{"id":"badbe9de-d360-43ce-b9c5-b808cc6c0781","arxiv_id":"1908.03620","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A quadratic reduced-order model learned in specific-volume and molar-concentration coordinates predicts single-injector combustion dynamics beyond the training interval at five to six orders of magnitude lower cost.","lead":"This paper learns a fast, low-dimensional surrogate model for a simulated single-injector combustor by fitting a quadratic model in transformed variables (specific volume and molar concentrations) to high-fidelity CFD snapshots. The model runs five to six orders of magnitude faster than the CFD solver and appears to stay accurate for two milliseconds beyond its training window.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Beyond-training evidence supports only pressure traces; species/temperature extrapolation rests on an untested quadratic fit to non-polynomial chemistry terms.","rationale":"The reader's weakest assumption is the same that I find most load-bearing: the quadratic ROM is an acknowledged approximation of non-polynomial combustion physics, and the beyond-training evidence does not cover the variables named in the central claim. The paper's own Appendix B and the conclusion explicitly state that the quadratic model form is an approximation, so this is not a manufactured objection. The numerical results genuinely demonstrate good pressure prediction and plausible training-end field agreement, but the claim that temperature, velocity, and species are accurately predicted 200% past training is not supported by the presented figures. This is an evidence gap, not a demonstrated failure, so the conditional verdict remains appropriate; a single quantitative check on the 2 ms test interval for all variables would either confirm or refute the strongest claim. I agree with the reader's assessment and do not recommend changing the verdict.","tokens_in":20509,"tokens_out":6821,"duration_ms":77344,"concrete_test":"Run the r=29, lambda=3e4 ROM for the full 3 ms test horizon and compute per-time-step error metrics for every state variable against GEMS: E_relative (Eq. 21) for p and T and E_nabs (Eq. 22) for vx, vy, CH4, O2, CO2, H2O, over t in [1,3] ms, at all monitor locations and as full-domain fields at t=3 ms. If the maximum T and CH4 errors over the 2 ms extrapolation exceed the levels shown at the end of training (Figs. 7-16), the headline extrapolation claim fails; if they remain comparable, the quadratic approximation is adequate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim requires that a quadratic ROM in q_L=[p,vx,vy,xi,c1..c4] remains accurate for 2 ms beyond the 1 ms training window for all reported variables. That is exactly where the paper's own derivation is weakest. Appendix B shows the pressure equation contains temperature-dependent enthalpies h_l(T) and their time derivatives, and the species equations contain Arrhenius source terms; a complete polynomial lifting (Appendix A) requires auxiliary variables and algebraic constraints, which the learned ROM does not include. The quadratic model therefore fits these non-polynomial effects to data inside the training window. The numerical evidence that this fit extrapolates is limited to pressure time traces at one monitor location (Figs. 5-6); species fields, velocity fields, and temperature are compared to GEMS only at the last training time step (Figs. 7-16). The abstract's claim that the ROM 'accurately predicts temperature, pressure, velocity, species concentrations... and is predictive 200% past the training interval' is thus supported by pressure alone. If the Arrhenius/enthalpy nonlinearities are not negligible on the 2 ms test interval, the learned quadratic operators will drift for temperature and species even while pressure matches.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a non-intrusive, operator-inference-based reduced-order modeling workflow for a two-dimensional single-injector combustor. The authors choose the learning variables q_L = [p, vx, vy, xi, c1,...,c4], transform GEMS snapshots accordingly, define a POD basis in these variables, and infer a quadratic ROM (Eq. (13)) by solving a regularized least-squares problem (Eq. (20)). The method is applied to 10,000 snapshots covering 1 ms of simulation; reduced orders r=24 and r=29 are used. Results compare ROM pressure traces with GEMS at one monitor location for 3 ms (1 ms training plus 2 ms prediction), field-level errors at the last training time step, integrated species, and CPU timings, reporting