{"id":"f6261210-0aae-442a-870b-fc57cad52049","arxiv_id":"2506.08560","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"For the Endurance CO2 storage site, adding 2-3 ocean-bottom seismometers yields the largest gain in locating small induced earthquakes, while land stations mainly lower the detection magnitude threshold.","lead":"This study uses Bayesian experimental design to compare ways to improve earthquake monitoring at the UK's offshore CO2 storage sites. It finds that ocean-bottom seismometers add the most location accuracy, while extra onshore stations mainly help detect smaller events.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The OBS-over-land conclusion hinges on the distance-proportional velocity uncertainty of Eq.","rationale":"The reader's weakest_assumption correctly flags both the picking-uncertainty model and the velocity-uncertainty model as jointly controlling the central distinction between 'detectable but poorly located' land events and 'well located' OBS events. My stress-test narrows this to the velocity-uncertainty term as the single most load-bearing element: Eq. 8 makes the uncertainty grow with travel time, which is precisely what gives nearby OBS their location-accuracy advantage. The paper's own sensitivity analysis (Figure 12) shows that when zeta_vel is small, a distant station performs nearly as well as an OBS, and the Discussion concedes that the regional CRUST1.0 estimate may be conservative for sites with local imaging. Because SCS sites are typically characterized in detail before injection, the relevant zeta_vel could be substantially smaller than 0.083. If so, the qualitative claim that OBS are 'most promising' would need qualification, and the quantitative station-count recommendations would lose support. This does not reject the paper; the Bayesian machinery is standard and the qualitative geometry-based argument for OBS is plausible. But the central claim is contingent on an uncertain scaling parameter, and the paper would be strengthened by reporting EIG values across a range of zeta_vel and by providing error bars or convergence diagnostics for the reported EIG differences. I therefore agree with the reader's CONDITIONAL verdict and see no reason to change it; the concern reinforces the conditionality rather than overturning the paper.","tokens_in":21915,"tokens_out":6172,"duration_ms":78216,"concrete_test":"Recompute the EIG comparisons underlying Figures 11-14 and the optimal OBS networks with sigma_vel^2 = t * zeta_vel^2 for zeta_vel = 0.02 and 0.04, and also with a constant sigma_vel = 0.02 s independent of distance (or with a site-specific tomographic velocity uncertainty if available). If the optimal single onshore station or the Boulby station then matches or exceeds one or two OBS stations in EIG, the central recommendation is not robust to velocity-uncertainty assumptions; if OBS remains dominant across all variants, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that OBS networks are the most promising monitoring strategy is primarily driven by the velocity-uncertainty model in Eq. 8, sigma_vel^2 = t * zeta_vel^2, with zeta_vel = 0.083 fitted from CRUST1.0 models. Because this term is proportional to travel time, distant land stations carry a large timing uncertainty penalty while nearby OBS stations carry almost none. Section 3.3.1 confirms that velocity-model uncertainty dominates the total uncertainty for detected events, and Figure 12 shows explicitly that as zeta_vel is reduced, the EIG of a distant low-noise station approaches that of a single OBS. The authors also acknowledge in the Discussion that local subsurface imaging is often available at SCS sites, so the regional CRUST1.0-derived value may be conservative. If the actual velocity uncertainty at Endurance is materially smaller, the relative benefit of OBS over an optimally placed onshore station or a low-noise Boulby station could shrink or disappear, which would undermine the station-count recommendations (2-3 OBS for Endurance, 3-5 for all licences). The picking-uncertainty model (Eq. 4, K=10) is also hand-tuned, but it mainly shifts detection thresholds for all network types; the velocity-uncertainty model is the one that directly controls the location-accuracy ranking that supports the paper's conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript uses Bayesian experimental design to compare the value of low-noise onshore stations, onshore arrays, an optimally placed land station, and ocean-bottom seismometers (OBS) for monitoring induced seismicity at the Endurance CCS site and neighbouring North Sea licence areas. The data likelihood combines SNR-dependent picking uncertainty with a travel-time-proportional velocity-model uncertainty, and the expected information gain (EIG) about source location is estimated with nested Monte Carlo. The