{"id":"b91f898f-354d-4aab-a49c-19c3faadcfcb","arxiv_id":"2607.17376","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A neural-network eddy-viscosity closure trained by adjoint-based optimization inside an LES solver improves mean-field and resolved-turbulence accuracy relative to Smagorinsky for premixed jet flames across Damköhler numbers.","lead":"This paper trains a neural-network turbulence model directly inside a large-eddy simulation of turbulent premixed flames, using the simulation itself to compute training gradients. On a planar jet test case it reports 25-50% better mean fields and over 60% better resolved turbulence statistics than a Smagorinsky baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported improvements are measured against a constant Cs=0.3 Smagorinsky baseline in one planar-jet configuration; without a dynamic/well-tuned baseline or a second geometry, the 25–50% and >60% gains and 'broadly applicable' conclusion are not established.","rationale":"The paper is a credible extension of DPM to reacting LES; the adjoint-based training is well-motivated, and the reported loss histories are internally consistent. The central claim, however, is an improvement claim relative to a baseline. That claim is only meaningful if the baseline is a fair reference. The constant Cs=0.3 choice is never justified against standard practice for jet flows, and the paper does not compare with dynamic Smagorinsky or other established SGS models. This is not a matter of consensus—it is a correctness risk: an over-dissipative baseline inflates all reported gains. The same issue underlies the out-of-sample generalization claim: if the baseline is poor in all regimes, a network that learns to reduce dissipation will appear to 'generalize.' The paper's own disclosure that adjoint optimization may implicitly correct baseline deficiencies (Sec. 6.1) reinforces this. The abstract's blanket 25–50% primitive improvement is also in tension with Table 2 (JP = −11% at Das=0), so the quantitative headline needs qualification. I therefore agree with the reader's conditional verdict; the concern does not require a different verdict, but it identifies the specific experiment that would strengthen or falsify the claim.","tokens_in":12683,"tokens_out":9809,"duration_ms":113577,"concrete_test":"Recompute Table 2's improvements for ME20-PRvv at Das=0, 20,000, and 35,000 using a dynamic Smagorinsky closure (or a properly tuned lower-Cs baseline) in the same LES solver. If the reported primitive-variable and stress/flux gains shrink to within noise or fall below roughly 10–15%, the headline percentages are an artifact of the Cs=0.3 comparator and the broad-applicability claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is calibrated against a single Smagorinsky baseline with Cs=0.3 (Eq. 7). This is a high-dissipation comparator: for free-shear/jet flows, standard and dynamic Smagorinsky formulations typically yield effective Cs well below 0.3 locally. Because the ME network is initialized as a small correction to this baseline (outputs scaled by 1e-5, Sec. 3) and is trained to match fDNS, the reported 25–50% primitive-variable and >60% stress/flux improvements may largely measure removal of excess baseline dissipation rather than closure superiority. This is compounded by the single temporally evolving planar jet, single-step chemistry, and the inclusion of the absolute time t as a network input (Eq. 8), which allows the model to key onto the specific training window (t=14.985–16.425). The paper's own Table 2 shows JP=−11% at Das=0 for the favored ME20-PRvv model, contradicting the abstract's unqualified '25–50%' primitive-variable improvement. Thus the evidence does not yet support the broad-applicability conclusion until a more competitive baseline and a second configuration are tested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a solver-in-the-loop (adjoint/DPM) training framework for neural-network subgrid-scale closures in LES of a temporally evolving planar turbulent premixed jet flame. The closure augments a Smagorinsky baseline with an MLP that predicts an eddy viscosity (and a variant that predicts SGS fluxes directly), with training by adjoint-based gradients matching filtered DNS data over optimization windows. Several objective functions (P, PRvv, PSF), training strategies (serial, parallel-window, parallel-data), and Damköhler-number training conditions are compared. The best model, ME20-PRvv, is reported to improve primitive-variable errors by 25–50% and resolved Reynolds stress and scalar-flux errors by more than 60% relative to the Smagorinsky baseline, including out-of-sample Damköhler-number cases.","tokens_in":12992,"tokens_out":4863,"duration_ms":55064,"significance":"If