five to six orders of magnitude speedup. The paper includes appendices detailing the lifting of chemical source terms and explicitly deriving the pressure and species equations, which are not quadratic in q_L.","tokens_in":20797,"tokens_out":5166,"duration_ms":49960,"significance":"If the claims hold, this is a useful demonstration that operator inference in transformed variables can produce stable low-dimensional ROMs for a reacting flow where classical POD-Galerkin requires hundreds of modes. The paper's explicit appendices are a strength: the authors identify the exact polynomial structure available and, importantly, state where the quadratic form is an approximation. The availability of code links is also positive. The central limitation is that the evidence for genuine extrapolation beyond the training interval is narrow: the held-out quantitative comparison is limited to pressure at one monitor, while temperature, species, and velocity fields are validated only at the end of the training interval. Thus the significance of the claimed 200% predictive capability is not yet fully established.","major_comments":[{"comment":"The abstract's sentence that the ROM 'accurately predicts temperature, pressure, velocity, species concentrations, and the limit-cycle amplitude ... and is shown to be predictive 200% past the training interval' is not supported by the beyond-training evidence shown in the paper. The only quantities compared against GEMS on the 20,000 time steps beyond the 1 ms training window are pressure time traces at a single monitor location (Figs. 5-6). Figures 7-16 compare the other variables only at the last training time step (K=10000), and the integrated species in Fig. 8 are not accompanied by an error measure on the prediction interval. The paper should either add quantitative held-out comparisons for temperature, velocity, and species (e.g., errors on a monitor set or over the spatial field during the 2 ms test window) or explicitly restrict the extrapolation claim to pressure.","section":"IV.C, Figs. 5-16 and Abstract"},{"comment":"The authors correctly acknowledge in the Conclusion that 'the quadratic model form is an approximation,' and Appendix B shows why: the pressure equation contains temperature-dependent enthalpies h_l(T) and their time derivatives, and the species equations contain Arrhenius source terms and terms proportional to rho c_l dxi/dx that are not quadratic in q_L. This is load-bearing because the paper motivates the method as exposing exact quadratic structure; the numerical extrapolation claim therefore rests on the empirical smallness of the omitted non-quadratic terms over the 2 ms test window. I ask for a direct assessment of this: for example, compute the residual of the exact GEMS equations in the q_L variables against the learned quadratic operators on the training snapshots, or report time-dependent errors of temperature and species over the test interval. Without such a check, the reader cannot distinguish an incidental fit from the claimed structure exploitation.","section":"Appendix B / Section III.C"}],"minor_comments":[{"comment":"The parenthetical 'K=10000 (which corresponds to t=0.0159999s)' is inconsistent with the stated time step dt=1e-7 and the 1 ms training interval; with these values K=10000 corresponds to t=0.001 s. Please correct the time label or clarify the simulation clock.","section":"IV.C"},{"comment":"These figures are titled 'Predictive results ... at the last time step of training data'; since this time step is inside the training window, the word 'predictive' is misleading and should be replaced by 'reconstructed' or 'training' results.","section":"Figures 9-16"},{"comment":"The paper states that 'another 2ms of testing data at the monitor locations' are available, but it would be helpful to state explicitly how many test snapshots are used and whether any of the test data were used in the L-curve or basis selection.","section":"IV.A / IV.C"},{"comment":"The text accurately says 'many terms in the governing equations take polynomial form,' but the abstract and introduction should avoid implying exact polynomial structure for the full system; consider using 'approximate quadratic structure' consistently throughout.","section":"III.C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest about its approximation, but the abstract overstates the extrapolation evidence. A major revision that either adds held-out field-level and species/temperature validation over the 2 ms test interval or sharpens the claims to pressure-only extrapolation would make the contribution publishable. I do not see a fundamental error in the operator inference methodology itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-executed application of operator inference with a genuinely useful variable transformation (specific volume plus molar concentrations) for a hard combustion problem. The learned ROM is stable at modest dimension (r = 24/29), gives five-to-six-orders-of-magnitude speedups, and beats the r = 200 Galerkin-POD baseline from the authors' earlier work. That is real credit.