{"id":"0d81b398-6a34-4852-a000-07375b260a39","arxiv_id":"2603.29079","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":6,"one_line_summary":"SINDy on Hankel-DMD coefficients recovers supercritical and subcritical dynamo normal forms that extrapolate better than weakly nonlinear analysis, including unstable branches and non-analytic nonlinearities.","lead":"Researchers recover sparse equations for stellar magnetic cycles from simulation data using DMD plus SINDy, and show these models often beat classic weakly nonlinear analysis far from onset, including on stiff and subcritical dynamos. The result matters because full stellar MHD is still too expensive for realistic parameters, so reliable reduced models are a practical route to cycle prediction.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"The claim that SINDy recovers a legitimate reduced-order model of the PDE rests on heavy filtering and library constraints that may isolate a smoothed signal rather than the true slow amplitude.","rationale":"The reader correctly isolates the weakest link: the pipeline that turns raw PDE snapshots into the two coefficients on which SINDy is trained. The paper itself documents that unrestricted libraries produce spurious terms and that filtering is required to suppress quadratic oscillations that otherwise cause overfitting (§§3.2, 4.1, 6.1). Because the headline robustness claims (far-from-onset saturation, unseen unstable branches, non-analytic regimes) all rest on the fidelity of those two coefficients, any material change under the proposed ablation would directly undercut the strongest claim. The existing Newton–Krylov checks and low integration errors are reassuring near onset but do not substitute for an unfiltered, unconstrained control. The concern therefore reinforces rather than overturns the CONDITIONAL verdict; a public repository that lets others run the ablation would convert the condition into acceptance for the 1-D setting.","tokens_in":26638,"tokens_out":644,"duration_ms":5743,"concrete_test":"Re-run the full (D,κ) SINDy pipeline of §6.2 on the identical snapshot set but with Savitzky–Golay filtering disabled (or window reduced to ≤50) and with the unrestricted fifth-order library that includes κ^{2} and κD products; if the recovered coefficients change by more than ~20 % relative to Eq. (31), or if the unstable-branch amplitudes in Fig. 5b deviate from the Newton–Krylov solutions by more than the present error, the claim that the models are robust reduced-order models of the PDE (rather than of the filtered signal) is weakened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Abstract; §§5–7; Figs. 4–5) is that constrained SINDy on Hankel-DMD coefficients yields models more robust than WNL normal forms, including recovery of unseen unstable subcritical branches and applicability when the nonlinearity 1/(1+κ_{2}B^{2}) is non-analytic. That claim is load-bearing on the assumption that the leading pair of Hankel-DMD modes, after Savitzky–Golay filtering (window 200–300), radius normalization, and a deliberately restricted library that omits κ^{2} and κD products (§§3.1–4.1, 6.1–6.2), still isolates the same slow amplitude H that WNL expands. Without those steps the regression overfits or fails to converge (explicitly noted for polar libraries and unrestricted fifth-order libraries). The recovered equations therefore risk being an accurate description of the pre-processed two-mode signal rather than a faithful reduced-order model of the full PDE; the Newton–Krylov agreement near onset does not fully rule this out far from onset or for stiff κ.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper develops a data-driven pipeline (Hankel DMD with adjoint modal coefficients, followed by SINDy) to recover reduced-order amplitude equations for oscillatory dynamos from a 1-D mean-field α–Ω model with magnetic-diffusion quenching. The recovered models are compared systematically to weakly nonlinear (WNL) normal forms at third and fifth order. For fixed (D,κ), for D-dependent families, and for a joint (D,κ) model, the authors report that SINDy equations integrate more accurately against DMD trajectories than WNL forms away from onset, recover unstable subcritical branches (cross-checked by Newton–Krylov), and remain usable when the quenching nonlinearity is stiff. Explicit coefficient tables, Pareto plots, bagging/constrained-SR3 tests, and bifurcation diagrams (Figs. 3–5, Tables 1–2) support the comparisons.","tokens_in":26941,"tokens_out":1516,"duration_ms":17516,"significance":"If the claims hold, the work offers a practical route to reduced dynamo-cycle models that do not require a full WNL expansion and that can be trained on simulation (or, in principle, observational) time series. Strengths that raise the contribution above a pure methods demonstration include: (i) side-by-side coefficient and error comparisons with analytically derived WNL forms; (ii) Newton–Krylov confirmation of unstable branches never present in the training trajectories; (iii) an adjoint-based extraction of DMD coefficients that improves isolation of the leading pair; and (iv) explicit tests of robustness (bagging, constrained SR3, polar vs Cartesian libraries). These make the paper a useful benchmark for data-driven reduced modeling of stellar cycles, even if the present demonstration remains on a 1-D mean-field system.","major_comments":[{"comment":"Abstract and §7 claim that SINDy finds equations 'where the nonlinearity is not analytic and WNL analysis cannot be applied.' In §7 the authors themselves choose κ₂=5·10⁻³ so that |B|<1/√κ₂∼14 remains empirically satisfied, and they defer 'stiff subcriticalities' to future work. The abstract claim should be narrowed to what is actually demonstrated (better extrapolation within a regime where the Taylor expansion is still marginally valid), or an additional experiment at larger κ₂ (where WNL fails) should be added.","section":"Abstract; §7"},{"comment":"The central identification pipeline (§3.1–4.1, 6.1–6.2) relies on Savitzky–Golay filtering (windows 200–300), radius/κ normalization, and deliberate library restrictions that omit κ² and κD products. Without these steps the authors report overfitting or non-convergence (polar libraries; unrestricted fifth-order libraries). The manuscript should quantify how much of the reported superiority over WNL survives under weaker preprocessing (e.g., no SG filter, or only Hankel delay), and state more clearly that the recovered ODEs are models of the filtered two-mode signal whose fidelity to the full PDE is validated mainly near onset by Newton–Krylov and by integration error on the same filtered coefficients.","section":"§3.1–4.1; §6.1–6.2"},{"comment":"For the subcritical case at fixed D (§4.2, Table 2, Eq. 24), SINDy systematically prefers a combination of r⁴ and r⁵ damping over the pure fifth-order term of the WNL form (22), even under bagging and constrained SR3. The paper treats this as acceptable because integration error remains <1%. That is fine for local prediction, but it weakens the claim that SINDy recovers the 'correct' normal-form structure. Either reframe the claim as structure-agnostic predictive accuracy, or show that the spurious r⁴ term vanishes under a cleaner isolation of H (e.g., exact adjoint-eigenvector projection without DMD).","section":"§4.2; Table 2; Eq. (24)"}],"minor_comments":[{"comment":"Eq. (28) contains the placeholder '[recalculate!]' in the manuscript text; this must be removed and the numerical bounds finalized before publication.","section":"§5.2; Eq. (28)"},{"comment":"Figure 1 caption refers to panel (a) as 'y=5/x' and then describes (a)–(d) inconsistently with the body text; renumber and clarify which panels show A, B, eigenvectors, and the DMD spectrum.","section":"Figure 1"},{"comment":"In §5.1 the Cartesian model is referred to once as (27) when the displayed equations are (26); check cross-references for (25)–(31) throughout §5–6.","section":"§5.1"},{"comment":"The abstract and introduction mention Hankel DMD, but the title and some early paragraphs say only 'DMD'; keep the terminology consistent (Hankel/HODMD vs standard DMD).","section":"Abstract; §1; §3.1"},{"comment":"Table 1 and the WNL appendices are valuable; a short statement of how the adjoint eigenproblem and solvability integrals were discretized in Dedalus would aid reproducibility.","section":"Table 1; Appendices A–B"},{"comment":"Typographical: 'a a grid-based resolution' (§2.1); 'mildly supercritical regime' used for a subcritical κ case (§4.2); 'intencity' in Fig. 5 caption.","section":"§2.1; §4.2; Fig. 5"}],"recommendation":"minor_revision","confidential_remarks":"The work is a solid methods paper on a well-chosen 1-D testbed; the overclaim on non-analytic nonlinearities is the main abstract-level issue and is easily fixed. Scope is appropriate for an astrophysical/fluid-dynamics journal that publishes reduced modeling. No concerns about novelty disclosure or citation pattern."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing: they recover sparse ODEs for a 1-D α–Ω dynamo that track saturation amplitudes better than classical weakly nonlinear normal forms once you leave the neighborhood of onset, and they recover unstable subcritical branches that never appear in the training data. That is concrete and useful inside reduced-model work.\n\nWhat is actually new is not DMD or SINDy, but the systematic side-by-side with a full fifth-order WNL expansion, the adjoint route to modal coefficients, and a joint (D, κ) model that switches supercritical to subcritical. They show the work: coefficient tables, Pareto fronts, integration errors of order 1–3%, and Newton–Krylov confirmation of the unstable branch. The claim that SINDy still produces usable equations when the quenching 1/(1+κ₂B²) is non-analytic—so WNL’s Taylor expansion fails—is the part that matters for more realistic dynamos.\n\nSoft spots, in proportion. The pipeline is heavy: Hankel delay, Savitzky–Golay windows of 200–300 points, radius and κ normalization, and a library deliberately stripped of κ² and κD products. Without those steps the regression overfits or fails; the paper says so. You are partly fitting a cleaned two-mode signal rather than raw PDE data. That does not kill the near-onset agreement or the branch checks they run, but it does mean “discovery of the PDE’s normal form from data” is a bit strong. The PDE is still a 1-D mean-field model with hand-chosen coefficients; no 3-D test and no public code. Mild circularity from knowing the target form is present but not fatal—the coefficients are fit to independent runs and the unstable branch is a genuine out-of-sample inference.