main conclusions are that land-based additions mostly improve detection of small events (M 0–2) without greatly improving location accuracy, while small OBS networks provide substantially better location accuracy, with 2–3 OBS recommended for Endurance and 3–5 OBS for all licence areas. The paper also concludes that velocity-model uncertainty is the key limiter on location precision once events are detected.","tokens_in":22231,"tokens_out":8402,"duration_ms":100397,"significance":"If the results are robust, the paper offers practically useful guidance for monitoring offshore CO2 storage: it quantifies the detection-vs-location distinction for land stations and arrays, and it gives specific OBS station-count recommendations. The methodology is a standard Bayesian OED framework used carefully, with a nested Monte Carlo estimator and convergence tests (Appendix C), an externally calibrated velocity-uncertainty factor derived from CRUST1.0 models, and explicit sensitivity analyses for the maximum frequency and the velocity-uncertainty factor. These strengths make the paper a credible contribution to the CCS monitoring literature, provided the robustness of the headline OBS recommendations to the calibrated noise-model parameters is demonstrated.","major_comments":[{"comment":"The headline station-count recommendations (2–3 OBS for Endurance, 3–5 OBS for all licence areas) are computed only at the nominal velocity-uncertainty factor ζ_vel = 0.083. The paper's own sensitivity analysis in §3.1.1 and §3.3.1 (Figs. 10 and 12) shows that reducing ζ_vel makes a distant low-noise station nearly as informative as a nearby OBS, and the Discussion concedes that local subsurface imaging may make the CRUST1.0-derived value conservative. Because the multi-OBS optimisations are not repeated at smaller ζ_vel, the reader cannot assess whether the recommended OBS counts, and the claim that onshore stations add negligible information beyond one OBS, are robust to the least certain component of the uncertainty model. Please repeat the optimisations at reduced ζ_vel (for example 0.04 and 0), or state the ζ_vel threshold at which the relative ranking of OBS versus onshore designs changes.","section":"§3.3.2, §3.3.3, Figs. 14 and 16"},{"comment":"The picking-uncertainty model is specified with K = 10 and f_max = 30 Hz, but no sensitivity test is reported for K. Because K controls the SNR level at which pick uncertainty rises sharply, and because the relative value of OBS versus land-based low-noise stations depends on how many small (M = 0–2) events are detectable, the configuration rankings could shift if K were changed, for example to the K = 20 value used in the original Aki/Fuggi formulation. Please add a sensitivity analysis for K, or at least discuss quantitatively how reasonable variations in K would affect the relative EIG of the compared designs.","section":"§2.2, Eq. (4)"}],"minor_comments":[{"comment":"Please ensure the typesetting places 2 f_max inside the inverse bracket, i.e. σ²_pick = [2 f_max log2(1 + SNR/K)]^{-2}, to match Figure 4 and the text's statement that lower frequencies increase picking uncertainty; the current layout could be misread as having the opposite frequency dependence.","section":"Eq. (4)"},{"comment":"The truncated Gutenberg-Richter density appears to contain a spurious '+M_lower' term; the expression should be exp[-γ(M_l - M_lower)] / (1 - exp[-γ(M_upper - M_lower)]) for M_l in [M_lower, M_upper].","section":"Eq. (1)"},{"comment":"The y-axis label reads 'noise level at Boulby site' but the caption and text refer to an OBS station; this should be corrected.","section":"Figure 12"},{"comment":"The first sentence says 'This section supplements section D' but the appendix supplements Section 3.2; please correct the cross-reference.","section":"Appendix D"},{"comment":"The Bayesian optimisation setup (acquisition function, number of iterations, search bounds, and treatment of the existing onshore stations) is not described; please add a short paragraph so the optimal designs in Figures 13 and 15 are reproducible.","section":"§3.3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the International Journal of Greenhouse Gas Control well. My main technical concern is the robustness of the quantitative OBS-count recommendations to the calibrated velocity-uncertainty factor ζ_vel; this is a natural extension of the sensitivity analysis the authors already provide for single stations, so I have framed it as a major revision rather than a rejection. I would also encourage the editor to ask the authors to make the implementation details of the Bayesian optimisation available, as the current description is too terse for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, useful paper for anyone planning seismic monitoring for UK North Sea CCS. The headline numbers (2–3 OBS