{"id":"873e7448-3460-4135-83c3-71baa60ef227","arxiv_id":"2505.13870","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A hybrid iterative fitting method combines the Bornitz formula with symbolic regression to derive a frequency-dependent ground vibration attenuation formula for the HEPS site, with MAE 3.13e-7 s2 and RMAE 90.47%.","lead":"This paper fits a site-specific ground vibration attenuation formula to field data from the High Energy Photon Source in Beijing, using an iterative method that alternates between classical Bornitz curve fitting and machine-learning symbolic regression. It reports better interpretability than black-box models and better accuracy than standard empirical formulas, though the accuracy claims rest on the same data used to fit the formula.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No held-out validation: the headline MAE/RMAE and the Bornitz/Yang comparison are in-sample fits, so the claimed 'discovered law' is not yet shown to transfer to unmeasured distances or frequencies.","rationale":"The paper's contribution is a workflow plus a site-specific formula; I read it as a case study, not a claim that Eq. (12)–(13) is universal. The method is clearly described and the data collection appears careful. The soft spot is that every accuracy number in the abstract and Section 4 is computed on the same data used for fitting. More importantly, the comparison with Bornitz/Yang is not apples-to-apples: those fixed-coefficient formulas have fewer parameters, so their higher in-sample MAE does not demonstrate that the discovery method is better. The FE simulation is only a trend check and explicitly disclaims physical significance of the simulated coefficients. Therefore, the single load-bearing condition—that the fitted n(f), α(f) capture the actual propagation, not just the training grid—is untested. My proposed hold-out-distance experiment directly tests this: if the near-field fit predicts far-field amplitudes, transferability is supported; if not, the formula is an interpolation. Because the concern is falsifiable and fixable with validation, the reader's CONDITIONAL verdict is unchanged.","tokens_in":13217,"tokens_out":6848,"duration_ms":68849,"concrete_test":"Refit Eq. (12)–(13) using only sensors S1–S50 (r ≤ ~50 m) for all available frequencies 1–100 Hz, then predict the held-out far-field sensors S51–S64 (r = 100–336 m) and compute the RMAE and the relative-error CDF on those held-out points. If the held-out RMAE is close to the in-sample 90.47% and the CDF does not degrade, the f-only Bornitz structure is supported; if far-field errors are substantially larger, the fitted coefficients have absorbed distance-dependent attenuation and the formula is an interpolation artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Eq. (11)–(13) is a high-precision attenuation law—is supported only by training-set error. The same 64-sensor, 1–100 Hz dataset is used to fit n(f) and α(f) and to report MAE=3.13e-7 s2 and RMAE=90.47% (Section 3.2); no distance range, frequency band, or second site is held out. The comparisons in Table 2 are also in-sample for every method, and the proposed formula has frequency-dependent coefficients, so a lower in-sample MAE than fixed-parameter Bornitz and Yang formulas is an expected flexibility effect rather than evidence of superiority. The only transferability check is an ABAQUS homogeneous-elastic model at four frequencies, which the paper itself says yields values 'without direct physical significance' (Section 3.3); it cannot validate that the adopted Bornitz structure, with n and α depending only on f, holds in the actual heterogeneous HEPS soil. Consequently, Eq. (12)–(13) may be a local interpolation over the measured r–f grid, and the claimed predictive advantage is unestablished.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid method that combines the classical Bornitz attenuation formula with an intelligent formula generation model (a sequence-to-sequence approach) to derive frequency-dependent geometric and material damping coefficients, n(f) and alpha(f), from field vibration measurements at the High Energy Photon Source (HEPS) site. The final expressions, Eq. (11)–(13), are reported with MAE 3.13e-7 s2 and RMAE 90.47% (Section 3.2). The authors argue that the structure is physically justified by energy-conservation arguments and by an ABAQUS/Explicit finite-element simulation that reproduces the increasing trend of n(f) with frequency. They then construct a distance-dependent error band, Eq. (18), using a log-normal assumption, and compare the formula's fitting accuracy