{"id":"d49e67d5-db01-46f8-9754-bcac7d4375a8","arxiv_id":"2411.11015","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"The authors present a code package and practical guide for Bayesian experimental design of volcano seismic networks, jointly optimizing travel-time, amplitude, and array-based source location.","lead":"This paper provides an open-source Python package that helps volcanologists decide where to place seismometers and seismic arrays around a volcano to best locate earthquakes. It combines travel-time, amplitude, and array data in one Bayesian design workflow and links to global maps so a first network design can be produced in minutes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DN-EIG equations hard-code the data-space dimension to 3, making EIG values and station-count comparisons potentially biased unless the code diverges from the paper.","rationale":"The reader's conditional verdict is appropriate: the paper presents a genuinely useful framework, an open-source package, and clear practical guidance, and the central claims are plausible. My stress-test does not find an independent fatal flaw that would require rejection. However, I identify a different load-bearing weak point than the reader's primary choice. The Gaussian-evidence assumption of the DN method is explicitly acknowledged and mitigated by the NMC fallback and by the recommendation to re-run optimisation with NMC. The equation-dimensionality and sign inconsistency is not acknowledged and is more directly connected to the paper's quantitative claims: the EIG is the objective being optimised, and if the printed equations are taken at face value they cannot all be true. Since the repository is public, the ambiguity is resolvable by inspection and by rerunning the figures. Thus the correct disposition remains CONDITIONAL, pending verification that the code matches a corrected, self-consistent EIG formula. I partially agree with the reader because the reader did flag the equation inconsistency as one of several technical issues, but the reader's stated weakest assumption was the DN Gaussian approximation rather than the algebra of the EIG equations.","tokens_in":549,"tokens_out":6289,"duration_ms":233898,"concrete_test":"Inspect the WoWED-volcano repository to determine whether EIG_DN uses a hard-coded 3/2 or derives k from the actual data dimension. Then re-run the Figure 8 experiment (1-10 receivers) using the implemented formula and compare the EIG curve and optimal designs to the paper. Independently re-derive Equation (6) from Equation (4) using the sign convention in Equation (15); if a consistent derivation yields a different expression, the equations need correction even if the code is correct.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's EIG formulas are internally inconsistent in a way that directly affects the central practical claim of near-real-time design. Equation (7) defines the information of a multivariate Gaussian evidence as k/2 (1+log(2π)) + 1/2 log|C_d|, with k the data dimension. However, Equation (6) replaces k with 3, and Equation (10) again uses 3/2 (1+log(2π)). In the example with three nodal stations plus one array, the data dimension is not 3: each nodal station contributes travel-time and amplitude, and the array contributes travel-time, amplitude and back azimuth, so k=9. In Figure 8 the data dimension changes with the number of receivers, so a constant 3/2 offset does not cancel across designs. If the code implements Equation (6) literally, all quoted EIG values are biased by a dimension-dependent offset, and the reported EIG-versus-N curve and the derived 'approx. standard deviation' in Figures 5, 8 and 10 are unreliable. If the code instead uses the true k, then the equations misrepresent the implementation, undermining reproducibility. There is also a sign-convention inconsistency: Equation (15) defines Shannon information as E[log p], i.e. negative entropy, whereas Equation (7) uses the positive entropy expression. The load-bearing issue is that the design objective itself is not unambiguously specified, and the open-source code is the only way to determine which formula was actually used.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a Bayesian experimental design framework and an accompanying Python package for optimizing seismic monitoring networks at volcanoes, combining travel-time, amplitude, and array back-azimuth observations in a single expected information gain (EIG) objective. The DN approximation is used for rapid design optimization and the NMC method for validation, and the workflow is demonstrated on Mount Etna with comparisons against random and Sobol reference designs, as well as an exploration of homogeneous, layered, and heterogeneous velocity models. The authors claim that this is the first simultaneous optimization of travel-time, amplitude, and array source location methods and that the package enables near-real-time design for any volcano through public databases.","tokens_in":21432,"tokens_out":10556,"duration_ms":90825,"significance":"If the software operates as described, it addresses a genuine operational need: rapid, accessible, and reproducible design of volcano monitoring networks. The integration of multiple data types and the connection to public topographic and volcanic databases are practical strengths. The authors also provide open-source code with a Jupyter notebook, which is commendable. However, the