the quantitative claims hold, this is a useful methodological advance: it demonstrates that PDE-consistent, adjoint-trained neural closures can be extended from non-reacting LES and RANS to LES of premixed flames, and it systematically compares loss formulations and training strategies. The cost analysis and GPU implementation are practical strengths, and the component-wise tables provide more information than a single aggregate metric. However, the broad-applicability conclusion and the headline percentage improvements are not yet fully supported because all results come from a single configuration with a single, possibly weak, comparator and without uncertainty quantification. The core idea is credible and worth publishing after strengthening the evidence and re-scoping the claims.","major_comments":[{"comment":"The only comparator is the constant-coefficient Smagorinsky model with Cs=0.3. This is a particularly dissipative baseline for jet/free-shear flows; standard dynamic or locally adjusted Smagorinsky coefficients are typically lower. Since the ME network is initialized as a small correction to this baseline and trained against filtered DNS, the reported 25–50% and >60% improvements may be dominated by removal of excess baseline dissipation rather than by a genuinely superior closure. The central 'broadly applicable framework' claim therefore requires comparison against a stronger baseline (e.g., dynamic Smagorinsky or a well-tuned constant-Cs model) and, ideally, an established reacting-flow SGS closure.","section":"Sec. 2, Eq. (7); Sec. 6"},{"comment":"All results are single-realization point estimates from one temporally evolving planar jet over the short interval t = 14.985–16.425 (Sec. 2). No uncertainty quantification, multiple initial conditions, or repeated training runs are reported. Because the optimization and evaluation involve chaotic LES trajectories, the specific percentage improvements are not established as statistically robust. At minimum, the authors should report ensemble/seed sensitivity or otherwise justify that the reported improvements are not within run-to-run variability.","section":"Sec. 6, Table 2; Figs. 2–5"},{"comment":"The abstract claims the best model 'improves a posteriori errors in the LES primitive variables by 25–50%', but Table 2 reports JP = -11% for ME20-PRvv at Das = 0, 29% in-sample, and 41% at Das = 35,000. The claim is therefore not accurate as stated; it should be qualified by case and by loss component. This is a load-bearing inconsistency because it directly concerns the headline quantitative result.","section":"Abstract and Table 2"},{"comment":"The out-of-sample Damköhler-number tests are still within the same temporal planar-jet configuration, same Re/Ma, same filter width, and same single-step chemistry; only Das is varied. Moreover, the network receives the absolute time t as an input and is trained on the window t = 14.985–16.425, so it can in principle key onto the specific training interval. This limits the strength of the 'broadly applicable framework' conclusion. I recommend either adding a second configuration/initial condition (or retraining/testing on an interval not containing the training t window) or re-scoping the conclusions to the same configuration across Damköhler numbers.","section":"Secs. 2 and 6.3, Eq. (8)"}],"minor_comments":[{"comment":"The vertical axis label 'JPRvv ×10^4' is ambiguous; specify whether this is the cumulative objective (13) or a single-window loss and give the normalization used for the percentage-improvement calculations.","section":"Sec. 6.3, Fig. 5"},{"comment":"CPU inference time is 642 s versus 100.5 s for Smagorinsky, a sixfold overhead, while GPU overhead is modest; the text should avoid implying the overhead is always modest without noting the CPU regime.","section":"Sec. 7"},{"comment":"The inclusion of Das and t as inputs is a design choice with potential generalization consequences. An ablation removing t, or at least a discussion of why the model does not simply memorize the training interval, would strengthen the out-of-sample interpretation.","section":"Eq. (8)"},{"comment":"The statement that the serial-window strategy is 'the most robust training configuration' is based on one reacting case. A caveat noting the limited evidence would be appropriate.","section":"Sec. 6.1"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is well-founded: the weak constant-Cs=0.3 baseline, the single planar-jet configuration, and the lack of UQ mean the headline quantitative claims are not yet established. The paper is technically sound enough to warrant revision rather than rejection. I would not necessarily require a second geometry for acceptance if the conclusions are re-scoped and the baseline is strengthened, but at least a dynamic-Smagorinsky comparison and run-to-run or ensemble sensitivity are needed to support the stated percentages."