{"id":"8599dbdd-97aa-47af-99d4-107b11e7a8f1","arxiv_id":"2607.27137","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Coupling a SINDy-discovered CHIKV ODE model to an EnKF corrects noisy forecasts and recovers unobserved host and vector states in synthetic partial-observation experiments.","lead":"A hybrid of sparse equation discovery (SINDy) and ensemble Kalman filtering reconstructs Chikungunya host–vector dynamics from synthetic noisy, partial observations. It matters as a practical template for epidemic forecasting when surveillance only sees a few compartments.","discovery_kind":"new_application","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"Table 6's >98% RMSE reduction may measure EnKF rescuing a structurally wrong propagator, not a working SINDy-EnKF hybrid: the bilinear library cannot represent (2.4)'s rational incidence terms, and Appendix C's own equations contradict the clean-data recovery claim.","rationale":"This sharpens rather than repeats the reader's weakest assumption. The reader worried the bilinear library is \"matched to the generator,\" inflating discovery success; the actual situation is the opposite and more damaging — the library cannot express the generator's rational incidence terms, so the clean-data exact-recovery claim in §3.1 is internally inconsistent with Appendix C (dVh=0, dTh=0). The reader did note appendix/prose tension but did not trace its consequence: the premise that SINDy supplies a meaningful forecast operator is unproven, and Table 6's headline number is computed against an open-loop baseline guaranteed to look terrible, so the demonstrated effect may be entirely attributable to the EnKF. This is a correctness/internal-consistency concern, not a consensus disagreement. I nonetheless agree with the reader's CONDITIONAL rather than REJECT: as a synthetic feasibility demo the EnKF correction evidence (Figs. 7–10, dtobs sensitivity) is coherent, and the proposed three-propagator comparison is cheap and would either rescue or deflate the hybrid attribution cleanly. The conditions for acceptance should include (i) the propagator-ablation test above, (ii) correcting or retracting the §3.1 recovery claim in light of Appendix C, and (iii) a threshold-selection rule that does not use ground-truth RMSE.","tokens_in":31658,"tokens_out":3266,"duration_ms":129596,"concrete_test":"On the identical 25%-noise, 4-channel synthetic setup of §3.3, run the EnKF three times with different forecast operators: (a) the 25%-noise SINDy model, (b) the true system (2.4), (c) a persistence propagator x_{n+1}=x_n. Report per-compartment posterior RMSE as in Table 6. If (c)≈(a), SINDy contributes nothing and the hybrid claim collapses to plain EnKF; if (a)≈(b), the claim holds. Additionally, re-run clean-data SINDy with a library augmented by rational incidence terms (Iv·Sh/Nv etc.); if the true equations are still not recovered, the §3.1 recovery claim should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — SINDy provides a forecast operator good enough that EnKF assimilation reconstructs the full 10-compartment state — requires the SINDy model to carry genuine dynamical information about (2.4). Two observations undercut this. (1) The forces of infection in (2.4) are rational in the state: λh = βv·Iv/Nv and λv = βh(ε1Ih+ε2Jh+ε3Th)/Nh with Nv, Nh state-dependent sums. The library (§3.1: constant, linear, pairwise bilinear monomials) contains no such terms, so SINDy cannot recover (2.4) even from clean data — it can only fit a polynomial proxy. Appendix C confirms non-recovery: the \"clean data\" discovered system has dVh = 0.0000 and dTh = 0.0000 identically, no Iv-dependent term anywhere in dSh or dEh, and a spray of spurious quadratics (e.g., −0.0237·Sv, +0.0133·Eh²). This directly contradicts §3.1's statement that SINDy \"reproduces both the structural form and the parameter values of the original model (2.4) to within numerical precision.\" (2) At 25% noise the forecast operator used for Table 6 is worse: dSv ≈ 12.03·Sh − 48.7·Eh − 289.6·ShIh + ..., bearing no resemblance to (2.4). The >98% RMSE \"reduction\" is then measured against that wrong model's open-loop divergence — a baseline any filter assimilating 4 directly observed channels at dtobs ≤ 1 would beat. Note the Ih row already shows the filter can hurt (−279% \"reduction\"). So the experiment as designed cannot distinguish \"SINDy-EnKF hybrid works\" from \"EnKF with frequent partial observations pins down the state regardless of propagator.