{"id":"834571a0-1795-4a4e-a9dc-56955713e91f","arxiv_id":"2411.16447","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper calibrates Gehlen's chloride diffusion model using Bayesian optimization and four corrosion wire sensor break times from a test bridge, and compares it with a neural network that is acknowledged to overfit.","lead":"A physics-based chloride diffusion model and a neural network were calibrated against wire sensor corrosion data from a concrete test bridge. The physics-based model gave interpretable, stable results, while the neural network overfitted with only four data points.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'increasingly accurate predictions' claim rests on calibration-fit accuracy, not held-out prediction: Table 4 and Figure 7 use the same wire-sensor points for fitting and validation, and the drilling-dust comparison is too sparse and unquantified to bear the claim.","rationale":"The paper has real virtues: it builds on a standard diffusion model, includes a clean-data sanity check of the Bayesian optimizer (Appendix C.1), performs a Sobol sensitivity analysis, and uses independent drilling-dust data, which is more than many calibration studies do. My concern is not that the method is invalid but that the paper's headline evidence conflates fitting accuracy with predictive accuracy. The progressive-calibration plots are all in-sample; Table 4 shows x_calc matching x_exp because those are the optimization targets. A three-parameter model fit to four points will generally interpolate those points, so the observed 'improvement' with more data is not evidence of better prediction. The one independent comparison, the drilling-dust DEff, is used to assert agreement, but no metric such as a confidence interval or factor-of-two tolerance is given, and the 2024 point appears to be off by roughly a factor of four under the reported four-point parameters. The unstable parameter estimates across calibration runs also signal non-identifiability, which matters because the paper interprets the effective diffusion coefficient physically. These issues are addressable with a leave-one-out analysis and an uncertainty-bounded comparison to the dust data, which is why I keep the reader's CONDITIONAL verdict rather than moving to REJECT or ACCEPT. I partially agree with the reader's weakest-assumption choice: the 24-day offset and the 0.6 M% critical chloride concentration are also load-bearing and should be sensitivity-tested, but the more decisive gap for the paper's own claim is the absence of any out-of-sample prediction.","tokens_in":13840,"tokens_out":10032,"duration_ms":92773,"concrete_test":"Run leave-one-out cross-validation on the four retained wire-sensor points: for each point i, fit (a, Dt, be) using the other three points and predict the depth at the held-out time; report the held-out depth error and compare it with the in-sample residual. If held-out errors are large or do not decrease as the training set grows, the progressive-accuracy claim fails. Using the same calibrated models, compute DEff,C(t) with the CCrit range 0.54-5.4 kg/m^3 at t=509 days and t=6692 days, and state whether the drilling-dust values 0.48e-12 and 0.12e-12 m^2/s fall inside the prediction band.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that progressive calibration with Bayesian optimization yields increasingly accurate predictions. But the evidence in Section 5.1 is calibration, not prediction: Table 4 lists x_calc equal to the x_exp values used as optimization targets, and Figure 7 plots the same four points against the fitted curve. With three free parameters (a, Dt, be) and at most four calibration points, close agreement at the calibration points is expected and does not demonstrate generalization to new wire depths or times. The only independent check is the drilling-dust comparison in Section 5.4, which provides two effective diffusion coefficients (0.48e-12 m^2/s at t=509 days and 0.12e-12 m^2/s at t=6692 days). Under the reported four-point parameters, the calibrated Gehlen model gives values on the order of 0.7e-12 and 0.5e-12 m^2/s at those times, so the 2024 point is off by roughly a factor of four; the paper asserts 'good agreement' without a quantitative tolerance or uncertainty band. The fitted parameters are also unstable across progressive calibrations (a ranges from about 0.1 to 0.78, Dt from 1.18e-12 to 8.69e-12), indicating weak identifiability that further undermines the physical interpretation of the diffusion coefficient. Thus the load-bearing part of the claim, that more sensor data leads to more accurate predictions, is not currently supported by any prediction on a datum not used in the fit.