with calibrated fixed-parameter Bornitz and Yang formulas, XGBoost, and a DNN (Table 2). The paper claims that the proposed method balances interpretability and accuracy and provides a transferable framework for vibration assessment at similar facilities.","tokens_in":13567,"tokens_out":2216,"duration_ms":22959,"significance":"If the central claim were established, the paper would offer a useful contribution: an interpretable, site-specific formula for ground-vibration attenuation that is frequency-dependent, based on a rich field dataset (64 sensors, 1–100 Hz, distances to 336 m). The strengths are the careful field experiment, the sensor consistency checks, the explicit iterative fitting procedure, the comparison with several baselines, and a probabilistic error model. However, the significance is diminished by a central methodological gap: all reported accuracies are in-sample, and the error model is calibrated and evaluated on the same residuals. The paper is honest about some of these limitations (e.g., the FE values 'do not carry direct physical significance'), but the abstract and conclusions state a stronger claim of 'high-precision' and 'predictive accuracy' than the evidence supports. The contribution is therefore better framed as a demonstration of a fitting methodology with an in-sample description of the HEPS dataset, not as a validated predictive law.","major_comments":[{"comment":"The reported MAE (3.13e-7 s2) and RMAE (90.47%) are computed on the same data used to fit n(f) and alpha(f); no distance range, frequency band, or second site is held out. The claim that Eq. (11)–(13) is a 'high-precision' attenuation law is therefore supported only by training-set error. Please provide a held-out evaluation, e.g., leave-one-sensor-out, a random 20% test split, or an external site, and report the test-set MAE/RMAE. Without such validation, the formula is a local interpolation over the measured r–f grid rather than a demonstrated predictive law.","section":"Section 3.2"},{"comment":"The error-band model sigma(r) in Eq. (18) is fitted to the residuals of Eq. (11)–(13), and then the same residuals are used to report that 71.5% of points fall within the 1-sigma range. This is a self-consistency check, not a predictive calibration. Under an assumed normal (or log-normal) distribution, fitting sigma to the data will by design produce coverage close to the nominal level when evaluated on those same data. Please validate the error band on independent data (e.g., a held-out set) or at least state explicitly that the reported coverage is an in-sample property and re-frame the claim accordingly. A chi-square or probability-integral-transform test on held-out data would be a more meaningful check.","section":"Section 4.1"},{"comment":"The finite-element simulation is used to support the assumption that n(f) depends on frequency and not on distance, but the text itself states that the simulated n values 'do not carry direct physical significance' because of the difficulty of modeling real, heterogeneous soil. The simulation only reproduces an increasing trend at four frequencies (1, 5, 10, 20 Hz) in a homogeneous elastic half-space, so it does not validate that the Bornitz structure with n and alpha depending only on f holds in the actual HEPS soil. The paper should temper the claim that the FE results 'support our assumption' and should discuss what additional data (e.g., a second site, varying source amplitudes, or a layered soil model) would be needed to test transferability of the discovered formula.","section":"Section 3.3"},{"comment":"The comparison with XGBoost and DNN is not a fair predictive comparison as presented. The ML models are trained on 75% of the data, but the evaluation set is not specified; if Table 2 reports errors on the training data for the ML models, that would understate their generalization error, while the proposed formula is evaluated in-sample on all data. Please specify the exact evaluation protocol (training/test split for ML and the identical split for the Bornitz/Yang calibrations) and report test-set metrics for all methods. Without this, the claimed 'significant advantages' of the proposed formula over black-box models are not established.","section":"Section 4.2"}],"minor_comments":[{"comment":"There are several notation inconsistencies: Eq. (1) uses A1 and r1, while the text and later equations use Ar and r; the subscripts in Eq. (9) are garbled (e.g., 'f f' and 'Sfx'); and the text around Eq. (2)–(4) refers to 'A0 and r0' and 'A1 and r1' in ways that are not consistently defined. Please unify the