\" Train and test trajectory are also the same synthetic run, with λ tuned per noise level.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript formulates a 10-compartment host–vector ODE model for chikungunya (2.4), then studies a pipeline in which SINDy with a constant/linear/bilinear library learns a surrogate from simulated trajectories at noise levels 0–50%, and the discovered system serves as the forecast operator of an EnKF that assimilates noisy observations of four host compartments (Vh, Ih, Jh, Th) to reconstruct the full state. Claims: (i) SINDy recovers (2.4) exactly from clean data; (ii) it degrades gracefully then badly with noise; (iii) the SINDy–EnKF hybrid reduces RMSE by >98% in most compartments (Table 6) and reconstructs unobserved states, with skill controlled by the assimilation interval dtobs (Fig. 7). All experiments are synthetic, generated from (2.4) with known observation operators. The pipeline is not definitionally circular, but the clean-data recovery claim is contradicted by the manuscript's own Appendix C, and the headline Table 6 numbers are computed against an open-loop baseline that any filter with frequent direct observations would beat.","tokens_in":32106,"tokens_out":5242,"duration_ms":116360,"significance":"A working recipe coupling sparse model discovery with sequential assimilation would be useful for compartmental epidemiology, where partial and noisy observation is the norm. The paper's strengths are a systematic noise-sensitivity sweep for SINDy on a realistic 10-compartment host–vector model (§3.2), an explicit dtobs-sensitivity analysis showing where the forecast model actually matters (Fig. 7), and publicly available code. However, all evidence is synthetic (self-generated RK4 data), the clean-recovery claim is internally contradicted, and the headline assimilation numbers lack a null-model baseline; as written, the contribution is a proof of concept rather than the surveillance-ready method the abstract suggests.","major_comments":[{"comment":"The claim that SINDy 'reproduces both the structural form and the parameter values of the original model (2.4) to within numerical precision' is contradicted by the manuscript's own appendix. The library (constant, linear, pairwise bilinear) cannot represent the rational forces of infection in (2.4), lambda_h = beta_v I_v/N_v and lambda_v = beta_h(eps_1 I_h+eps_2 J_h+eps_3 T_h)/N_h, with state-dependent N_v, N_h. Appendix C's 'Clean data' column confirms non-recovery: dV_h = 0.0000 and dT_h = 0.0000 identically, no I_v-dependent term in dS_h or dE_h, and spurious quadratics (e.g., +0.0133 E_h^2). Moreover, dV_h = 0 is inconsistent with Table 4's clean-data RMSE of 0.0001 for V_h, since the true dV_h ~ theta*S_h ~ O(10^2)/day with theta=0.4. Either the discovered-equation table is mislabeled or the RMSE table is. This must be resolved and the recovery claim corrected; the authors cite S S","section":"§3.1 and Appendix C (Table 7)"},{"comment":"The headline >98% RMSE reduction is measured against open-loop integration of a badly misspecified 25%-noise SINDy model (Appendix C: dS_v ~ 12.03 S_h - 48.75 E_h - 289.6 S_h I_h + ...). With four channels directly observed at dtobs <= 1, essentially any propagator is pinned near the truth, so this number cannot distinguish 'SINDy-EnKF hybrid works' from 'EnKF with frequent partial observations pins the state down.' The I_h row already illustrates the metric's fragility (-279% 'reduction'). A control is needed: e.g., EnKF with a persistence/random-walk forecast or a deliberately wrong linear model at the same dtobs, plus reporting Table 6 across dtobs. Figure 7 suggests the forecast model genuinely matters only for dtobs >= 5 — quantify this rather than relying on the high-frequency regime.","section":"§3.3 and Table 6"},{"comment":"The STLSQ threshold is hand-tuned per noise level, and the tuning is non-monotone (20% noise: lambda=1e-6, smaller than the 5% value 2e-5 and the 10% value 2.8e-5), which suggests selection against ground-truth error. In real surveillance no ground truth is available, so the pipeline that produces the forecast operator behind Table 