\n\nThe method is cleanly presented. The transformation to q_L = [p, vx, vy, xi, c1..c4] makes the volume and velocity equations quadratic, and the operator inference least-squares formulation is standard but applied carefully here: redundant Kronecker terms are removed, scaling is handled, and L2 regularization is used to stabilize the fit. The appendices are honest about the main limitation: a complete lifting would require auxiliary variables and algebraic constraints (Appendix A), and the pressure and species equations are not quadratic in q_L because of temperature-dependent enthalpies and Arrhenius factors (Appendix B). The conclusion acknowledges that the quadratic model is an approximation. That honesty strengthens the paper.\n\nThe soft spot is the extrapolation claim. The abstract says the ROM accurately predicts temperature, pressure, velocity, species concentrations, and is predictive 200% past the training interval. The only held-out evidence is pressure traces at one monitor location (Figs. 5–6). The field plots (Figs. 9–16) and integrated species (Fig. 8) are at the last training time step or over the training interval; they do not demonstrate extrapolation of temperature, species, or velocity. Given that the non-quadratic chemistry and enthalpy terms are fit by the quadratic operator, the concern is real: pressure can match while species drift. This is not a fatal flaw, but the abstract should either be softened or backed by held-out error metrics for all reported variables over the full 3 ms.\n\nMinor points: the regularization parameter and basis dimension are selected using the same data, and the L-curve choice is a bit hand-wavy, but that is acceptable for a proof of concept. The speedup claim ignores the 200 h CFD training-data cost, which is fair for a ROM paper but should be stated as offline cost.\n\nWho this is for: ROM practitioners and combustion CFD researchers building non-intrusive surrogates. It deserves a serious referee; a good referee should push for the missing extrapolation evidence and a clearer statement of which variables are validated on held-out data. I would conditionally accept after that revision.","headline":"Operator inference plus specific-volume/molar-concentration variables is a genuinely useful recipe for combustion ROMs, but the paper overstates the extrapolation evidence: only pressure traces are actually shown on the 200% held-out interval.","tokens_in":21271,"tokens_out":3896,"would_cite":true,"duration_ms":40217,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a quadratic reduced-order model, learned from 10,000 CFD snapshots in transformed variables, predicts the single-injector combustor's pressure, temperature, velocity, and species fields accurately, runs five to six…","keywords":["reduced-order modeling","operator inference","proper orthogonal decomposition","combustion dynamics","quadratic structure","variable transformation","non-intrusive model reduction","single-injector combustor"],"falsifier":"Run the learned ROM with the forcing frequency changed from 5 kHz to 4 kHz, or the forcing amplitude changed, without retraining, and compare the predicted pressure phase and limit-cycle amplitude against a new high-fidelity CFD simulation; if the phase or amplitude error grows sharply, the claim that the quadratic form captures the dominant dynamics beyond the training interval is refuted.","tokens_in":20295,"feed_emoji":"🔥","tokens_out":9096,"duration_ms":80500,"temperature":0.7,"pith_summary":"This paper establishes that a low-dimensional quadratic model can reproduce the unsteady dynamics of a single-injector combustor. The method trains on a set of transformed physical variables—pressure, velocities, specific volume, and species molar concentrations—in which much of the governing physics becomes quadratic in form. Operator inference learns the reduced operators from projected snapshot data, without touching the CFD solver, while proper orthogonal decomposition fixes the low-dimensional subspace. The