for Endurance, 3–5 for the wider licences) are credible as a first quantitative pass, but they are model outputs, not stable facts. The main reason is the velocity-uncertainty term, sigma_vel^2 = t*zeta^2 with zeta fitted from CRUST1.0. That term penalises distant land stations almost exactly in proportion to their distance, so OBS wins largely because the model says distant stations carry large velocity-error penalties. The authors know this: Figure 12 shows that when zeta shrinks, a low-noise onshore station's EIG approaches that of the OBS, and the Discussion admits the regional zeta may be conservative when local tomography exists. The qualitative recommendation—nearby OBS improves both detection and location accuracy—survives even if zeta is smaller, because proximity also helps SNR and azimuthal coverage. But the station-count numbers should not be treated as precise.\n\nNew content is real. The same group published the Bayesian OED machinery before, but the SNR-dependent detectability and the distance-scaled velocity uncertainty are new, and this is the first time optimised OBS numbers are given for these specific UK licences. The EIG estimator is standard nested MC, and the convergence tests in Appendix C are appropriate.\n\nSoft spots, in order of importance. (1) Equation (4) as printed does not match the Shannon-Hartley formula shown in Figure 4 and described in the text; one of them has the reciprocal behaviour. The plot and the text agree with each other, so it is likely a typesetting error in the equation, but it must be fixed before review. (2) The picking model uses K=10 and fmax=30 as hand-picked values; that is acknowledged, but it directly affects the detectable-magnitude thresholds, so the absolute EIG comparisons between magnitude ranges are approximate. (3) No code or data are provided, and EIG values have no error bars. The convergence tests show relative comparisons are stable, but the reader cannot check the station-count differences.\n\nCitation pattern looks fair; they cite the relevant Kettlety, Jerkins, and their own previous work without overselling. Serious thinker: yes.\n\nWho is this for? CCS operators, regulators, and network planners who need a defensible starting point for choosing between land arrays and OBS. It deserves a serious referee. My recommendation: send it to review, require the Eq. (4) fix, a short sensitivity analysis for zeta_vel at Endurance (e.g., using a local 1D model), and ask for code/data release or at least the parameter values and station noise inputs.","headline":"A useful applied Bayesian OED paper for UK offshore CCS monitoring, but the OBS-over-land ranking is more sensitive to the assumed velocity-uncertainty model than the paper lets on—though the authors do flag this.","tokens_in":22718,"tokens_out":5503,"would_cite":true,"duration_ms":63674,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62K05","62F15","86A15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Of the tested monitoring upgrades, only seabed seismometers substantially improve where small induced earthquakes at offshore CO2 storage sites are located: 2-3 sensors cover Endurance, 3-5 all North Sea licences.","keywords":["induced seismicity monitoring","carbon capture and storage","Bayesian experimental design","expected information gain","ocean-bottom seismometers","microseismic monitoring","North Sea","seismic network optimisation"],"falsifier":"Deploy a network of two to three ocean-bottom seismometers plus the existing land stations at a North Sea storage site, record calibration sources with known locations such as airgun shots, and compare the mislocation errors: the paper predicts the seabed network should roughly halve location uncertainty for magnitude 1-2 events while an additional land station would barely change it. Alternatively, re-run the analysis with picking-error statistics measured from modern automated phase pickers on real local data, or with the velocity-uncertainty factor replaced by local tomographic estimates; if magnitude 0-2 events then locate as well from land as from the seabed, the central ranking would be overturned.","tokens_in":21714,"feed_emoji":"🌊","tokens_out":23274,"duration_ms":227845,"temperature":0.7,"pith_summary":"The paper asks how the UK should upgrade its earthquake monitoring so that small induced earthquakes near offshore carbon dioxide storage sites in the North Sea can be detected and, more importantly, located. Using Bayesian experimental design, it compares four candidate additions to the existing land-based British network: a deep low-noise station in Boulby mine, the North York Moors seismic array, an optimally placed extra onshore station, and ocean-bottom seismometers (OBS). Its central claim is that land-based additions mainly lower the detection threshold for magnitude 0-2 events while barely