notation.","section":"Throughout"},{"comment":"Fig. 3(b) is described as an 'on-site consistency test' but the caption and text refer to 'Fig. 2(b)' in one place; please correct the cross-reference.","section":"Section 3.1"},{"comment":"The Shapiro–Wilk test is described, but the p-values are not reported; only W statistics are shown in Fig. 12. Reporting p-values (or at least stating the sample size per r) would allow the reader to judge whether the normality assumption is actually supported at each distance.","section":"Section 4.1"},{"comment":"The sentence 'The distribution of these data points does not exhibit any clear pattern ... This finding demonstrates the robust generalization capability of the formula within the tested range' is an overinterpretation: a random-looking pattern of large errors does not demonstrate generalization, which requires a held-out evaluation.","section":"Section 4.1"},{"comment":"The description of the iterative algorithm would benefit from a formal statement of the optimization objective and a convergence criterion. As written, it is unclear when the iterations stop and whether the final result depends on the starting point n(f)=0.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an honest and careful in-sample fitting study, but the central claims of 'discovered law' and 'high precision' rest on training-set errors and a circular error-band calibration. The requested revisions are substantial but achievable within the paper's scope: add held-out validation, re-frame the claims, and clarify the ML comparison protocol. I would not recommend reject, but the current form overstates the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on vibration assessment or symbolic-regression workflows. The new piece is the iterative wrapper: fix one Bornitz coefficient, solve the other by linear regression, fit the frequency dependence with the formula generator, then alternate. It is a modest but workable idea, and the HEPS field dataset is genuinely valuable—64 stations, 1–100 Hz swept sines, coherence gating, transients discarded. The authors also do the reader the courtesy of showing the error decrease across iterations and of comparing against standard formulas and black-box models with the same inputs.\n\nThe soft spots are real but not fatal. Most importantly, there is no held-out validation. The reported MAE of 3.13e-7 s2 and RMAE of 90.47% are training-set errors; the comparison against Bornitz and Yang is also in-sample, so the frequency-dependent form has an expected flexibility advantage that does not by itself prove superiority. The error-band analysis is self-referential: sigma(r) is fit to the same residuals it is then used to explain, so the '71.5% within 1-sigma' is a consistency check, not a predictive statement. The FE validation only reproduces the trend of n(f) and the authors explicitly say the coefficient values lack direct physical significance—fair of them, but it means the structural assumption that n and alpha depend only on f, not on r, remains weakly checked. The RMAE numbers are high, and the paper leans on distribution arguments to soften them; that is reasonable, but it undercuts the 'high-precision' wording.\n\nThe citation pattern is fine: the prior Bornitz variants and the authors' own formula-generation model are properly credited. The thinking is generally clear, and the authors are honest about several limitations. The paper would benefit from a leave-one-distance-out or train-on-low-frequencies/test-on-high-frequencies split, from reporting out-of-sample metrics, and from releasing the data and code. With those additions it could be a useful reference for similar large-facility sites. As it stands, the formula is best treated as a site-specific interpolation; the claim of transferable law is unproven.\n\nSend it to review. It deserves referee time, but the revision should make the validation honest and the language less celebratory.","headline":"A solid, carefully executed case study that delivers a site-specific attenuation formula, but the headline accuracy claims rest on in-sample fit and the 'discovered law' needs out-of-sample validation before it earns that name.","tokens_in":14023,"tokens_out":2674,"would_cite":false,"duration_ms":27006,"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":"A hybrid iterative loop that alternates least-squares fitting with a symbolic-regression model discovers a compact, frequency-dependent Bornitz attenuation law from field vibration data, with a mean absolute error of…","keywords":["ground vibration attenuation","Bornitz formula","formula discovery","symbolic