6 currently relies on oracle information. A data-only selection rule (held-out trajectory cross-validation, Pareto-front selection, or ensemble-SINDy stability as in ref. [10]) should be specified and the main results shown to survive it.","section":"Tables 4–5 captions; §3.2.1"},{"comment":"Two load-bearing methodological gaps. (i) The derivative-estimation method for SINDy is never described: with dt = 0.0001 (Table 3) and 25–50% multiplicative noise, raw finite differences would produce enormous derivative noise, so any smoothing (Savitzky–Golay, weak form, etc.) must be stated. (ii) The EnKF configuration is unspecified: ensemble size N_e, model-error covariance Q_n in (2.16), calibration of R_o, and any inflation/localization are all absent, yet these control the results in §3.3. Without these details the noise-robustness and assimilation conclusions cannot be reproduced or independently assessed; the GitHub repository should also be cited with a URL.","section":"§3.2 setup; §3.3.1; Table 3"},{"comment":"All validation is on synthetic RK4 trajectories generated from (2.4) itself, with a known linear observation operator and Gaussian noise. This is appropriate for a method stress test, but the abstract's framing around 'a common constraint in real-world epidemiological surveillance' overstates what is demonstrated. Either temper the abstract/conclusion to a proof-of-concept claim, or add at least one experiment with realistic surveillance features — underreporting, reporting delays, non-Gaussian or aggregated (weekly) counts, or a historical CHIKV outbreak dataset.","section":"Abstract; §3; §4"}],"minor_comments":[{"comment":"The Supporting Information lists 'Appendix B. Discovered SINDy equations; Appendix C. Numerical results ... partial and noisy data,' but the actual layout has them reversed (Appendix B contains the partial-data figure, Appendix C the equation table). Please fix the cross-referencing.","section":"Supporting information / Appendices B–C"},{"comment":"After Eq. (2.24): 'where the observation operator G is assumed to be linear' — G is undefined; presumably H. Also, u^k_{m+1|m}, k=1,...,K for an ensemble of J members clashes with K already denoting the state dimension.","section":"§2.3, Eq. (2.24)"},{"comment":"The observed subset is described as 'vaccinated (V_h), infectious (I_h), treated (J_h), and recovered-under-treatment (T_h),' but per Table 1 J_h is asymptomatic infectious and T_h is treated. Please correct the labels.","section":"§3.3.1"},{"comment":"It is unclear what the plotted curves represent (per-compartment NRMSE/correlation? averages over host and vector groups?) — no legend or per-compartment identification is given.","section":"Figure 7 caption"},{"comment":"'eps 0.2' appears without subscript — state whether eps_1 = eps_2 = eps_3 = 0.2, which sits oddly with the text's description of eps_i as distinct reduction factors.","section":"Table 3"},{"comment":"Several panels have unlabeled y-axes and awkwardly placed 10^4 multipliers; row/column organization is hard to parse.","section":"Figures 8–10"},{"comment":"'we systematically add clean trajectories with state-dependent Gaussian noise' — grammatical error (presumably 'add ... to'). In Eq. (2.7), the state matrix is written with rows x_1^T,...,x_K^T although there are m time points and K states — the indexing is inconsistent.","section":"§3.2; Eq. (2.7)"}],"recommendation":"major_revision","confidential_remarks":"The methodological novelty over EKF-SINDy [23, 24] is modest — the contribution is the epidemiological application and the noise-sensitivity study. The Appendix C / §3.1 discrepancy (Major 1) is serious enough that I would want the authors to re-verify the clean-data experiment end-to-end, not just relabel a table. Several 2025–2026 references could not be independently verified; the editor may wish to spot-check."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a solid methods demo on synthetic CHIKV trajectories—SINDy under graded noise, then EnKF with four observed host channels reconstructing six unobserved states—not a validated surveillance tool, and the equation-recovery prose overreaches what the appendices show.