{"id":"f63980b5-e0ca-4b6f-809d-4d207db3136d","arxiv_id":"2607.19241","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Replacing enthalpy with an ideal-gas temperature/density estimate as neural-network input improves prediction of real-fluid T, ρ, ψ in supercritical combustion surrogates by factors up to 14.5.","lead":"A neural-network surrogate for real-fluid thermodynamics in supercritical combustion becomes more accurate when the raw enthalpy input is replaced by a simple physical estimate: an ideal-gas temperature or density tailored to each output. The paper shows this input reparameterization cuts prediction error by 1.5–7.5x on held-out data and by up to 14.5x on a new flame, while adding almost no computational cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unseen-strain-rate transfer factors rest on unquantified coverage of the flame manifold by the hand-specified augmentation envelope; exact bounds and coverage analysis are missing.","rationale":"The paper's controlled comparison is well designed: identical architectures, training data, and optimization isolate the effect of the input coordinate; the cross-reparameterization controls support the causal interpretation; and the limitations (no inline coupling, single fuel, no extrapolation beyond the envelope) are stated explicitly. The held-out RMSE reductions and the monotonic projections in Fig. 5 are internally consistent and do not depend on the augmentation procedure being perfect. However, the transfer-to-unseen-strain-rate claim is a headline result, and it hinges on the assertion that the a=3000 flame is inside the augmented thermodynamic envelope. The text provides no quantitative evidence for this: the augmentation bounds, the power-law exponent, and the species-perturbation details are not given, and the unseen flame is not shown in the coverage plots. This is precisely the reader's weakest assumption, and it is load-bearing for the transfer factors. The concern does not invalidate the central methodological claim about TAIR's benefit on the held-out distribution, but it does warrant the CONDITIONAL verdict: the paper should either provide the augmentation specification and a coverage analysis or soften the transfer claim. Since the reader already recommended CONDITIONAL, no verdict change is needed.","tokens_in":14396,"tokens_out":6537,"duration_ms":74313,"concrete_test":"Request the exact augmentation bounds and power-law exponent (or code) from the authors. Simulate the a=3000 s^-1 flame and compute, for each state, the local density of the augmented training set in the full (h,p,Y) or (T,p,Y) space, e.g., k-nearest-neighbor distance to augmented training samples. Overlay the unseen flame on Fig. 3. If >95% of flame states fall within high-density regions of the augmented distribution, the in-envelope claim holds; if a substantial fraction lie in low-density tails, the transfer factors are out-of-distribution comparisons and the paper must be revised. A secondary check: retrain using only a=1000 and a=10000 flames (excluding 5000) and test at 3000; if the raw-to-TAIR gap collapses, the result is sensitive to envelope details.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central transfer claim (Sec. III D) reports reductions of 3.6x, 14.5x, and 6.0x on an 'unseen' a=3000 s^-1 flame. These numbers appear in the abstract and are load-bearing for the generalization claim. Their validity depends entirely on the assertion that this flame lies 'within the augmented thermodynamic envelope.' However, the augmentation procedure (Sec. II C) is described only qualitatively: temperature and pressure are 'randomly varied within prescribed bounds' and species mass fractions are perturbed by a 'common power-law exponent,' with the bounds and exponent not specified. The paper does not provide any quantitative coverage measure—such as the local density of augmented samples at the unseen flame's states—and Fig. 3 does not overlay the a=3000 flame. If the hand-chosen augmentation fails to reproduce the joint correlations among h, p, and Y present in the actual flame, the augmented training set may have low density on the test-flame manifold. In that case, the large raw-input errors at 3000 s^-1 could be an out-of-distribution artifact, and the improvement factors would overstate TAIR's benefit. The held-out test-set improvements (1.5x, 2.0x, 7.5x) would remain valid, but the transfer generalization claim—a headline contribution—would be unjustified. The missing specification also makes the experiment irreproducible as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes target-aligned input reparameterization (TAIR) for neural surrogates of real-fluid thermodynamic closure in supercritical combustion. Instead of feeding raw enthalpy h to three MLPs predicting temperature T, density ρ, and compressibility ψ, TAIR uses an ideal-gas estimated temperature T̃ for the temperature network and an ideal-gas density ρ̃ for the density and compressibility networks, while retaining pressure p and composition Y. These transformations are explicit, use only solver-available state and species constants, and are invertible