{"id":"60dce601-66b3-423e-9f57-14304784e1d8","arxiv_id":"1908.05297","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A data-driven linear correction fitted by least squares stabilizes low-dimensional reduced order models of quasi-geostrophic ocean flow, and adding a dissipativity constraint improves accuracy at very low dimensions.","lead":"This paper tests a data-driven correction reduced order model (DDC-ROM) and a physically constrained version (CDDC-ROM) on a quasi-geostrophic ocean circulation model. For low-dimensional models, both outperform the standard Galerkin reduced order model, with the constrained version most accurate at very small dimensions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'CDDC-ROM provides the best results' claim is supported only by in-sample Table 1; the predictive Table 2 omits CDDC-ROM, so the paper's own out-of-sample setting does not test the second half of the central claim.","rationale":"The reader's weakest_assumption is the linear ansatz in Eq. (21). I agree that this is a limitation for extrapolation, but it is not the first thing that threatens the central claim as stated. The claim has two conjuncts: (i) DDC/CDDC beat G-ROM for low r; (ii) CDDC is best. Conjunct (i) is supported out-of-sample by Table 2 for DDC (though not at r=45 and r=50, where G-ROM is better), so the linear ansatz is at least empirically adequate on the tested trajectory. Conjunct (ii) is supported only by Table 1, which is in-sample. The predictive Section 3.3.3, which is the paper's main argument against overfitting, omits CDDC entirely. A reader cannot tell whether CDDC's advantage persists when trained on a shorter interval. Since the abstract and conclusion present CDDC-best as a general finding, this missing evidence is more directly load-bearing than the form of the ansatz. The concrete test is simply to include CDDC in the existing Table 2 protocol; this is a run-level check, not a new theory. If CDDC remains best out-of-sample, the linear ansatz is the main remaining risk; if not, the paper's headline needs qualification. Either way, the reader's CONDITIONAL verdict is appropriate, and the additional condition is to report CDDC-ROM in the predictive experiments.","tokens_in":75,"tokens_out":5292,"duration_ms":351584,"concrete_test":"Run Table 2's two predictive cases (training on [10,45] and [10,35], testing through t=80) with the CDDC-ROM, using the same SVD tolerance and m=3r settings reported in Section 3.2, and report the time-averaged streamfunction relative error (26) alongside the existing G-ROM and DDC-ROM entries. If CDDC-ROM is not the most accurate at least in the low-r regime (r <= 20), the abstract's 'CDDC-ROM provides the best results' should be narrowed to the in-sample setting. As a secondary check, compute the relative errors using only times t > t_p^* to exclude any transient overlap with the training interval.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: DDC-ROM and CDDC-ROM outperform G-ROM for low-dimensional ROMs, and CDDC-ROM provides the best results. The first part receives genuine out-of-sample support in Section 3.3.3 (Table 2), where DDC-ROM beats G-ROM for r up to 40 when trained on [10,45] or [10,35] and tested on [10,80]. The second part, however, is supported only by Table 1, which is in-sample: the POD basis and the least-squares operator \\tilde{A} in Eq. (22) are constructed from snapshots on [10,80], and the error metric (26) is evaluated on the same interval. The predictive experiments in Table 2 do not report CDDC-ROM results at all. Thus the paper's headline conclusion that 'CDDC-ROM provides the best results' has not been tested in the paper's own predictive setting, and the factor-2000 improvement quoted in Section 4 is an in-sample number. This is load-bearing because Section 3.3.3 is specifically designed to demonstrate behavior beyond the training interval; omitting CDDC-ROM from that demonstration leaves the strongest part of the abstract unsupported. The linear ansatz (21) is a related generalization concern, but it is secondary: even a misspecified linear closure could beat G-ROM on the tested trajectory, so the decisive missing evidence is empirical.