\n\nWhat is actually new is narrow but real. SINDy, EnKF-in-epi, and EKF–SINDy hybrids already exist (they cite Rosafalco). Pairing a SINDy surrogate as the EnKF forecast operator on a partially observed host–vector compartmental system is less covered. They build a detailed ten-compartment CHIKV ODE, map SINDy failure as noise rises, and show that with frequent enough assimilations the filter pulls both observed and unobserved compartments back toward truth (Figs. 7–10). That trajectory-level story is coherent and useful for people who build hybrid discovery–assimilation pipelines.\n\nCredit where due: the noise ladder is systematic, partial-observability design matches real surveillance constraints, and they are honest that standalone SINDy collapses under moderate noise. Table 6’s large RMSE drops versus open-loop noisy SINDy are real numbers for that comparison.\n\nSoft spots, in proportion. First, the library is constants, linears, and bilinears; the true forces of infection are rational in Nh and Nv. So SINDy cannot recover (2.4) even from clean data—only a polynomial proxy. Appendix C’s clean-data system has dVh = 0 and dTh = 0, missing Iv terms in the host infection equations, and a spray of spurious quadratics. That contradicts §3.1’s claim of structural and parametric recovery “to within numerical precision.” Trajectory match under tiny λ is not equation recovery. Second, at 25% noise the forecast operator used for Table 6 is structurally wrong; the >98% “reduction” is mostly EnKF stopping open-loop divergence when four channels are observed often. The Ih row already shows the filter can slightly hurt a good compartment. Third, λ is hand-tuned per noise level against ground truth, train and test are the same synthetic run, and there is no real outbreak series and almost no baseline beyond open-loop SINDy. GitHub is mentioned without a firm reproducibility claim in the text.\n\nWho it is for: computational epi and DA people who want a worked hybrid on a nontrivial arbovirus model. Not for someone needing equation discovery guarantees or operational forecasting evidence.\n\nI would send it to peer review. Referees should force a fix of the recovery claim versus Appendix C, an automatic or cross-validated threshold protocol, a null propagator baseline, and ideally one real-data sketch. Engage if you work on SINDy–filter hybrids; skim the figures if you only need the noise-sensitivity lesson.","headline":"Useful synthetic demo of SINDy-as-forecast plus EnKF on a 10-compartment CHIKV model, but the clean-data “recovery” claim does not match Appendix C and Table 6 partly credits the filter for fixing a bad open-loop model.","tokens_in":31907,"tokens_out":698,"would_cite":false,"duration_ms":20120,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Pairing sparse equation discovery with ensemble filtering recovers Chikungunya host–vector trajectories from noisy, partial surveillance-style data.","keywords":["Chikungunya virus","SINDy","Ensemble Kalman Filter","data assimilation","infectious disease dynamics","partial observability","computational epidemiology"],"falsifier":"Apply the full pipeline to a held-out real CHIKV surveillance series (only the reported host compartments, with realistic noise and gaps) and check whether unobserved states and short-horizon forecasts stay within the error levels claimed on synthetic 25% noise data; systematic failure or non-physical reconstructed mosquito trajectories would falsify the central claim.","tokens_in":31532,"feed_emoji":"🦟","tokens_out":911,"duration_ms":22166,"temperature":0.7,"pith_summary":"Mechanistic models of Chikungunya spread need parameters and full compartment data that real surveillance rarely supplies. This paper shows that sparse identification of nonlinear dynamics (SINDy) can rediscover the governing mass-action equations from clean epidemic trajectories, but breaks down once observational noise is added, inventing spurious terms and distorting forecasts. Embedding the SINDy-derived model as the forecast operator inside an Ensemble Kalman Filter lets noisy partial observations correct the state and fill in unobserved compartments, including mosquito states never reported. In synthetic tests with only a few host compartments observed at 25% noise, the hybrid cut root-mean-square error by more than 98% for most states relative to noisy SINDy alone. The practical stake is a data-driven path to interpretable epidemic models that still