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Closure training inside the LES solver is the right direction. This is the first application of the solver-in-the-loop DPM approach to reacting LES, and the paper does a genuinely careful job of comparing closure forms, training strategies, and objective functions. The loss curves and component-wise tables make the in-sample improvements plausible, and the authors are candid that the direct-closure model is unstable and that the unresolved reaction source is left unmodeled. Those are real disclosures.\n\nThe stress-test note lands. The baseline is Smagorinsky with Cs=0.3, a high-dissipation choice for a planar jet; a dynamic or lower-Cs comparator could absorb a large chunk of the reported gains. All results come from one temporally evolving planar jet with single-step chemistry. The network receives absolute time t as an input, which lets it memorize the training window, and the Damköhler number is given explicitly, so 'generalization' is more a labeled interpolation test than an extrapolation test. Table 2 shows JP=-11% at Das=0 for the preferred model, so the abstract's unconditional '25-50%' phrasing is wrong. There is no UQ, no code or data release, and the loss weights are empirically tuned.\n\nNone of this sinks the paper as a proof-of-concept. The method is well explained, the serial-vs-parallel and objective-function comparisons are valuable for the community, and the authors are honest about failures. The main fix is to align claims with evidence: either test a second configuration and a stronger baseline, or qualify the broad-applicability language.\n\nI'd send it to peer review, because it deserves referee time. I'd want the authors to address the baseline and the time input before publication. I'd also cite the training-methodology results if I worked in this area.","headline":"A credible proof-of-concept for solver-embedded closure training in reacting LES, with claims running ahead of the evidence.","tokens_in":13484,"tokens_out":2926,"would_cite":true,"duration_ms":30621,"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 neural-network eddy-viscosity closure trained inside the LES solver reduces resolved-field errors by 25–50% and resolved Reynolds-stress and scalar-flux errors by over 60%, while generalizing across Damköhler numbers.","keywords":["large-eddy simulation","premixed combustion","turbulence closure","neural network","adjoint-based optimization","differentiable programming","Damköhler number generalization","subgrid-scale modeling"],"falsifier":"Run the trained ME20-PRvv closure in a posteriori LES of a spatially developing premixed jet at a Damköhler number outside {0, 20,000, 35,000} and at a higher Reynolds number, comparing against filtered DNS or experimental data; if the reported 25–50% and 60% error reductions do not reproduce or the solver becomes unstable, the generalization claim is falsified.","tokens_in":12525,"feed_emoji":"🔥","tokens_out":4658,"duration_ms":47918,"temperature":0.7,"pith_summary":"The paper tries to establish that a deep-learning turbulence closure for large-eddy simulation of premixed flames can be trained directly inside the solver—through adjoint-based gradients that respect the governing equations—rather than on precomputed data. The best model, a small neural network that augments the standard eddy-viscosity closure, is claimed to cut primitive-variable errors by 25–50% and resolved Reynolds-stress and scalar-flux errors by more than 60% relative to a constant-coefficient baseline. Crucially, the paper claims the learned closure remains stable and accurate at Damköhler numbers it was not trained on, including both the nonreacting and strongly burning limits. If true, this would give combustion LES a route to closures that adapt to turbulence–flame interactions across regimes without retuning per case.","feed_headline":"Neural closure trained inside LES cuts premixed-flame errors by 25-50%","feed_subtitle":"The same model improves resolved Reynolds-stress and scalar-flux errors by more than 60%, even at Damköhler numbers unseen in training.","key_machinery":"The machinery is an augmented eddy-viscosity closure: the baseline Smagorinsky eddy viscosity is retained and a multilayer perceptron (four layers, 100 hidden units, GELU activations, about 35,000 parameters) adds a learned correction, with one output channel for momentum/species transport and one for heat flux. The network is evaluated pointwise using Galilean-invariant local gradients of density, velocity, internal energy, product