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript investigates the data-driven correction reduced order model (DDC-ROM) and its physically constrained variant (CDDC-ROM) for the quasi-geostrophic equations with double-gyre wind forcing. The DDC-ROM adds a data-fitted linear correction term \\tilde{A} a to the standard Galerkin ROM, and the CDDC-ROM enforces a negative-semidefiniteness constraint on \\tilde{A} in Eq. (25). The numerical study compares G-ROM, DDC-ROM, and CDDC-ROM at dimensions r = 5 to 50 using two metrics: kinetic energy time series (Section 3.3.1) and the relative L2 error of the time-averaged streamfunction defined in Eq. (26) (Section 3.3.2). It also reports predictive experiments (Section 3.3.3) in which the ROMs are trained on [10,45] or [10,35] and simulated through t=80. The central claim is that, for low-dimensional ROMs, DDC-ROM and CDDC-ROM both outperform G-ROM and that CDDC-ROM provides the best results.","tokens_in":13826,"tokens_out":5993,"duration_ms":58050,"significance":"The QGE test problem is substantially more challenging than the Burgers and cylinder-flow benchmarks used in prior DDC-ROM work, so demonstrating stabilization and error reduction on this problem is a useful contribution. The paper's strengths include the systematic sweep over r, the explicit constrained formulation in Eq. (25), and the inclusion of predictive experiments for the DDC-ROM in Table 2. However, the paper's strongest advertised conclusion, that CDDC-ROM provides the best results, rests entirely on in-sample Table 1, while the predictive Table 2 omits CDDC-ROM. In addition, the hyperparameters of the method are selected on the same DNS interval used for the main evaluation. These gaps are load-bearing for the abstract's central claim and should be addressed before the paper can be accepted.","major_comments":[{"comment":"The central claim that CDDC-ROM provides the best results is supported only by in-sample evidence. In Table 1, the POD basis and the operator \\tilde{A} in Eq. (22) are constructed from snapshots on [10,80], and the error metric (26) is evaluated on the same interval. The predictive experiments in Table 2, in which training on [10,45] or [10,35] is followed by simulation through t=80, do not report CDDC-ROM results at all. Thus the second half of the abstract's claim has not been tested outside the training interval, and the 'factor of about 2000' improvement quoted in Section 4 is also an in-sample number. Please add CDDC-ROM to the predictive experiments, or explicitly qualify the CDDC-ROM ranking as in-sample only.","section":"Section 3.3.2 (Table 1), Section 3.3.3 (Table 2), Eq. (26)"},{"comment":"The truncated SVD tolerance is chosen 'with a tolerance that yields the most accurate results,' and m=3r is chosen because 'numerical experiments suggest' it balances accuracy and efficiency. These are free parameters selected on the same DNS interval used to evaluate Table 1, so part of the reported advantage may reflect tuning to the test interval. The paper should either fix these choices a priori, provide a sensitivity analysis of Tables 1 and 2 to these parameters, or use a nested validation procedure so that the reported errors are not selected post hoc.","section":"Section 3.2 (ROM construction)"}],"minor_comments":[{"comment":"In Case I at r=50, the DDC-ROM error (2.09e-01) is larger than the G-ROM error (1.67e-01), so the statement that 'the DDC-ROM is significantly more accurate than the G-ROM' should be qualified to the range r <= 45, or to 'small and moderate r values.'","section":"Section 3.3.3, Table 2"},{"comment":"The statement that modeling the commutation error 'does not significantly change' the DDC-ROM and CDDC-ROM results is not supported by any table, figure, or quantitative comparison; a brief numerical summary would improve reproducibility.","section":"Section 3.2"},{"comment":"In Eq. (21), the notation mixes a generic function g(\\omega_r) with a vector component (\\tilde{A} a)_i; please clarify whether g denotes a vector-valued operator or whether the right-hand side should be interpreted componentwise.","section":"Section 2.3, Eq. (21)"},{"comment":"The replacement of (\\omega_R,\\psi_R) by (\\omega_m,\\psi_m) with m=3r is mentioned only briefly; since this is a modeling choice with a direct effect on the learned correction, a sentence explaining why m=3r is sufficient would be helpful.","section":"Section 3.2"},{"comment":"There is a typo in the text: 'Furrthermore' should be 'Furthermore.'","section":"Section 3.3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a numerical investigation in the journal's scope, and the DDC-ROM results in Table 2 are encouraging. The main issue is that the CDDC-ROM ranking is untested in the predictive setting, and the method's free parameters are selected on the evaluation interval. Both issues are fixable in a revision by adding CDDC-ROM to the predictive table and by documenting the parameter-selection procedure more rigorously. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead Mou et al. on data-driven correction ROMs for the QGE. Bottom line: worth your time if you work on ROM closure for geophysical turbulence, but the abstract overstates what is actually demonstrated.\n\nWhat is genuinely new: this is the first application of DDC-ROM/CDDC-ROM to the quasi-geostrophic double-gyre, which is a harder test than the Burgers/cylinder cases in the earlier papers. The paper shows the correction term can stabilize a 10-mode Galerkin ROM that otherwise blows up, and the time-averaged streamfunction errors drop by orders of magnitude. The out-of-sample experiments in Section 3.3.3 are the real contribution: training on [10,35] or [10,45] and simulating through t=80, the DDC-ROM beats G-ROM at essentially every r, and the error stays low beyond the training interval. That is a legitimate predictive demonstration. The finding that the dissipativity constraint helps at small r and hurts at large r is a useful piece of practical knowledge.\n\nThe soft spots are concentrated on the CDDC-ROM part of the claim. Table 1, which shows CDDC-ROM as best at r=5-20, is fully in-sample: same [10,80] interval for POD basis, least-squares fit, and error metric. The factor-2000 number in the conclusions is that in-sample number. Table 2, the out-of-sample test, omits CDDC-ROM entirely. So the abstract's sentence 'the CDDC-ROM provides the best results' is not supported by the paper's own predictive setting. The stress-test note is correct on this.\n\nOther issues are minor in comparison. The SVD tolerance is chosen 'with a tolerance that yields the most accurate results' and m=3r is chosen from numerical experiments, with no sensitivity analysis. No error bars, no comparison to the stabilized ROM from San & Iliescu [41], which is the natural baseline for this test problem. The linear ansatz (21) is acknowledged as a modeling choice with higher-order alternatives left for future work, which is honest but means we don't know how much of the success depends on that particular closure form.\n\nOn balance, this is a solid numerical study with one clear strength (out-of-sample DDC-ROM) and one clear weakness (missing out-of-sample CDDC-ROM). I'd send it to peer review but require: include CDDC-ROM in the predictive tables, report the SVD tolerance and m values, add sensitivity runs, and compare against the stabilized ROM. Ask for code/data too. The paper can be fixed without a rewrite.\n\nRecommendation: serious referee, major revision.","headline":"A useful numerical application of a known closure idea to the quasi-geostrophic double-gyre, where the out-of-sample DDC-ROM results are convincing but the headline CDDC-ROM claim only has in-sample support.","tokens_in":14353,"tokens_out":2204,"would_cite":true,"duration_ms":21052,"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 data-fitted linear correction stabilizes low-dimensional quasi-geostrophic ocean-circulation ROMs.","keywords":["reduced order model","data-driven correction","quasi-geostrophic equations","proper orthogonal decomposition","ROM closure","physical constraints","ocean circulation","Galerkin ROM"],"falsifier":"Train the DDC-ROM on $[10,35]$ for a quasi-geostrophic run at a higher Reynolds number where eddy feedback is stronger, then integrate through $t=80$ and compare the least-squares residual of Eq. (22) on the unseen interval $[50,80]$; if the residual