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper calibrates two models for chloride ingress into concrete using wire-sensor breakage times from the 'Concerto' test bridge: a physics-based Gehlen diffusion model with three parameters (a, Dt, be) estimated by Bayesian optimization, and a feed-forward neural network that maps time and temperature to an effective diffusion coefficient. The authors claim that progressive calibration with additional wire-sensor data yields increasingly accurate predictions of chloride penetration depth and that the calibrated diffusion coefficients agree with independent drilling-dust measurements. The paper also includes a Sobol sensitivity analysis, an optimizer sanity check, and a comparison against a rapid chloride migration (RCM) test.","tokens_in":14276,"tokens_out":8360,"duration_ms":69481,"significance":"If the progressive-calibration claim were established, the work would be practically valuable for non-destructive corrosion monitoring of concrete infrastructure. The manuscript has genuine strengths: it uses multi-year real sensor data from a near-real-scale bridge, presents a transparent physics-based model, checks the Bayesian optimizer on synthetic data (Appendix C.1), and honestly acknowledges that the neural network is overfitted with four training points (Section 5.2). The commitment to publish data and code on Zenodo is also a positive feature. However, the central evidence for 'increasingly accurate predictions' is currently in-sample calibration, and the only independent comparison is sparse and not quantitatively evaluated. These load-bearing issues prevent the paper from supporting its main claim in its present form.","major_comments":[{"comment":"The demonstration of 'increasingly accurate predictions' is an in-sample calibration result, not a prediction on held-out data. In Table 4, the listed x_calc values equal the x_exp values by construction, because the optimization objective in Eq. (10) is the mean squared error over those exact depth-time pairs. Figures 6 and 7 plot the fitted model against the same points used for fitting. This does not support generalization to new wire depths or later exposure times. To substantiate the progressive-calibration claim, the authors should perform a genuine out-of-sample evaluation, such as leave-one-out calibration over the four points, or predict the deepest wire (0.03 m) using only the first three calibration points, and report quantitative predictive errors.","section":"Section 5.1, Eq. (10), Table 4, Figures 6-7"},{"comment":"The only non-circular validation is the comparison with drilling-dust-derived effective diffusion coefficients, but the text merely asserts 'good agreement' without a numerical tolerance or uncertainty band. Using the four-point calibrated parameters from Table 4 (a = 0.16, Dt = 1.29e-12 m2/s, be = 1010.7 K) in Eq. (4) gives DEff,C values on the order of 0.9e-12 and 0.6e-12 m2/s at t = 509 and 6692 days, respectively, whereas the chloride-profile regressions in Section 5.4 report 0.48e-12 and 0.12e-12 m2/s. The later value is therefore overestimated by roughly a factor of five. The paper should either quantify the discrepancy, provide confidence intervals from the parameter uncertainties, or explicitly state whether the discrepancy is expected within the CCrit uncertainty range shown in Figure 8.","section":"Section 5.4, Figure 13"},{"comment":"The calibrated parameters are strongly unstable across the progressive-calibration steps: a varies from 0.10 to 0.78 and Dt from 1.18e-12 to 8.69e-12 m2/s. With only four data points and three free parameters, this indicates weak identifiability. This instability undermines the physical interpretation of the calibrated Dt and weakens the reliability claim made in Section 5.1. The authors should report parameter uncertainties or posterior ranges (for example, using the GP surrogate in the Bayesian optimizer), and they could implement the dimensionality reduction they propose in Section 3.1 by fixing be, then check whether the remaining parameters and predictions are stable.","section":"Section 5.1, Table 4"},{"comment":"The conversion of each resistance jump into a calibration point is load-bearing: the paper assumes that corrosion started exactly 24 days before the observed wire break and that at that earlier time the chloride concentration at the wire depth equaled 0.6 M% (1.62 kg/m3). This assumption turns every resistance jump into a depth-time observation used in Eq. (10). No sensitivity analysis or validation of the 24-day offset is provided, and the offset is not treated as an uncertain quantity. The authors should justify the offset with a quantitative reference or include it as an additional calibration parameter (or at least perturb it over a plausible range and show the effect on the fitted parameters and on the Section 5.4 comparison).","section":"Section 3, Assumption 2"}],"minor_comments":[{"comment":"The cosine argument in Eq. (6) is ambiguous because of the commas in the large numbers and the missing parentheses: '2π · t + 2, 542, 453.44 / 32, 407, 303.30' could be read as (2πt + 2,542,453.44) / 32,407,303.30 or as 2πt + (2,542,453.44 / 32,407,303.30). Please rewrite with clear parentheses and without thousands separators inside the mathematical expression.","section":"Section 3, Eq. (6)"},{"comment":"The text states that the true value of Dt is '2e-11', but Table 6 lists the true value as '2 × 10−12'. These are