mass fraction, and pressure, together with the baseline eddy viscosity, the chemical source term, and time. Because the network is embedded in the PDE residual, parameter gradients are obtained by integrating an adjoint equation backward over each optimization wi","core_discovery":"The central claim is that embedding an untrained neural network into the filtered LES equations and fitting it with adjoint-based gradients yields a closure that recovers both mean fields and resolved turbulence statistics in turbulent premixed jet flames across Damköhler-number regimes. The best model, trained with a primitive-variable objective augmented by the flame-normal resolved Reynolds stress (the PRvv variant), improves primitive-variable errors by 25–50% and resolved Reynolds-stress and scalar-flux errors by more than 60% relative to the Smagorinsky baseline, while staying stable at out-of-sample conditions. The paper also finds that a two-output augmented eddy-viscosity form gener","pith_inferences":["Because the demonstration uses one temporally evolving planar jet with single-step chemistry, the same recipe would need validation on more complex geometries and fuels before it can be called broadly applicable; the paper's framing leaves that step implicit.","The success of a small pointwise network with gradient inputs suggests a path toward online adaptive closures that update as the flow regime changes, rather than offline per-case tuning.","The comparison against only one fixed-coefficient Smagorinsky baseline leaves open whether the reported margins would shrink against dynamic or otherwise more advanced subgrid closures; a head-to-head test would be a useful extension.","The unresolved reaction source is not modeled explicitly but is assumed to be absorbed by the learned species-transport closure, so accuracy may degrade in cases where filtered-chemistry error dominates—an implicit limitation worth testing directly."],"forward_implications":["Solver-embedded neural closures can replace or augment constant-coefficient eddy-viscosity models in reacting-flow LES without becoming unstable during deployment.","Including second-order statistics in the training objective is necessary: models trained only on primitive variables leave resolved Reynolds stresses worse than the baseline.","The two-output augmented eddy-viscosity formulation is more robust than a high-dimensional direct-closure formulation, which becomes unstable in reacting regimes.","Serial-window training, where the LES evolves continuously across windows, generalizes better than parallel-window training for long-horizon predictions.","The learned closure maintains accuracy at out-of-sample Damköhler numbers, including the nonreacting limit and the strongly burning thin-flame regime."],"fun_headline_variants":["PDE-consistent neural closure improves LES flame errors 25-50%","Adjoint-trained neural closure cuts premixed-flame LES errors 25-50%","Neural eddy-viscosity model trained in LES generalizes across Damköhler numbers","Solver-in-the-loop neural closure improves premixed flame LES by up to 50%","Neural LES closure beats Smagorinsky in premixed flames, generalizes out-of-sample"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim rests on a single temporally evolving planar premixed jet with one-step chemistry and one fixed Smagorinsky baseline; if that configuration does not represent practical premixed burners, the reported margins and cross-Damköhler generalization may not carry over.","fun_headline_variants_meta":{"raw":{"variants":["PDE-consistent neural closure improves LES flame errors 25-50%","Adjoint-trained neural closure cuts premixed-flame LES errors 25-50%","Neural eddy-viscosity model trained in LES generalizes across Damköhler numbers","Solver-in-the-loop neural closure improves premixed flame LES by up to 50%","Neural LES closure beats Smagorinsky in premixed flames, generalizes out-of-sample"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000983,"raw_usage":{"total_tokens":4013,"prompt_tokens":753,"completion_tokens":3260,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":3147}},"tokens_in":497,"tokens_out":3260,"duration_ms":24179,"temperature":1.0,"reasoning_tokens":3147,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T18:08:52.739952+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained ME20-PRvv closure in a posteriori LES of a spatially developing premixed jet at a Damköhler number outside {0, 20,000, 35,000} and at a higher Reynolds number, comparing against filtered DNS or experimental data; if the reported 25–50% and 60% error reductions do not reproduce or the solver becomes unstable, the generalization claim is falsified.","supporting_citations":[],"review_version":1}