learned 29-mode model matches the CFD pressure trace, temperature, velocity, and species fields, captures the limit-cycle amplitude, and remains predictive for an additional 2 ms of simulation beyond the 1 ms training data. A reliable model of this cost would make repeated combustor evaluations practical for design and uncertainty quantification.","feed_headline":"Quadratic ROM predicts combustor dynamics 200% past training","feed_subtitle":"A 29-mode model reproduces 5,000-Hz pressure phase and runs five orders of magnitude faster than CFD.","key_machinery":"The central object is the quadratic reduced-order model $\\dot{\\hat{q}} = \\hat{A}\\hat{q} + \\hat{F}\\hat{q}^{(2)} + \\hat{B}u + \\hat{c}$ in POD coordinates, where $\\hat{q}^{(2)}$ is the vector of unique quadratic monomials (the collapsed Kronecker product). The enabling mechanism is the variable transformation to $\\mathbf{q}_L$: using specific volume and molar concentrations converts most of the combustion dynamics into quadratic form, which projection preserves exactly. Operator inference—a linear least-squares fit of the reduced operators to projected snapshot data—does the learning, aided by randomized SVD for the large snapshot matrix and L2/Tikhonov regularization to keep the learned operators stable.","core_discovery":"The central discovery is that the single-injector combustor's governing equations expose quadratic structure when written in the learning variables $\\mathbf{q}_L = [p, v_x, v_y, \\xi, c_1, c_2, c_3, c_4]$, where $\\xi = 1/\\rho$ is the specific volume and $c_l$ are the molar concentrations of CH$_4$, O$_2$, CO$_2$, and H$_2$O. The continuity, momentum, and species evolution equations become quadratic in these variables, and this polynomial form is preserved under Galerkin projection. Operator inference then identifies the ROM operators $\\hat{A}, \\hat{F}, \\hat{B}, \\hat{c}$ by solving a regularized least-squares problem from projected snapshot data, without access to the high-fidelity model's implementation. The learned ROM of dimension $r=29$ predicts the pressure trace at the monitor location with accurate phase and a good approximation of the limit-cycle amplitude over a 3 ms simulation (1 ms training plus 2 ms extrapolation), at a computational speedup of five to six orders of magnitude relative to the CFD solver.","pith_inferences":["The paper's own derivations show that the pressure and species equations contain non-polynomial terms (temperature-dependent enthalpies, Arrhenius rates, and mixed density terms) that the quadratic model simply fits away; a natural extension would test how the 2 ms predictive horizon degrades as the operating point moves away from the training conditions.","The same methodology could be tested on a parameterized family of simulations (varying inlet mass-flow rates, forcing frequency, or equivalence ratio) to see whether a single quadratic ROM, or a small library of them, interpolates between operating points; the paper only considers one fixed operating point with one forcing frequency.","Because the approach is non-intrusive, it could be applied to time-resolved experimental measurements (for example, from a laboratory combustor) if field data or suitable sensor projections are available, although the paper uses simulated data only.","The ROM's ability to hold the limit-cycle phase beyond training hints that the learned linear operator $\\hat{A}$ may encode the acoustic eigenstructure of the combustor; extracting its eigenvalues and comparing with linear stability analysis of the full system is a testable check of physical consistency."],"forward_implications":["If the central claim holds, repeated combustion model evaluations for design, optimization, and uncertainty quantification become practical at seconds of compute rather than hundreds of CPU-hours per run.","The 29-dimensional ROM resolves the 5 kHz pressure forcing with correct phase, so it could be embedded in real-time control or active-instability-suppression loops.","The transformed-variable-plus-operator-inference recipe transfers to other reacting-flow or multi-physics systems whose equations can be expressed (or lifted) in polynomial form.","Demonstrating accuracy for 2 ms of pure prediction beyond a 1 ms training window suggests the learned operators encode the dominant nonlinear coupling, at least at the tested operating point and forcing condition."],"supporting_citations":[{"why":"Supplies the operator inference framework that learns reduced operators from snapshot data via least squares.","marker":"[34]"},{"why":"Introduces