stays as large as the correction itself while the streamfunction error returns to G-ROM levels, the linear ansatz is the limiting factor.","tokens_in":13273,"feed_emoji":"🌊","tokens_out":7478,"duration_ms":68527,"temperature":0.7,"pith_summary":"The paper asks whether a small, data-fitted correction can rescue low-dimensional reduced-order models of the quasi-geostrophic double-gyre ocean circulation, a flow whose active spatial scales are far too many for a ten-mode Galerkin truncation. It answers yes for this test case: at $r=10$ the standard Galerkin ROM (G-ROM) has a time-averaged streamfunction relative error of $2.06\\times 10^2$, the data-driven correction ROM (DDC-ROM) cuts this to $3.25\\times 10^{-1}$, and the physically constrained CDDC-ROM improves it to $9.58\\times 10^{-2}$. The improvement persists when the correction is trained on a shorter interval, $[10,35]$ or $[10,45]$, and then integrated through $t=80$. The paper concludes that, at the low resolutions where the G-ROM blows up or locks onto the wrong statistical state, learning the missing unresolved-mode interactions from data is a practical route to stable and accurate ROMs.","feed_headline":"Learned correction stabilizes low-mode ocean-circulation models","feed_subtitle":"At 10 modes, the data-fitted term cuts the mean-flow error from about 200 to 0.1.","key_machinery":"The load-bearing object is the correction term $\\mathrm{Correction}\\approx (\\tilde A a)_i$, the difference between projecting the full-rank dynamics $f(\\omega_R)$ and the resolved dynamics $f(\\omega_r)$ onto the first $r$ modes. The paper determines $\\tilde A$ by minimizing the least-squares residual in Eq. (22) over DNS training snapshots, using truncated SVD to control ill-conditioning and replacing $\\omega_R$ with an $m$-mode approximation with $m=3r$ for efficiency. The constrained variant solves Eq. (25) with the dissipation constraint $a^\\top \\tilde A a\\le 0$. What carries the argument is that the correction enters only the linear part of the ROM, $\\dot a = b + (A+\\tilde A)a + a^\\top B a$, so it stabilizes the most violent instabilities without altering the resolved nonlinear interactions.","core_discovery":"The central claim is that the effect of the unresolved POD modes on the resolved modes can be captured, for low-dimensional quasi-geostrophic ROMs, by a linear data-driven term added to the Galerkin system. Starting from the exact identity $\\dot a_i = (f(\\omega_r),\\phi_i) + [(f(\\omega_R),\\phi_i)-(f(\\omega_r),\\phi_i)]$, the paper closes the system with the ansatz $\\mathrm{Correction}\\approx (\\tilde A a)_i$ and fits $\\tilde A\\in\\mathbb{R}^{r\\times r}$ to DNS snapshots through the least-squares problem in Eq. (22). Requiring $a^\\top \\tilde A a\\le 0$ in the fit yields the CDDC-ROM, whose learned linear operator is dissipative. The paper reports that the CDDC-ROM is the most accurate ROM for $5\\le r\\le 20$, the DDC-ROM is the most accurate for $25\\le r\\le 50$, and the G-ROM is consistently worse across the whole range; at $r=10$ the CDDC-ROM error is more than three orders of magnitude below the G-ROM error.","pith_inferences":["The fact that a purely linear closure produces such large gains suggests that, at these truncations, the dominant missing interaction acts like a linear damping or transport mechanism; quadratic or higher-order closures would matter most in regimes with stronger nonlinear eddy feedback.","Because the error plateau at larger $r$ is blamed on rough basis functions, smoothing or regularizing the vorticity POD modes may be a more direct fix than enlarging the ROM once the correction is in place.","The dissipation constraint helps exactly where the uncorrected ROM is most unstable; a relaxed or data-adaptive constraint might preserve the low-$r$ stabilization without the overdamping seen at $r\\ge 25$.","A natural transfer test is to train the correction at one Rossby or Reynolds number and evaluate at nearby parameter values; if the linear operator must be retrained for every regime, the practical value of the method is limited to fixed-parameter settings."],"forward_implications":["At $r=10$ the DDC-ROM reduces the time-averaged streamfunction error