work when surveillance is incomplete and noisy.","feed_headline":"Filter plus sparse discovery rebuilds hidden mosquito states","feed_subtitle":"SINDy alone fails under noise; coupling it to EnKF cuts most compartment errors over 98% in synthetic CHIKV tests.","key_machinery":"Hybrid SINDy–EnKF: a sparsely identified bilinear ODE model serves as the EnKF forecast operator, while the filter’s ensemble covariance assimilates noisy partial observations and transfers information to unobserved compartments.","core_discovery":"Standalone SINDy recovers the CHIKV compartmental equations from noise-free data but fails under moderate-to-high noise; coupling that same SINDy model to an Ensemble Kalman Filter restores accurate trajectories for both observed and unobserved host and vector compartments under partial, noisy observations, with reported RMSE reductions above 98% for most of the ten states in the 25%-noise experiments.","pith_inferences":["If compartment-wise or scale-weighted sparsity thresholds replace a single global λ, the standalone SINDy failure mode on large vector populations may shrink before filtering is applied.","The same cross-covariance mechanism should transfer to other arboviruses with analogous host–vector mass-action structure (e.g., dengue, Zika) without redesigning the library class.","Real underreporting is often multiplicative and delayed, not additive Gaussian; testing that noise model is a natural next falsification step the paper leaves open.","Online updating of SINDy coefficients inside the filter would blur offline discovery and assimilation, potentially reducing dependence on the initial noisy identification."],"forward_implications":["Interpretable CHIKV models can be learned and corrected online from partial noisy surveillance rather than fully specified a priori parameters.","Unobserved mosquito compartments can be inferred from host observations when host–vector couplings are retained in the forecast model.","Observation frequency bounds filter skill: high-frequency assimilation keeps NRMSE low; long gaps let SINDy model error dominate.","The same hybrid pattern extends, on the authors’ account, to weak-form or ensemble SINDy, Neural-ODE/PINN residuals, and joint state–parameter updates.","Outbreak monitoring and short-term forecasting become feasible targets once the framework is validated on real outbreak records."],"fun_headline_variants":["SINDy-EnKF hybrid rebuilds hidden CHIKV states from noisy data","Sparse discovery plus filter cuts CHIKV errors over 98%","EnKF steadies SINDy to recover unobserved mosquito compartments","Hybrid SINDy-EnKF tracks partial noisy Chikungunya dynamics","Noise-proof SINDy via EnKF restores full CHIKV trajectories"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that a hand-tuned sparsity threshold and a candidate library built from the same mass-action structure that generated the synthetic data produce a forecast model good enough for the filter—without real outbreak data or an automatic threshold rule.","fun_headline_variants_meta":{"raw":{"variants":["SINDy-EnKF hybrid rebuilds hidden CHIKV states from noisy data","Sparse discovery plus filter cuts CHIKV errors over 98%","EnKF steadies SINDy to recover unobserved mosquito compartments","Hybrid SINDy-EnKF tracks partial noisy Chikungunya dynamics","Noise-proof SINDy via EnKF restores full CHIKV trajectories"]},"model":"grok-4.5","effort":"low","cost_usd":0.003277,"raw_usage":{"total_tokens":1102,"prompt_tokens":725,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":32768000,"prompt_tokens_details":{"text_tokens":725,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":299,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":725,"tokens_out":78,"duration_ms":5720,"temperature":1.0,"reasoning_tokens":299,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T11:31:40.993212+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Apply the full pipeline to a held-out real CHIKV surveillance series (only the reported host compartments, with realistic noise and gaps) and check whether unobserved states and short-horizon forecasts stay within the error levels claimed on synthetic 25% noise data; systematic failure or non-physical reconstructed mosquito trajectories would falsify the central claim.","supporting_citations":[],"review_version":1}