inconsistent and should be reconciled.","section":"Appendix C.1, Table 6"},{"comment":"The wire is described as having a 'cross section area' of 0.065 mm, but mm is a unit of length, not area. Please specify either a diameter of 0.065 mm or give the cross-sectional area in mm2 (e.g., π·(0.0325 mm)2).","section":"Section 1.2, Section 3"},{"comment":"Several table titles contain the typo 'T able' instead of 'Table' (e.g., Tables 1-3), and the heading 'Calibration using one Data pair' in Table 4 should read 'one data point'.","section":"General formatting"},{"comment":"The neural network training details are incomplete: the number of epochs, learning rate, and final loss value are not reported. Since the paper acknowledges overfitting, these details would help the reader judge the convergence claim.","section":"Section 5.2, Algorithm 2"},{"comment":"The description of fitting DEff,C from the chloride profiles says 'adjusting the calculated chloride concentration CS,Δx to the assumed chloride profile', which is unclear; it should clarify which parameters are adjusted in the nonlinear regression and how the uppermost layers are excluded.","section":"Section 4"},{"comment":"Reference [11] is cited as 'DIN EN 12390-11:2015-11' but the standard number is 'DIN EN 12390-11'; also reference [6] is listed as a BAW data sheet without a full title or year, making it difficult to locate.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is probably better positioned as an application case study for structural health monitoring than as a general method paper; the novelty should be framed around the sensor-data calibration workflow rather than a fundamental methodological advance. The in-sample validation issue in Section 5.1 and the unquantified independent comparison in Section 5.4 are the key blockers. The data-availability statement is welcome, but the code and data are not yet available for verification; this should be clarified in the final version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nRead of arXiv:2411.16447. The worthwhile core is the calibration workflow: taking wire-sensor break times from the Concerto bridge and fitting Gehlen's chloride diffusion model with Bayesian optimization, including a clean-data sanity check and an explicit acknowledgment that the neural-network comparison overfits with four points. The paper also produces a non-circular external comparison against independent drilling-dust profiles, which is more than many sensor papers do.\n\nWhat is actually new: a set of calibrated values for a, Dt, and be on this specific sensor dataset, and a demonstration that a physics-based model can be fit progressively as sensor data arrive. That is a legitimate but modest contribution to corrosion monitoring.\n\nThe soft spots are real and mostly where the stress-test lands. First, the central claim—'progressive calibration ... leads to increasingly accurate predictions'—is supported by Table 4 and Figures 6–7, which essentially show the fitted depths matching the depths used as optimization targets. That is calibration accuracy, not prediction accuracy. No wire depth or time not used in the fit is held out. Second, the fitted parameters are unstable across the progressive calibrations (a goes from about 0.78 to 0.16, Dt from 8.7e-12 to 1.3e-12), which signals weak identifiability and undermines the physical interpretation of the effective diffusion coefficient. Third, the independent drilling-dust check is sparse: two effective diffusion coefficients, 0.48e-12 at 509 days and 0.12e-12 at 6692 days. Under the four-point parameters, the Gehlen model gives roughly 0.7e-12 and 0.5e-12 at those times—the later point is off by about a factor of four. The paper calls this 'good agreement' without an uncertainty band or tolerance. That is a load-bearing weakness.\n\nTo the authors' credit, they do not hide the overfitting of the neural network, and the data/code promise is good. The Sobol dimensionality reduction is explicitly left unimplemented, which is fine as a suggestion but should not be read as part of the contribution.\n\nWho gets value? People working on sensor-based chloride monitoring and maintenance decisions. They get a real case study and a workflow, not a validated predictive model. It deserves a serious referee—the dataset and the external comparison are worth engaging—but the authors should be pushed to either add a held-out point or provide a quantitative uncertainty analysis around the drilling-dust comparison. Recommend peer review with major revision.","headline":"A competent calibration study whose 'increasingly accurate predictions' claim is really about fitting the same four data points, not predicting new ones.","tokens_in":14696,"tokens_out":3751,"would_cite":false,"duration_ms":33067,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Progressive calibration of a physics-based chloride diffusion model with Bayesian optimization, using wire-sensor break times, yields increasingly accurate predictions of chloride penetration depth and effective diffusion coefficients…","keywords":["chloride-induced