regression","machine learning","frequency-dependent damping","field vibration testing","vibration-sensitive facilities"],"falsifier":"Apply the published $n(f)$ and $\\alpha(f)$ to a second, geologically different site, or to a held-out set of frequencies and distances from the same site, and check whether relative errors grow systematically with distance or amplitude; alternatively, fit a model in which $n$ also depends on $r$ and see whether it substantially improves the original data—either outcome would show the frequency-only Bornitz separation is not a transferable law.","tokens_in":13066,"feed_emoji":"📊","tokens_out":15379,"duration_ms":127291,"temperature":0.7,"pith_summary":"The paper aims to show that a frequency-dependent ground-vibration attenuation law can be discovered automatically from field measurements, rather than assumed from theory or delivered as a black box. It locks the overall shape of attenuation to the classical Bornitz power-exponential decay law, then runs an iterative loop in which a symbolic-regression model proposes functional forms for the two damping coefficients while ordinary least squares pins down their numerical values. Applied to a 336-meter measurement line at a synchrotron facility in Beijing under 1–100 Hz sinusoidal excitation, the loop converges in four rounds to a closed-form formula with mean absolute error $3.13\\times10^{-7}\\,\\mathrm{s}^2$ and relative mean absolute error $90.47\\%$. If the claim holds, site-specific interpretable vibration-prediction formulas can be produced quickly for vibration-sensitive scientific infrastructure.","feed_headline":"Four fitting rounds turn field data into a ground-vibration law","feed_subtitle":"Iterative symbolic regression plus least squares yields an interpretable 1–100 Hz attenuation law for a synchrotron site.","key_machinery":"The engine is the frequency-parameterized Bornitz equation $A(r,f)=A_0(r_0/r)^{n(f)}\\exp(-\\alpha(f)(r-r_0))$, which keeps the classical multiplicative separation of geometric spreading and material absorption while making both coefficients functions of frequency. The method alternates two cheap steps: fix $\\alpha(f)$, solve for the optimal $n(f)$ at each frequency by unweighted least squares on the linearized log-amplitude relation, then use an intelligent formula generation model—a generative symbolic-regression network that emits candidate formulas token by token—to fit a readable expression to those per-frequency values; then fix $n(f)$ and repeat for $\\alpha(f)$. Four iterations starting from $n(f)=0$ reduce the fitting error and terminate in explicit formulas for both coefficients, so the final object is an inspectable equation rather than a trained network.","core_discovery":"The central claim is that the hybrid iterative fitting procedure yields, from field data alone, an explicit frequency-dependent Bornitz law of the form $A(r,f)=A_0(r_0/r)^{n(f)}\\exp(-\\alpha(f)(r-r_0))$, with $n(f)$ a piecewise expression and $\\alpha(f)$ a compact expression in frequency, giving a mean absolute error of $3.13\\times10^{-7}\\,\\mathrm{s}^2$ and a relative mean absolute error of $90.47\\%$ on the measured data. The paper further claims that this law outperforms both a calibrated fixed-parameter Bornitz formula and a prior frequency-dependent Bornitz variant in MAE and RMAE, that its reported minimum geometric damping coefficient of $0.47$ is consistent with the theoretical lower bound of $0.5$, and that its structure agrees with finite-element simulations showing $n$ is essentially independent of distance and grows with frequency. A log-normal residual model supplies distance-dependent $\\sigma$ bands that contain $98.7\\%$ of the data within $3\\sigma$, giving the deterministic formula a probabilistic error envelope.","pith_inferences":["If the frequency-only separation is an approximation rather than an exact law, the derived formula is best read as a compact interpolation over the measured 1–100 Hz and 2–336 m grid; extrapolating to other distances, frequencies, amplitudes, or soil columns would likely require refitting or additional dependence on soil properties.","The discovered frequency dependence of the coefficients invites a physical mapping against independent soil measurements: $\\alpha(f)$ could be compared with viscoelastic damping models, and $n(f)$ with wave-front geometry, to test whether the fitted coefficients carry genuine material meaning.","The same alternating least-squares-plus-symbolic-regression loop could be applied to other physically anchored formulas, for example with coefficients depending on source