at fixed p and Y, so no information is lost. On a 1.2M-sample augmented database from supercritical methane–oxygen counterflow flames at 100 bar, TAIR reduces held-out RMSE by factors of about 1.5, 2.0, and 7.5 for T, ρ, and ψ relative to a raw-input baseline, and by factors of about 3.6, 14.5, and 6.0 on an unseen a=3000 s⁻¹ flame. Target-inconsistent cross-reparameterization controls perform worse, supporting the claim that the benefit comes from target-matched input design rather than generic preprocessing. The paper also reports a closure-level speed-up of about 89× relative to the iterative Cantera reference.","tokens_in":14758,"tokens_out":4137,"duration_ms":47069,"significance":"The methodological idea is simple, thermodynamically motivated, and potentially general: if a target property admits a cheap ideal-gas or reduced-order approximation, that approximation can serve as an input coordinate, leaving the network to learn only the departure. The controlled comparison—identical architecture, data, and training across raw-input, TAIR, and cross-reparameterization configurations—is well designed and directly isolates the effect of input reparameterization. The explicit statement of invertibility and the use of target-inconsistent controls are commendable and rule out the most obvious circularity concerns. If the quantitative claims are confirmed with full specification of the data-generation protocol and uncertainty quantification, this would be a useful contribution to neural thermodynamic surrogates for reacting-flow simulations.","major_comments":[{"comment":"The unseen-strain-rate transfer factors (3.6×, 14.5×, 6.0×) are load-bearing for the generalization claim and depend entirely on the assertion that the a=3000 s⁻¹ flame lies 'within the augmented thermodynamic envelope.' However, the augmentation procedure is not specified quantitatively: no bounds on temperature and pressure, no power-law exponent for species perturbation, and no acceptance/rejection thresholds are given. Fig. 3 does not overlay the a=3000 flame, and no coverage measure (e.g., local density of augmented samples on the unseen flame manifold, nearest-neighbor distances, or kernel density ratio) is provided. As reported, the experiment is not reproducible, and the improvement factors could reflect out-of-distribution artifact rather than a TAIR-specific benefit. Please specify all augmentation constants and provide quantitative coverage diagnostics for the a=3000 state dis","section":"§II C, §III D, Abstract"},{"comment":"Two numerical constants required for exact reproducibility are missing. First, the relative pressure perturbation ε used in the reference ψ evaluation is never given; because ψ is a supervised target, this value directly affects the reported RMSE numbers. Second, the augmentation 'prescribed bounds' and 'common power-law exponent' are referenced but not defined. These are not cosmetic details: the ψ target and the training distribution are determined by them. The authors should report ε and the full augmentation parameter set, including the ranges and the exponent, as part of the experimental setup.","section":"§II A, Eq. (10)–(11); §II C"},{"comment":"All RMSE values are single-run point estimates with no repeated-seed statistics. Given stochastic optimization (random initialization, data shuffling, augmentation randomness) and the single data partition, the improvement factors—especially the 14.5× density improvement on the unseen flame—need uncertainty quantification to establish that the rankings are robust. I recommend reporting mean ± standard deviation (or minima/confidence intervals) over at least 5–10 independent training runs per configuration, with the same seeds across configurations for paired comparisons.","section":"§III C, §III D, §II D"}],"minor_comments":[{"comment":"Fig. 3 shows only 2D projections with point clouds; the local density of augmented samples is not visible. A density plot or histogram overlay would help assess coverage, especially for the unseen flame.","section":"§II C / Fig. 3"},{"comment":"The transfer result is reported at a single time instant (t = 7×10⁻⁴ s). Since the flame is transient, a single snapshot may not represent the full manifold. Reporting error statistics over multiple snapshots or time-averaged fields would strengthen the transfer claim.","section":"§III D"},{"comment":"The notation for the constant heat capacity at T0 is introduced as c̊_p,i in the text but the equation appears to use a slightly different symbol. Please unify notation.","section":"§II B, Eq. (14)"},{"comment":"The sensitivity analysis selects 'one representative database state in each temperature bin,' but the binning procedure is not described (number of bins, selection rule). Please specify.","section":"§IV B / Fig. 9"}],"recommendation":"major_revision","confidential_remarks":"The core controlled comparison is well executed and the negative controls are persuasive. The main risk is the transfer claim resting on an unquantified and incompletely specified augmentation envelope. If the authors provide the missing constants, coverage diagnostics, and seed variance, the paper would likely be publishable at a good venue. I do not see a circularity problem, as the input coordinates are derived from solver-available state and not from the targets."