corrosion","effective diffusion coefficient","Bayesian optimization","wire sensor","model calibration","concrete durability","chloride migration","neural network surrogate"],"falsifier":"Run the same optimization while varying the assumed corrosion-start lag (for example, ±7 and ±30 days) and the critical chloride concentration (0.2 to 2 M%) and observe whether the fitted diffusion coefficients remain within the drilling-dust values; alternatively, place additional wires at the same depths on a new exposure and compare predicted versus measured break times without re-fitting.","tokens_in":13660,"feed_emoji":"🧪","tokens_out":5055,"duration_ms":45334,"temperature":0.7,"pith_summary":"The paper asks whether a physics-based chloride diffusion model can be kept accurate over time by recalibrating it every time a cheap wire sensor in the concrete breaks. It shows that, as more wire-break events are added to a Bayesian optimization routine, the model's predicted chloride penetration depths converge to the observed ones, and the fitted effective diffusion coefficients land in the same range as values obtained from independent drilling-dust chloride profiles. In contrast, a neural network trained on the same four events overfits and cannot generalize, although it reproduces the calibration points. The practical payoff would be a monitoring system where corrosion risk is updated continuously from simple resistance measurements, reducing the need for destructive sampling until visible damage appears.","feed_headline":"Each wire break sharpens chloride corrosion forecasts","feed_subtitle":"Progressive Bayesian recalibration matches drilling-dust measurements and reduces need for destructive sampling.","key_machinery":"The central object is the analytical chloride diffusion model of reference [1], which gives the chloride concentration profile $C(x,t)$ as an error-function expression with an effective diffusion coefficient that depends on time and temperature. By inverting the error function, the model becomes a deterministic map from the parameters (aging exponent $a$, temperature sensitivity $b_e$, and migration-coefficient factor $D_t$) to the depth at which the critical chloride content is reached at a given sensor time. Bayesian optimization—a Gaussian-process surrogate with expected-improvement acquisition—finds the parameter triple that minimizes the squared error between calculated and observed depths. The wire sensor supplies the observed depths and times: each resistance jump is converted into a calibration point by assuming the wire began corroding 24 days earlier at a chloride concentration of 0.6 M% by cement mass, equivalent to 1.62 kg/m$^3$.","core_discovery":"The paper demonstrates that progressively calibrating a physics-based chloride diffusion model with wire-sensor break times via Bayesian optimization yields increasingly accurate predictions of chloride penetration depth and effective diffusion coefficients. Starting from the analytical solution of Fick's second law for chloride ingress, the authors isolate the penetration depth expression via the inverse error function and treat the measured (depth, time) wire-break pairs as data for fitting the parameters $a$, $b_e$, and $D_t$ in the effective diffusion coefficient. They use Bayesian optimization with a Gaussian-process surrogate and expected-improvement acquisition to find the parameter triple that minimizes the mean squared error in predicted depth. Calibrating progressively with one, two, three, and four wire-break events makes the predicted time-versus-depth curve align ever more closely with the observed events, and the effective diffusion coefficients extrapolated to the sampling dates agree with the drilling-dust values ($0.48 \\times 10^{-12}$ m$^2$/s at 509 days, $0.12 \\times 10^{-12}$ m$^2$/s at 6692 days) when the critical chloride content is taken at the lower bound of its uncertainty range. The authors conclude that the progressive calibration of the model, paired with Bayesian optimization, leads to increasingly accurate predictions.","pith_inferences":["The 24-day lag and the 0.6 M% critical concentration are the main assumptions that convert resistance jumps into calibration pairs; replacing them with direct electrochemical depassivation measurements would make the method transferable to other structures without re-estimating these offsets.","Because the neural network overfits with only four data points, adding more wire sensors at intermediate depths would allow a meaningful test of whether the data-driven approach catches up to the physics-based model as data grow.","The model assumes one-dimensional, homogeneous diffusion; extending the calibration to cracked or layered concrete with a numerical solver would check whether the fitted parameters remain physically stable.","A direct cross-validation on a second exposure cycle would test whether the calibrated diffusion coefficients predict the timing of new sensor breaks before they occur."],"forward_implications":["If the calibration is correct, a concrete structure's corrosion