depth or excitation amplitude, giving a general recipe for transparent predictive laws in engineering geophysics.","Because the symbolic-regression step is stochastic, different runs may converge to different but statistically equivalent expressions; reporting the spread of candidate formulas would clarify how much of the discovered form is forced by the data and how much is a modeling choice."],"forward_implications":["Site-specific attenuation formulas can be derived automatically from a single field campaign, replacing manual calibration of the damping coefficients.","Because the final law is a closed-form equation, vibration impact assessments can trace predictions to physically meaningful geometric and material damping coefficients.","The discovered formula beats calibrated fixed-parameter Bornitz and earlier frequency-dependent formulas in both MAE and RMAE on this site, while black-box models that fit slightly better produce physically impossible negative amplitudes.","A distance-dependent probabilistic error band can be attached to the deterministic formula, giving engineers a quantified confidence interval for vibration limits.","The iterative scheme is not tied to one site; the same loop can be rerun wherever a vibration-sensitive facility is planned or monitored."],"supporting_citations":[{"why":"Supplies the original Bornitz attenuation formula whose multiplicative power-exponential shape the whole method preserves.","marker":"[22]"},{"why":"Provides the intelligent formula generation model that the hybrid loop uses to propose symbolic expressions for the damping coefficients.","marker":"[31]"},{"why":"A previous frequency-dependent Bornitz formula derived from theory and experiments; the main formula-based baseline the new law must beat.","marker":"[37]"},{"why":"Earlier work establishing a positive correlation between material damping and frequency, used to justify the form of $\\alpha(f)$.","marker":"[36]"}],"fun_headline_variants":["Hybrid ML derives explicit ground-vibration attenuation law","From field data to an interpretable Bornitz law","Interpretable vibration formula outperforms black-box models","Auto-derived Bornitz law for synchrotron site","Data-driven formula beats empirical vibration models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The procedure assumes the attenuation law has the exact multiplicative form $A = A_0 (r_0/r)^{n} e^{-\\alpha(r-r_0)}$ and that the two decay coefficients $n$ and $\\alpha$ depend only on frequency, not on distance, vibration amplitude, soil layering, or source character; if attenuation is more entangled than that, the discovered formula is only an interpolation over the measured distances and frequencies at one site.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid ML derives explicit ground-vibration attenuation law","From field data to an interpretable Bornitz law","Interpretable vibration formula outperforms black-box models","Auto-derived Bornitz law for synchrotron site","Data-driven formula beats empirical vibration models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000529,"raw_usage":{"total_tokens":2555,"prompt_tokens":952,"completion_tokens":1603,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":1530}},"tokens_in":568,"tokens_out":1603,"duration_ms":10314,"temperature":1.0,"reasoning_tokens":1530,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:08:27.095856+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the published $n(f)$ and $\\alpha(f)$ to a second, geologically different site, or to a held-out set of frequencies and distances from the same site, and check whether relative errors grow systematically with distance or amplitude; alternatively, fit a model in which $n$ also depends on $r$ and see whether it substantially improves the original data—either outcome would show the frequency-only Bornitz separation is not a transferable law.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the intelligent formula generation model that the hybrid loop uses to propose symbolic expressions for the damping coefficients."},{"cited_title":"Bornitz, Über die Ausbreitung der von Großkolbenmaschinen erzeugten Bodenschwingungen in die Tiefe, Springer-Verlag, 2013","cited_arxiv_id":null,"evidence_quote":"A previous frequency-dependent Bornitz formula derived from theory and experiments; the main formula-based baseline the new law must beat."},{"cited_title":"Verma, T.N","cited_arxiv_id":null,"evidence_quote":"Earlier work establishing a positive correlation between material damping and frequency, used to justify the form of $\\alpha(f)$."}],"review_version":1}