by a factor of about 600 relative to the G-ROM; the CDDC-ROM reduces it by about 2000.","At $r=10$ and $r=15$ the corrected ROMs keep kinetic energy close to the DNS range, while the uncorrected G-ROM grows to unphysical values.","The CDDC-ROM is the best choice for heavily truncated models ($5\\le r\\le 20$), while the DDC-ROM is better for larger dimensions because the dissipation constraint tends to overdamp.","Training the correction on $[10,35]$ or $[10,45]$ and simulating through $t=80$ still gives large gains over the G-ROM, so the learned linear term is not simply memorizing the training window.","For all ROMs the errors plateau at large $r$ rather than vanishing, which the paper attributes to roughness in the higher vorticity basis functions."],"supporting_citations":[{"why":"analyzes the commutation error that the paper chooses to neglect, justifying the form of the correction term.","marker":"[22]"},{"why":"introduces the physically constrained data-driven correction ROM whose dissipation constraint is used in Eq. (25).","marker":"[32]"},{"why":"supplies the QGE DNS benchmark, POD snapshot generation, and the four-gyre time-averaged streamfunction used as the main metric.","marker":"[41]"},{"why":"provides the numerical setup and validation for the barotropic ocean circulation model at the Reynolds and Rossby numbers used here.","marker":"[44]"},{"why":"introduces the original DDC-ROM formulation, the least-squares parameter fitting, and the truncated-SVD treatment of ill-conditioning.","marker":"[53]"}],"fun_headline_variants":["Data-driven correction outperforms Galerkin in QG ROMs","Learned ROM cuts low-mode ocean error by 3 orders","Constrained DDC-ROM best for low-dimensional QG model","Corrected ROMs beat standard model in QG simulation","Data-fitted term stabilizes low-mode ocean-circulation ROMs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction assumes the correction is linear in the resolved coefficients ($\\mathrm{Correction}\\approx \\tilde A a$); if the true unresolved-mode feedback is strongly nonlinear in those coefficients, the learned term cannot represent it and the reported gains may not generalize.","fun_headline_variants_meta":{"raw":{"variants":["Data-driven correction outperforms Galerkin in QG ROMs","Learned ROM cuts low-mode ocean error by 3 orders","Constrained DDC-ROM best for low-dimensional QG model","Corrected ROMs beat standard model in QG simulation","Data-fitted term stabilizes low-mode ocean-circulation ROMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000866,"raw_usage":{"total_tokens":3742,"prompt_tokens":922,"completion_tokens":2820,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":2731}},"tokens_in":538,"tokens_out":2820,"duration_ms":21437,"temperature":1.0,"reasoning_tokens":2731,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:18:30.147612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the DDC-ROM on $[10,35]$ for a quasi-geostrophic run at a higher Reynolds number where eddy feedback is stronger, then integrate through $t=80$ and compare the least-squares residual of Eq. (22) on the unseen interval $[50,80]$; if the residual stays as large as the correction itself while the streamfunction error returns to G-ROM levels, the linear ansatz is the limiting factor.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"analyzes the commutation error that the paper chooses to neglect, justifying the form of the correction term."},{"cited_title":"Mohebujjaman, L","cited_arxiv_id":null,"evidence_quote":"introduces the physically constrained data-driven correction ROM whose dissipation constraint is used in Eq. (25)."},{"cited_title":"San and T","cited_arxiv_id":null,"evidence_quote":"supplies the QGE DNS benchmark, POD snapshot generation, and the four-gyre time-averaged streamfunction used as the main metric."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the numerical setup and validation for the barotropic ocean circulation model at the Reynolds and Rossby numbers used here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the original DDC-ROM formulation, the least-squares parameter fitting, and the truncated-SVD treatment of ill-conditioning."}],"review_version":1}