improving location accuracy, whereas a small OBS network sharpens locations substantially because the sensors sit close to the expected events. If the paper is right, operators should direct monitoring money at small seabed deployments rather than further onshore stations; the bottleneck that remains is uncertainty in the seismic velocity model, not station coverage.","feed_headline":"Two or three seabed sensors beat extra land stations for CO2 quakes","feed_subtitle":"The UK land network detects, but cannot pin down, small induced quakes; a few seabed stations change that.","key_machinery":"The engine of the comparison is the expected information gain (EIG) of Bayesian experimental design: the expected reduction in Shannon entropy of the event-location posterior, estimated by nested Monte Carlo with 10,000 outer samples and 500 likelihood samples. Two modelling choices carry the argument. Picking uncertainty follows the Shannon-Hartley theorem, $\\sigma_{\\mathrm{pick}}^2 = \\left[2 f_{\\max} \\log_2\\left(1 + \\mathrm{SNR}/K\\right)\\right]^{-2}$, with $K=10$ calibrated to empirical pick-error studies and $f_{\\max}=30$ Hz; this term makes an event effectively undetectable as the signal-to-noise ratio approaches one and nearly constant above $\\mathrm{SNR}=5$. Velocity-model uncertainty is modelled as $\\sigma_{\\mathrm{vel}}^2 = t\\, \\zeta_{\\mathrm{vel}}^2$, scaling the differential travel time $t$ by a factor $\\zeta_{\\mathrm{vel}}=0.083$ fitted to 25 CRUST1.0 crustal models beneath the North Sea, making velocity error the dominant uncertainty wherever events are detectable. Signal amplitudes come from published magnitude-distance relations for P- and S-waves, which is why a seabed sensor's proximity to the expected events raises SNR and thereby both detectability and location precision; the EIG values are converted into an approximate posterior standard deviation in kilometres by assuming an isotropic Gaussian posterior.","core_discovery":"Stated on the paper's own terms, the result is a ranking of monitoring strategies for the Endurance site and the wider UK North Sea. The existing land network already detects and adequately locates magnitude-2 and larger events, so the gap is the magnitude 0-2 range, where events are either missed or seen but cannot be pinned down. A low-noise mine station such as Boulby, or an onshore array such as North York Moors, primarily converts undetectable small events into detectable ones without improving where they occurred, and an optimally located additional land station adds little beyond detection. An ocean-bottom seismometer behaves differently: proximity raises signal amplitudes, so even one OBS at ordinary noise levels improves location uncertainty for magnitude-1 and larger events, and optimised networks of two to three OBS stations deliver most of the achievable information gain for Endurance. Across all licence areas off England's east coast, three to five OBS stations give robust monitoring, with a standalone three-station OBS network performing nearly as well as one combined with the onshore network. Velocity-model uncertainty remains the dominant error source in every configuration, so reducing it, for example by local seismic tomography, is predicted to shrink location uncertainty by a factor of two to three.","pith_inferences":["A cost-conscious operator could read the results as a two-tier strategy: keep the existing land network for magnitude-2-plus events and add two to three OBS for locating smaller induced events, rather than paying for ultra-quiet mine installations or wide seabed arrays; the paper itself only notes that the OBS route is likely more expensive.","The framework amounts to a quantitative forecast that a field deployment could test: with two to three OBS at Endurance, location errors for magnitude 1-2 events should fall toward the predicted posterior standard deviation of roughly 1.7-2 km, whereas the land network alone should scatter them much more widely.","Because the velocity-uncertainty factor is fitted from coarse regional crustal models, sites that already have detailed three-dimensional seismic velocity models may see better land-station performance than this study's conservative estimate suggests, which would narrow the OBS advantage; the paper acknowledges the factor may be conservative where local imaging exists.","The same machinery could be re-run as a cost-benefit optimisation: instead of asking which network gives most information, ask which network maximises information per pound, or whether near-coastal or mixed onshore-offshore stations reproduce OBS performance at lower cost, an option the paper flags because several optimal OBS locations lie close to the coast."],"forward_implications":["Events of magnitude 2 and above near Endurance are already detected and located adequately by the existing