\n\nWho it is for: people already building reduced dynamo models or applying SINDy to oscillatory fluids. Not yet a general stellar-cycle method. Math, data comparisons, and citation pattern look solid. It deserves a serious referee; desk rejection would be wrong. I would engage with it if I were writing reduced cycle models; I would not treat it as finished for 3-D stellar dynamos. Send to peer review; expect requests for code and a clearer statement of how much the filtering buys.","headline":"Solid 1-D demo that constrained SINDy on Hankel-DMD amplitudes can outperform fifth-order WNL on subcritical and stiff dynamos, with real checks—but heavy filtering and a toy PDE keep the claim local.","tokens_in":27570,"tokens_out":588,"would_cite":true,"duration_ms":14461,"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":"Data-driven equations for stellar dynamo cycles outperform classic weakly nonlinear analysis, even far from onset and on unseen unstable branches.","keywords":["stellar dynamo","mean-field dynamo","Dynamic Mode Decomposition","SINDy","weakly nonlinear analysis","subcritical bifurcation","reduced-order model","α-Ω dynamo"],"falsifier":"Apply the same DMD–SINDy pipeline to a dynamo whose nonlinearity is strongly non-analytic (large κ2) or to a three-dimensional convective dynamo, and check whether the recovered equations still reproduce saturation amplitudes and the correct supercritical-to-subcritical transition when compared with direct numerical solutions outside the training window.","tokens_in":27487,"feed_emoji":"☀️","tokens_out":646,"duration_ms":5943,"temperature":0.7,"pith_summary":"Stars like the Sun generate oscillating magnetic fields that drive activity cycles and space weather, yet full magnetohydrodynamic simulations are too expensive and multi-scale to be practical for many purposes. Reduced-order models that capture the essential cycle dynamics are therefore valuable. This paper shows that a combination of Hankel Dynamic Mode Decomposition (to extract the coherent magnetic wave) and Sparse Identification of Nonlinear Dynamics (to fit a sparse polynomial ODE to its amplitude) recovers the governing amplitude equations for a canonical one-dimensional mean-field dynamo directly from simulation snapshots. The recovered models are compared systematically with the classic weakly nonlinear normal forms obtained by asymptotic expansion about the onset of dynamo instability. The data-driven equations remain accurate far beyond the weakly nonlinear regime, correctly predict saturation amplitudes when the underlying nonlinearities are stiff or non-analytic, and even reconstruct unstable subcritical solution branches that never appear in the training data. The result is a practical route to cycle models that can be extrapolated in dynamo strength and magnetic quenching parameters without requiring an analytic expansion.","feed_headline":"Data-driven dynamo equations beat classic analysis far from onset","feed_subtitle":"SINDy models recover saturation, subcritical branches, and non-analytic regimes from snapshots alone","key_machinery":"Hankel Dynamic Mode Decomposition followed by adjoint projection and Savitzky–Golay filtering yields a clean complex amplitude time series for the primary dynamo wave; constrained SINDy then identifies a sparse polynomial ODE for that amplitude as a function of the dynamo number D and the magnetic-dissipation parameter κ.","core_discovery":"Sparse polynomial models fitted by SINDy to the leading Hankel-DMD modal coefficients of a one-dimensional mean-field dynamo are more robust than the corresponding weakly nonlinear normal forms: they correctly predict magnetic saturation amplitudes far from onset, recover both stable and unstable subcritical branches, and remain usable when the nonlinearity is non-analytic so that weakly nonlinear analysis cannot be applied.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["SINDy dynamo models beat WNL far from onset","Data-driven equations recover dynamo saturation and subcritical branches","Sparse models from DMD-SINDy predict magnetic states where WNL fails","SINDy yields robust oscillatory dynamo equations beyond analytic regimes","Data-only SINDy recovers stiff nonlinear dynamo cycles and unstable branches"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The leading pair of filtered Hankel-DMD modes is assumed to isolate the same slow amplitude that weakly nonlinear theory expands, so that a sparse polynomial fit on those two coefficients is a legitimate reduced model of the full system rather than an over-fit to the filtered signal.","fun_headline_variants_meta":{"raw":{"variants":["SINDy dynamo models beat WNL far from onset","Data-driven equations recover dynamo saturation and subcritical branches","Sparse models from DMD-SINDy predict magnetic states where WNL fails","SINDy yields robust oscillatory dynamo equations beyond analytic regimes","Data-only SINDy recovers stiff nonlinear dynamo cycles and unstable branches"]},"model":"grok-4.5","effort":"low","cost_usd":0.005432,"raw_usage":{"total_tokens":1512,"prompt_tokens":862,"num_sources_used":0,"completion_tokens":90,"cost_in_usd_ticks":54320000,"prompt_tokens_details":{"text_tokens":862,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":560,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":862,"tokens_out":90,"duration_ms":5385,"temperature":1.0,"reasoning_tokens":560,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T15:55:36.872351+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Apply the same DMD–SINDy pipeline to a dynamo whose nonlinearity is strongly non-analytic (large κ2) or to a three-dimensional convective dynamo, and check whether the recovered equations still reproduce saturation amplitudes and the correct supercritical-to-subcritical transition when compared with direct numerical solutions outside the training window.","supporting_citations":[],"review_version":1}