood. This yields an input-dependent variance $\\sigma_f(x_\\ast)^2$ over the underlying velocity function, which the paper interprets as uncertainty due to disagreement between velocity models. The covariance structure of the GP then allows spatially coherent velocity samples to be drawn from the predictive distribution. Each sample is passed through a finite-difference solver for the 1-D acoustic wave equation with a free-surface boundary condition at the surface and a perfectly matched layer at depth, and the peak surface displacement is recorded.","core_discovery":"The paper demonstrates that fitting a parametric predictive Gaussian process regressor (PPGPR) to two synthetic 1-D velocity profiles simultaneously yields a predictive distribution whose uncertainty tracks the local differences between the models. This contrasts with standard sparse variational Gaussian processes, which absorb disagreement into a single observational noise term and produce noisy, spatially incoherent samples. Drawing 200 samples from the PPGPR predictive distribution and simulating the acoustic wave equation with a Ricker source produces a histogram of peak ground displacement whose median and middle 70% span far more than the PGD values obtained from the two input models alone. The paper concludes that using only the input models under-represents possible ground motions, and that the workflow is a proof-of-concept for probabilistic physics-based hazard analysis.","pith_inferences":["A natural calibration test would run the same fusion on many synthetic model pairs and check whether the GP's predictive intervals match the true spread of PGD, rather than relying on a single realization.","If real velocity models carry correlated errors, the diagonal Gaussian likelihood used here will likely understate uncertainty; encoding cross-model correlation in the process covariance would be the next step.","Because the input models' own uncertainties are currently ignored, incorporating them as heteroscedastic observation noise would presumably widen the final PGD distribution further."],"forward_implications":["Physics-based hazard analyses that rely on a single velocity model will systematically understate the range of possible peak ground displacements, because the two input models' PGD values fall well inside the middle 70% of the distribution produced from GP samples.","The GP fusion step accepts any number of overlapping velocity models, so adding more input models does not change the workflow, although unequal data density would require weighting.","Because the predictive samples carry spatial covariance, the velocity profiles are plausible structures rather than pointwise noise, which is what makes the downstream wave simulations physically meaningful.","For real hazard assessment the workflow must be extended from 1-D synthetic profiles to 2-D and 3-D elastic media, as the paper itself notes, so the direct consequences here are limited to the proof-of-concept setting."],"supporting_citations":[{"why":"Supplies the Gaussian-process regression framework that the fusion workflow builds on.","marker":"[16]"},{"why":"Provides the sparse variational GP formulation whose predictive variance PPGPR modifies.","marker":"[18]"},{"why":"Introduces the PPGPR objective that targets the predictive distribution and yields input-dependent variance.","marker":"[20]"},{"why":"Gives the ideal maximum-likelihood two-observation Gaussian used as the reference for the predictive mean and variance.","marker":"[25]"},{"why":"Supplies the perfectly matched layer boundary condition used in the acoustic wave-propagation simulations.","marker":"[26]"}],"fun_headline_variants":["GP fusion reveals wider quake ground-motion spread","Earthquake shaking: GP fusion expands uncertainty set","1-D GP proof: velocity model mismatch widens ground motion","Probabilistic GP shakes up ground-motion estimation beyond single models","Ground-motion uncertainty grows when GPs fuse velocity models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two input velocity models can be treated as conditionally independent, zero-mean noisy observations of a single true velocity profile; the synthetic construction makes them correlated, and real velocity models likely share systematic biases, so if this premise fails the predictive variance can no longer be read as uncertainty due to model disagreement.","fun_headline_variants_meta":{"raw":{"variants":["GP fusion reveals wider quake ground-motion spread","Earthquake shaking: GP fusion expands uncertainty set","1-D GP proof: velocity model mismatch widens ground motion","Probabilistic GP shakes up ground-motion estimation beyond single models","Ground-motion uncertainty grows when GPs fuse velocity models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001085,"raw_usage":{"total_tokens":4508,"prompt_tokens":891,"completion_tokens":3617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":3536}},"tokens_in":507,"tokens_out":3617,"duration_ms":23696,"temperature":1.0,"reasoning_tokens":3536,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:33:22.493349+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the second synthetic model with a copy of the first; if the workflow is calibrated, the predictive variance should collapse and the PGD distribution should concentrate on a single value. If substantial predictive variance or a wide PGD spread remains, the uncertainty attributed to model disagreement is not actually measuring disagreement.","supporting_citations":[{"cited_title":"Scivier, Tarje Nissen-Meyer, Paula Koelemeijer, and Atılım Güne¸ s Baydin","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian-process regression framework that the fusion workflow builds on."},{"cited_title":"When Gaussian Process Meets Big Data: A Review of Scalable GPs.IEEE Transactions on Neural Networks and Learning Systems, 31(11):4405–4423,","cited_arxiv_id":null,"evidence_quote":"Provides the sparse variational GP formulation whose predictive variance PPGPR modifies."},{"cited_title":"Variational Learning of Inducing Variables in Sparse Gaussian Processes","cited_arxiv_id":null,"evidence_quote":"Introduces the PPGPR objective that targets the predictive distribution and yields input-dependent variance."},{"cited_title":"Large-scale Heteroscedastic Regression via Gaussian Process","cited_arxiv_id":"1811.01179","evidence_quote":"Gives the ideal maximum-likelihood two-observation Gaussian used as the reference for the predictive mean and variance."},{"cited_title":"A reflectionless discrete perfectly matched layer","cited_arxiv_id":null,"evidence_quote":"Supplies the perfectly matched layer boundary condition used in the acoustic wave-propagation simulations."}],"review_version":1}