UK network, so the meaningful monitoring gap is the magnitude 0-2 range.","Adding a Boulby-type low-noise station or an onshore array would primarily lengthen the catalogue of small detected events without saying where they occurred, making land additions a detection strategy rather than a location strategy.","An optimised network of two to three OBS stations near Endurance captures most of the achievable location information, with diminishing returns beyond three stations and the onshore network contributing little once two OBS are in place.","Across all UK North Sea carbon storage licence areas off eastern England, three to five OBS stations give robust monitoring, and a standalone three-station OBS network performs nearly as well as a combined onshore-plus-OBS network.","Reducing velocity-model uncertainty, for example through local seismic tomography, is predicted to reduce location uncertainty by a factor of two to three across all configurations."],"supporting_citations":[{"why":"Defines the expected information gain that all network designs are ranked by.","marker":"Lindley (1956)"},{"why":"Supplies the nested Monte Carlo EIG estimator with reused inner samples that makes the design calculations tractable.","marker":"Huan and Marzouk (2013)"},{"why":"The source of the Shannon-Hartley picking-uncertainty model that links signal-to-noise ratio to arrival-time pick error.","marker":"Fuggi et al. (2024)"},{"why":"Origin of the signal-to-noise based picking-uncertainty formula that Equation 4 adapts.","marker":"Aki (1976)"},{"why":"Empirical pick-error estimates used to calibrate the constant K=10 in the picking-uncertainty formula.","marker":"Zelt and Forsyth (1994)"},{"why":"Provides the UK P-wave magnitude-amplitude relation used for signal strength and the 30 Hz maximum frequency.","marker":"Green et al. (2020)"},{"why":"Provides the local-magnitude scale for S-wave amplitudes used to convert magnitudes to signal strength.","marker":"Luckett et al. (2019)"},{"why":"CRUST1.0 crustal models used to fit the velocity-model uncertainty factor of 0.083.","marker":"Laske et al. (2012)"},{"why":"Open-source ray-tracing code that computes the differential P- and S-wave travel times forming the likelihood.","marker":"Heimann et al. (2017)"},{"why":"Defines the new low and high noise model curves that set reference levels for judging which station noise levels are achievable.","marker":"Peterson (1993)"}],"fun_headline_variants":["Seabed sensors beat extra land stations for pinpointing CO2 quakes","A few OBS, not land arrays, sharpen location of small induced quakes","Two to three seabed stations best for monitoring CO2 storage quakes","For CO2 storage offshore, seabed sensors improve quake location more"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything hinges on two assumed error models, neither measured directly at the site: how sharply arrival-time picking accuracy degrades as the signal-to-noise ratio falls (a Shannon-Hartley formula with hand-chosen $K=10$ and $f_{\\max}=30$ Hz) and how large the regional velocity-model error is taken to be (a factor $\\zeta_{\\mathrm{vel}}=0.083$ fitted from coarse crustal models and scaled by travel time), because together these determine why land stations are judged 'detectable but poorly located' while seabed stations are judged 'well located'.","fun_headline_variants_meta":{"raw":{"variants":["Seabed sensors beat extra land stations for pinpointing CO2 quakes","A few OBS, not land arrays, sharpen location of small induced quakes","Two to three seabed stations best for monitoring CO2 storage quakes","For CO2 storage offshore, seabed sensors improve quake location more"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1687,"prompt_tokens":1110,"completion_tokens":577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":726,"completion_tokens_details":{"reasoning_tokens":495}},"tokens_in":726,"tokens_out":577,"duration_ms":6735,"temperature":1.0,"reasoning_tokens":495,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:08:07.225209+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy a network of two to three ocean-bottom seismometers plus the existing land stations at a North Sea storage site, record calibration sources with known locations such as airgun shots, and compare the mislocation errors: the paper predicts the seabed network should roughly halve location uncertainty for magnitude 1-2 events while an additional land station would barely change it. Alternatively, re-run the analysis with picking-error statistics measured from modern automated phase pickers on real local data, or with the velocity-uncertainty factor replaced by local tomographic estimates; if magnitude 0-2 events then locate as well from land as from the seabed, the central ranking would be overturned.","supporting_citations":[],"review_version":1}