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core claim holds up. Replacing h with an ideal-gas-derived temperature for the T-network and an ideal-gas density for the rho- and psi-networks is a cheap, physically motivated preprocessing step, and the controlled comparison—same architecture, same data, only the first input coordinate changes—is exactly the right way to test it. The cross-reparameterization controls are the key piece: they show the gain is not just \"any analytic input\" but target-matched coordinates. Held-out RMSE reductions of 1.5–7.5x and the large compressibility improvement are credible, and the projected input–target relations in Fig. 5 support the mechanism.\n\nThe soft spots are where the reader and stress-test point. The unseen-strain-rate results (3.6x, 14.5x, 6.0x) are load-bearing for the generalization claim, and they rest on the assertion that the a=3000 flame lies inside the \"augmented thermodynamic envelope.\" The augmentation is described only qualitatively: no numerical bounds, no power-law exponent, no quantitative coverage measure of the unseen flame. If the hand-chosen augmentation misses the joint correlations among h, p, and Y in the real flame, those transfer factors could be optimistic. That doesn't invalidate the held-out comparison, but it does mean the abstract's strongest numbers are not fully supported as reported. The other gaps are standard: no code/data/artifacts, no repeated-seed variance, and a single fuel system. The paper is honest about its own limits—inline coupling, extrapolation, and differential consistency are all flagged explicitly.\n\nThis is a paper that deserves a serious referee. The idea is simple enough to be broadly applicable, and the controlled study is the right shape. The main demands in revision should be: specify the augmentation procedure, add a coverage analysis for the unseen flame, report variance across seeds, and release the training data or code. If those are provided, the transfer claim goes from plausible to solid. I would cite this if I worked on neural surrogates for real-fluid closure, and I would bring it to a reading group as a clean example of input design versus architecture changes.","headline":"Controlled study shows a cheap thermodynamic input reparameterization helps on held-out data; the transfer claim needs better evidence about coverage of the unseen flame.","tokens_in":15220,"tokens_out":1879,"would_cite":true,"duration_ms":20794,"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":"Neural surrogates for real-fluid thermodynamics become substantially more accurate when their leading input coordinate is matched to the target property: replacing raw enthalpy with an ideal-gas temperature estimate for T and an ideal-gas d","keywords":["supercritical combustion","real-fluid thermodynamics","neural-network surrogate","input reparameterization","enthalpy-based closure","Peng–Robinson equation of state","compressibility coefficient","target-aligned input reparameterization"],"falsifier":"Compute the TAIR-versus-raw-input RMSE separately for test points where the compressibility factor Z deviates most from unity (strongly non-ideal states). If the TAIR advantage disappears or reverses on these points, the claim that TAIR guides networks to learn real-fluid departures would be limited to states near the ideal-gas baseline.","tokens_in":14265,"feed_emoji":"🔥","tokens_out":4172,"duration_ms":41174,"temperature":0.7,"pith_summary":"The paper argues that the learnability of neural surrogates for real-fluid thermodynamic closure in supercritical combustion depends less on network capacity than on which thermodynamic coordinate is supplied as the primary input. In the enthalpy-based pressure-correction formulation, the closure maps (h, p, Y) to (T, ρ, ψ), but raw enthalpy is a poorly informative coordinate for density and compressibility because it conflates caloric and volumetric information. The authors propose target-aligned input reparameterization (TAIR): for the temperature network, replace h with T̃, the temperature obtained by inverting a constant-c_p ideal-gas enthalpy approximation; for the density and compressibility networks, use ρ̃ = p/(R_m T̃), the ideal-gas density. These algebraic transformations are free of iteration and preserve the solver-available state, leaving the networks to learn only the real-fluid departure from an ideal-gas baseline. On supercritical methane–oxygen counterflow flame data, TAIR reduces held-out RMSE by factors of about 1.5, 2.0, and 7.5 for T, ρ, and ψ, with even larger gains at an unseen strain rate, while target-inconsistent cross-reparameterization performs worse—evidence that the benefit comes from thermodynamically matched input design rather than generic preprocessing.","feed_headline":"Thermo-matched inputs cut neural property error up to 7.5x","feed_subtitle":"Ideal-gas T and rho coordinates improve supercritical flame surrogate accuracy at ~90x