the transform-and-learn idea of choosing transformed variables to expose polynomial structure before learning.","marker":"[35]"},{"why":"Shows how Arrhenius-type reaction terms can be lifted to polynomial form, guiding the source-term treatment.","marker":"[37]"},{"why":"Prior reduced-order model of the same combustor that required over 100 modes for stability; serves as the baseline this work improves on.","marker":"[11]"},{"why":"Documents the GEMS CFD solver that generated the training and test snapshot data.","marker":"[39]"},{"why":"Defines the single-injector combustor geometry, operating conditions, and chamber-pressure forcing used in the experiment.","marker":"[42]"},{"why":"Provides the one-step methane/oxygen reaction mechanism used for the chemical source terms.","marker":"[43]"},{"why":"Supplies the randomized SVD algorithm used to compute the POD basis for the large snapshot matrix.","marker":"[49]"},{"why":"Provides the L-curve criterion used to select the regularization parameter for the least-squares solve.","marker":"[50]"}],"fun_headline_variants":["Quadratic ROM predicts combustor 200% beyond training","29-mode ROM matches pressure phase, >5 orders faster","ROM learns combustor dynamics without CFD code","Quadratic structure enables 29-mode combustion ROM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a quadratic model in the variables $[p, v_x, v_y, \\xi, c_1, \\ldots, c_4]$ adequately represents the true combustor dynamics over the prediction horizon, even though the governing pressure and species equations contain non-polynomial terms (temperature-dependent enthalpies, Arrhenius factors, and mixed density terms) that the quadratic form simply fits away.","fun_headline_variants_meta":{"raw":{"variants":["Quadratic ROM predicts combustor 200% beyond training","29-mode ROM matches pressure phase, >5 orders faster","ROM learns combustor dynamics without CFD code","Quadratic structure enables 29-mode combustion ROM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001517,"raw_usage":{"total_tokens":6128,"prompt_tokens":1047,"completion_tokens":5081,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":5019}},"tokens_in":663,"tokens_out":5081,"duration_ms":42855,"temperature":1.0,"reasoning_tokens":5019,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:07:17.695856+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the learned ROM with the forcing frequency changed from 5 kHz to 4 kHz, or the forcing amplitude changed, without retraining, and compare the predicted pressure phase and limit-cycle amplitude against a new high-fidelity CFD simulation; if the phase or amplitude error grows sharply, the claim that the quadratic form captures the dominant dynamics beyond the training interval is refuted.","supporting_citations":[{"cited_title":"Data-drivenoperatorinferencefornonintrusiveprojection-basedmodelreduction,","cited_arxiv_id":null,"evidence_quote":"Supplies the operator inference framework that learns reduced operators from snapshot data via least squares."},{"cited_title":"Investigations and Improvement of Robustness of Reduced-Order Models of Reacting Flow,","cited_arxiv_id":null,"evidence_quote":"Prior reduced-order model of the same combustor that required over 100 modes for stability; serves as the baseline this work improves on."},{"cited_title":"Coupling between hydrodynamics, acoustics, and heat release in a self-excited unstable combustor,","cited_arxiv_id":null,"evidence_quote":"Documents the GEMS CFD solver that generated the training and test snapshot data."},{"cited_title":"Experimentalstudyofhigh-frequencycombustioninstabilityinacontinuously variable resonance combustor (CVRC),","cited_arxiv_id":null,"evidence_quote":"Defines the single-injector combustor geometry, operating conditions, and chamber-pressure forcing used in the experiment."},{"cited_title":"Simplified reaction mechanisms for the oxidation of hydrocarbon fuels in flames,","cited_arxiv_id":null,"evidence_quote":"Provides the one-step methane/oxygen reaction mechanism used for the chemical source terms."},{"cited_title":"A randomized algorithm for the decomposition of matrices,","cited_arxiv_id":null,"evidence_quote":"Supplies the randomized SVD algorithm used to compute the POD basis for the large snapshot matrix."},{"cited_title":"The L-curve and its use in the numerical treatment of inverse problems,","cited_arxiv_id":null,"evidence_quote":"Provides the L-curve criterion used to select the regularization parameter for the least-squares solve."}],"review_version":1}