central quantitative results—EIG values, approximate standard deviations, and the receiver-count comparison in Figure 8—depend on information-theoretic equations that are internally inconsistent in the manuscript. The inconsistencies affect the sign and dimension of the evidence term, so the reported numbers cannot be considered reliable unless the code resolves the discrepancies. The paper's practical contribution is therefore potentially valuable but is not yet supported by a correct and unambiguous presentation of the objective function.","major_comments":[{"comment":"Equations (6) and (10) replace the term k/2(1+log(2π)) from Equation (7) with the constant 3/2(1+log(2π)), even though Equation (7) defines k as the dimension of the data space. In the main example (Section 3.0.5), three nodal stations and one array yield data dimension k = 9 (each nodal station contributes a travel time and an amplitude; the array contributes a travel time, an amplitude, and a back azimuth). Because k changes with the number of receivers, the error does not cancel when comparing designs of different sizes, as is done in Figure 8. Please correct the equations to use the actual data-space dimension and state explicitly which expression is implemented in the released code.","section":"2.1.2, Eqs. (6), (7), (10)"},{"comment":"The sign convention for Shannon information is inconsistent. Equation (15) defines I[p] = E_p[log p], which is the negative of the differential entropy, whereas Equation (7) gives the positive entropy of a multivariate Gaussian. Depending on which convention is used, the evidence term in Equations (4) and (6) has the opposite sign. The design objective is therefore not unambiguously specified. Please adopt a single convention, apply it consistently, and confirm that the code implements the corrected form.","section":"2.1.2 and Appendix A, Eqs. (4), (6), (15), (7)"},{"comment":"The travel-time uncertainty formula σt(t)^2 = σp^2 + t·σv^2 is dimensionally inconsistent if σv is a dimensionless relative velocity uncertainty as stated in the text. The example (standard deviation of 0.1–0.2 s for a travel time of 1 s) implies the intended formula is σt^2 = σp^2 + (t·σv)^2. Please correct the equation and verify that the code implements the intended form, since this term enters the likelihood for all travel-time based data.","section":"3.0.4, Eq. (12)"},{"comment":"The derivation of the approximate standard deviation appears to contain an algebraic error. For a three-dimensional isotropic Gaussian posterior with variance σ^2, the entropy is H = (3/2)(1+log(2π)) + 3 log σ. Substituting this into the relation arI_post = EIG - I_prior and the exponential in Equation (14) yields arσ = exp( -arI_post/3 - 1/2(1+log(2π)) ), not arσ^2. Please re-derive Equation (14) and ensure that the values reported in Figures 5, 8, and 10 are correct.","section":"3.0.6, Eq. (14)"}],"minor_comments":[{"comment":"The phrase \"we define an array as seismic array\" should read \"we define an array as a seismic array\".","section":"2.2.3"},{"comment":"The word \"confimation\" should be \"confirmation\".","section":"3.0.6"},{"comment":"The axis simultaneously shows EIG in nats and approximate standard deviation in meters, which makes the relative scales difficult to interpret. Consider using separate panels or a secondary axis.","section":"Figure 8"},{"comment":"The description of Sobol reference designs is somewhat confusing: it would be clearer to state that the Sobol sequence is scaled by a random factor sampled from 0 to 20 km, and to explain why this approximates a \"reasonable, uniformly distributed random design.\"","section":"3.0.6"},{"comment":"The code is available only through a GitHub link. For reproducibility, please also provide a versioned archive (e.g., Zenodo DOI) and state the software license.","section":"7, Data Availability"}],"recommendation":"major_revision","confidential_remarks":"The review was based on the manuscript text only; I did not inspect the GitHub repository. Given that the paper is a software contribution, the inconsistency between the printed EIG equations and the likely implementation is a particularly important issue, since the code is the only way to determine which formula was actually used. If the code uses the correct data-space dimension and sign convention, the manuscript still needs a careful revision of all information-theoretic expressions. I would encourage the authors to re-run the examples after any correction and to make archival of the code a condition of acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The code package is a real contribution: it is the first open-source implementation that optimizes travel-time, amplitude, and array-based source location designs simultaneously, and it plugs into GVP and OpenTopography so a user can start from a volcano name and get a design in minutes. The paper is honest about the DN method's limitations, including the multimodal prior failure it references from prior work, and it provides an NMC fallback. The step-by-step Etna example is clear, and the inclusion of a heterogeneous velocity model section shows the code can be extended.