risk can be updated continuously from resistance measurements alone, without destructive dust sampling.","The progressive calibration procedure makes model-predicted penetration depths trustworthy only after the first few sensor events; earlier predictions are less reliable.","Uncertainty in critical chloride content can be propagated through the model to give a band of predicted corrosion times, improving maintenance planning.","The comparison shows that the physics-based model needs less data than the flexible neural network to generalize, and the two approaches could be combined to exploit both interpretability and flexibility."],"supporting_citations":[{"why":"Supplies the analytical chloride diffusion model with time- and temperature-dependent effective diffusion coefficient that the paper calibrates.","marker":"[1]"},{"why":"Documents the wire sensor and provides the 24-day lag and 0.6 M% critical-concentration assumption that converts resistance jumps into calibration pairs.","marker":"[5]"},{"why":"Defines the rapid chloride migration test used as an independent reference for the diffusion coefficient.","marker":"[6]"},{"why":"Provides the weathering-modified diffusion formulations that the model extends.","marker":"[7]"},{"why":"Supplies the state-of-the-art durability modeling framework that the diffusion expression is adapted from.","marker":"[8]"},{"why":"Provides the Bayesian optimization package used for parameter fitting.","marker":"[9]"},{"why":"Defines the standard method for least-squares regression of chloride profiles from drilling dust, used to derive the comparative effective diffusion coefficients.","marker":"[11]"}],"fun_headline_variants":["Wire breaks fine-tune corrosion model","Bayesian tuning sharpens rust depth forecasts","Progressive sensor data hones chloride predictions","Smart calibration cuts need for concrete drilling","Physics plus data boosts corrosion forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole calibration rests on the assumption that a resistance jump implies the wire began to corrode exactly 24 days earlier, when the chloride concentration at the wire depth was 0.6 M% by cement mass; if the corrosion-start lag or that critical concentration is wrong, every fitted parameter and every subsequent prediction shifts.","fun_headline_variants_meta":{"raw":{"variants":["Wire breaks fine-tune corrosion model","Bayesian tuning sharpens rust depth forecasts","Progressive sensor data hones chloride predictions","Smart calibration cuts need for concrete drilling","Physics plus data boosts corrosion forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1374,"prompt_tokens":990,"completion_tokens":384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":336}},"tokens_in":606,"tokens_out":384,"duration_ms":4238,"temperature":1.0,"reasoning_tokens":336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:05:18.129354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same optimization while varying the assumed corrosion-start lag (for example, ±7 and ±30 days) and the critical chloride concentration (0.2 to 2 M%) and observe whether the fitted diffusion coefficients remain within the drilling-dust values; alternatively, place additional wires at the same depths on a new exposure and compare predicted versus measured break times without re-fitting.","supporting_citations":[{"cited_title":"Gehlen, Probabilistische Lebensdauerbemessung von Stahlbetonbauwerken – Zu- verl¨ assigkeitsbetrachtungen zur wirksamen Vermeidung von Bewehrungskorrosion","cited_arxiv_id":null,"evidence_quote":"Supplies the analytical chloride diffusion model with time- and temperature-dependent effective diffusion coefficient that the paper calibrates."},{"cited_title":"Korrosionsmonitoring von stahlbetonbauwerken als element des lebensdauermanagements,","cited_arxiv_id":null,"evidence_quote":"Documents the wire sensor and provides the 24-day lag and 0.6 M% critical-concentration assumption that converts resistance jumps into calibration pairs."},{"cited_title":"BA W - Merkblatt: Dauerhaftigkeitsbemessung bei Carbon- atisierung und Chlorideinwirkung (MDCC)","cited_arxiv_id":null,"evidence_quote":"Defines the rapid chloride migration test used as an independent reference for the diffusion coefficient."},{"cited_title":"Chloride transport in concrete: theory and application,","cited_arxiv_id":null,"evidence_quote":"Provides the weathering-modified diffusion formulations that the model extends."},{"cited_title":"Lausanne, Switzerland: CEB Bulletin No","cited_arxiv_id":null,"evidence_quote":"Supplies the state-of-the-art durability modeling framework that the diffusion expression is adapted from."},{"cited_title":"DIN EN 12390-11:2015-11, Pr¨ ufung von Festbeton - Teil 11: Bestimmung des Chloridwiderstandes von Beton - Einseitig gerichtete Diffusion; Deutsche Fassung EN 12390-11:2015","cited_arxiv_id":null,"evidence_quote":"Defines the standard method for least-squares regression of chloride profiles from drilling dust, used to derive the comparative effective diffusion coefficients."}],"review_version":1}