\n\nThe soft spot is the EIG equations. Equation (7) defines k as the data-space dimension, but equations (6) and (10) hard-code 3/2 (1+log 2π). In the Etna example, three nodal stations and one array give k=9, not 3. That matters: the offset does not cancel when comparing designs with different station counts, so the EIG-versus-N curves and the approximate standard deviations in Figures 5, 8 and 10 are not trustworthy as printed. There is also a sign inconsistency: equation (15) defines information as E[log p] (negative entropy), while equation (7) uses the positive entropy expression. The reported EIG values are positive, so the code probably implements the correct formula and the paper misprints it, but the reader cannot tell without reverse-engineering the repo. This is a reproducibility problem, not just a typo.\n\nThe other limitations are less severe. The demonstration is entirely synthetic; the claim that optimized designs beat random or Sobol designs is a consistency check of the optimizer, not an independent test. For a software paper that is acceptable, but it should be labeled as such. The approximate standard deviation relies on an isotropic Gaussian posterior, which they acknowledge. The lack of real-data validation is a missed opportunity, but not a blocker.\n\nThe paper deserves a serious referee. The underlying math is standard, the code is open source, and the practical goal is valuable. The revision should fix the equations, pin a commit, and ideally add a small real-data example. I would not desk-reject this.","headline":"Useful open-source tool, but the EIG equations are internally inconsistent and must be fixed before the quantitative claims are reliable.","tokens_in":21978,"tokens_out":5813,"would_cite":true,"duration_ms":63218,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62K05","62F15","86A15"],"pacs":["91.30.-f"],"model":"deepseek-v4-flash","headline":"A new open-source package lets volcano seismologists compute near-optimal monitoring networks in minutes by jointly optimizing travel-time, amplitude, and array observations.","keywords":["Bayesian experimental design","volcano seismic monitoring","seismometer network optimization","expected information gain","travel-time source location","amplitude source location","seismic array back-azimuth","genetic algorithm"],"falsifier":"Take a volcano with two distinct plausible source regions (multimodal prior), optimize with both the fast Gaussian-evidence method and the slow exact method, and compare the resulting station layouts and their true expected information gain; if the fast layout differs substantially or performs clearly worse, the near-real-time design claim fails for that scenario.","tokens_in":20884,"feed_emoji":"🌋","tokens_out":4371,"duration_ms":42726,"temperature":0.7,"pith_summary":"The paper claims that Bayesian experimental design can be made practical for volcano seismology: a code package lets any seismologist produce an optimal station layout in minutes using generic volcano databases, and refine it within hours. It is the first to optimize travel-time, amplitude, and array back-azimuth observations simultaneously, combining them into a single expected-information-gain objective. This matters because the network geometry determines how well earthquake locations can be inferred, and most existing networks are not optimized at all. Using Mount Etna as an example, the paper shows that optimal designs outperform random and space-filling layouts and quantifies how the number of receivers trades against expected location uncertainty.","feed_headline":"Optimal volcano seismometer layouts in minutes, not days","feed_subtitle":"Open-source package jointly optimizes travel-time, amplitude, and array data for any volcano's earthquake monitoring.","key_machinery":"The engine is the expected information gain (EIG) written in data space as EIG(ξ) = E_{p(m)}[I[p(d|m,ξ)]] − I[p(d|ξ)], where p(m) is the prior over source locations, p(d|m,ξ) is the Gaussian data likelihood, and I is Shannon information. During optimization the EIG is evaluated with the DN method, which approximates the evidence p(d|ξ) as a multivariate Gaussian and uses the log-determinant of the data covariance, while the nested Monte Carlo (NMC) method provides an unbiased but slower estimate for validation. A genetic algorithm searches over a design space built from digital elevation models and user-defined constraints, and the prior is defined on a grid so each cell can be weighted by elevation or other volcano-specific information.","core_discovery":"The central contribution is a unified Bayesian experimental design framework and its open-source implementation, which optimizes a seismic monitoring network for hybrid data types—travel times, amplitudes, and array back-azimuths—at the same time. The paper demonstrates that the fast DN approximation of expected information gain is accurate enough for near-real-time optimization, while the slower NMC method serves to validate the result. It also shows that the optimal design process is relatively robust to the choice of velocity model, and that going to a 3D heterogeneous model yields a modest improvement, mainly for deeper events. By translating the expected information gain into an expected posterior standard deviation, the paper gives practitioners an intuitive number for comparing designs and deciding how many stations they need.","pith_inferences":["The authors' homogeneous-velocity assumption is likely the main hidden cost of the 'minutes' promise; moving to a 3D velocity model shifts computation from minutes to hours, so the speed claim applies most directly to the simplified forward model.","A natural testable extension is to back-test the optimizer on a volcano with an existing seismic catalog, comparing simulated location errors of the recommended network against those of the actual network.","Because the code treats an array as a single instrument measuring back-azimuth, it does not optimize the internal geometry of the array aperture itself, so users needing fine array control would have to extend the design vector beyond what the paper demonstrates."],"forward_implications":["A volcano seismologist with no design expertise can obtain an initial optimal network layout within minutes and refine it within hours, using only public topography and volcano databases.","Optimal layouts consistently beat random and quasi-random space-filling layouts, matching the mean uncertainty of Sobol designs with one fewer receiver.","Optimizing travel-time, amplitude, and array back-azimuth data together produces designs suited to diverse event types, including long-period and tremor signals that lack clear phase arrivals.","The expected location uncertainty can be summarized by a single standard-deviation number, so practitioners can read off how many receivers are needed to reach a desired precision.","The choice of velocity model has modest influence on the optimal layout; a heterogeneous 3D model yields the best performance, mainly by improving resolution of deep events."],"supporting_citations":[{"why":"Supplies the DN method, the fast Gaussian-evidence approximation used as the default EIG estimator during design optimization.","marker":"Coles and Curtis (2011)"},{"why":"Provides the computationally efficient nested Monte Carlo formulation used to estimate EIG without bias.","marker":"Huan and Marzouk (2013)"},{"why":"Introduced the nested Monte Carlo estimator that the paper uses as its accurate validation method.","marker":"Ryan (2003)"},{"why":"Supports the claim that the DN method performs well for seismic monitoring network design despite its upward bias.","marker":"Bloem et al. (2020)"},{"why":"Provides prior evidence on DN method behavior and an example where it fails under multimodal priors, which motivates the validation step.","marker":"Strutz and Curtis (2023)"},{"why":"Supplies the SRTM15+ digital elevation model used to build the design space for any volcano.","marker":"Tozer et al. (2019)"},{"why":"The Global Volcanism Program database gives basic volcano information such as location, type, and eruption history used to set up the prior and design space.","marker":"Venzke (2024)"},{"why":"Provides the amplitude source location forward model and uncertainty treatment that the paper adopts for amplitude data.","marker":"Kumagai et al. (2013)"},{"why":"Supplies the eikonal solver used to extend the forward modeling to layered and heterogeneous velocity models.","marker":"White et al. (2020)"}],"fun_headline_variants":["Volcano seismic design: minutes, not days","Optimize volcano networks with hybrid data","Near-real-time volcano monitoring layout","Fast Bayesian design for volcano seismology","Unified sensor layout optimization for volcanoes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fast design method used by default assumes that the spread of predicted data across all possible earthquake locations is roughly bell-shaped; if a volcano has several well-separated possible source regions, that assumption breaks and the quick designs can be misleading unless the slower, exact method is used.","fun_headline_variants_meta":{"raw":{"variants":["Volcano seismic design: minutes, not days","Optimize volcano networks with hybrid data","Near-real-time volcano monitoring layout","Fast Bayesian design for volcano seismology","Unified sensor layout optimization for volcanoes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000538,"raw_usage":{"total_tokens":2540,"prompt_tokens":861,"completion_tokens":1679,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":1617}},"tokens_in":477,"tokens_out":1679,"duration_ms":12341,"temperature":1.0,"reasoning_tokens":1617,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:00:54.898166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a volcano with two distinct plausible source regions (multimodal prior), optimize with both the fast Gaussian-evidence method and the slow exact method, and compare the resulting station layouts and their true expected information gain; if the fast layout differs substantially or performs clearly worse, the near-real-time design claim fails for that scenario.","supporting_citations":[{"cited_title":"Experimental design for fully nonlinear source location problems: which method should I choose? Geophys","cited_arxiv_id":null,"evidence_quote":"Supports the claim that the DN method performs well for seismic monitoring network design despite its upward bias."},{"cited_title":"T., Smith, W","cited_arxiv_id":null,"evidence_quote":"Supplies the SRTM15+ digital elevation model used to build the design space for any volcano."},{"cited_title":"Global Volcanism Program [Database] Volcanoes of the World (v","cited_arxiv_id":null,"evidence_quote":"The Global Volcanism Program database gives basic volcano information such as location, type, and eruption history used to set up the prior and design space."},{"cited_title":"S., Yamashina, T., Ruiz, M., Palacios, P., Ortiz, H., and Yepes, H","cited_arxiv_id":null,"evidence_quote":"Provides the amplitude source location forward model and uncertainty treatment that the paper adopts for amplitude data."}],"review_version":1}