speedup.","key_machinery":"The central object is the TAIR coordinate transformation: T̃ = T0 + (h − Σ_i Y_i h_f,i) / (Σ_i Y_i c_p,i) and ρ̃ = p/(R_m T̃), both computed algebraically from solver-available variables and species constants. The paper also derives the ideal-gas identity (∂ρ̃/∂p)_{h,Y} = ρ̃/p, which motivates supplying ρ̃ to the compressibility network: since pressure is retained as an input, the pair (ρ̃, p) carries the leading ideal-gas baseline for the isenthalpic density response that ψ represents. The transformation functions as a thermodynamic preconditioner, exposing the dominant ideal-gas dependence of each target and leaving the networks to model only the real-fluid departure from that baseline.","core_discovery":"The paper establishes that a simple, physically motivated reparameterization of the input coordinate can substantially improve neural prediction of real-fluid properties. For the temperature network, replacing raw enthalpy h with an estimated ideal-gas temperature T̃—obtained by setting h equal to a mixture enthalpy with constant heat capacities—reduces the regression to a near-monotonic, weakly nonlinear map. For the density and compressibility networks, replacing h with the ideal-gas density estimate ρ̃ = p/(R_m T̃) aligns the leading pressure and composition dependence of the target, making the learned map close to linear over most of the sampled range. The cross-reparameterization contro","pith_inferences":["The paper's results suggest a general 'baseline-preconditioning' design rule for neural surrogates: wherever a cheap physical approximation exists for a target, feed that approximation as an input so the network learns the residual. This likely applies beyond thermodynamics to chemical kinetics and turbulence-related regressions.","A testable extension would be to apply TAIR to other equations of state (e.g., SRK or multi-parameter) and other fuel–oxidizer pairs; the benefit should scale with how much of the target variance the ideal-gas baseline captures, and should shrink in strongly non-ideal near-critical regions.","The large improvement for ψ hints that finite-difference-derived quantities are especially hard to learn from raw enthalpy; since the paper does not enforce differential consistency between independently predicted ρ and ψ, a natural next step is to use the relation (∂ρ/∂p)_h,Y = ψ as a regularization or consistency loss.","The transfer test is a priori only; running TAIR surrogates inline in the pressure-correction solver would test whether the accuracy advantage survives error accumulation, pressure–density feedback, and excursions outside the training envelope."],"forward_implications":["If TAIR's accuracy gains hold, neural thermodynamic surrogates can achieve the same closure error with smaller networks, freeing capacity for other parts of reacting-flow surrogates.","The accuracy improvement transfers to an unseen strain-rate flame within the augmented thermodynamic envelope, suggesting the method is useful for off-design operating conditions without retraining.","TAIR adds negligible computational overhead (about 8.8% over the raw-input network) while preserving a roughly 90-fold speedup over iterative real-fluid closure, making it directly compatible with expensive CFD solvers.","The target-matching principle is general: any target with an inexpensive analytical baseline could be reparameterized similarly, potentially extending to transport properties and other closure quantities."],"fun_headline_variants":["Thermo-matched inputs cut neural error up to 7.5x","Ideal-gas coordinates boost neural surrogates 7.5x in supercritical flames","Physics-informed input reparameterization slashes neural property error","TAIR: inputs aligned with thermodynamics improve neural accuracy 7.5x","Neural surrogates gain 7.5x accuracy via target-matched thermodynamic inputs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The training data are generated by random augmentation in (T, p, Y) space around flame states at 1000, 5000, and 10000 s⁻¹, and the paper assumes this augmented distribution covers the thermodynamic states encountered in the unseen 3000 s⁻¹ flame; if the augmentation bounds miss correlations among h, p, and Y that occur in that flame, the reported transfer improvements could be optimistic.","fun_headline_variants_meta":{"raw":{"variants":["Thermo-matched inputs cut neural error up to 7.5x","Ideal-gas coordinates boost neural surrogates 7.5x in supercritical flames","Physics-informed input reparameterization slashes neural property error","TAIR: inputs aligned with thermodynamics improve neural accuracy 7.5x","Neural surrogates gain 7.5x accuracy via target-matched thermodynamic inputs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000539,"raw_usage":{"total_tokens":2479,"prompt_tokens":860,"completion_tokens":1619,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":1518}},"tokens_in":604,"tokens_out":1619,"duration_ms":17248,"temperature":1.0,"reasoning_tokens":1518,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T12:58:31.368568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the TAIR-versus-raw-input RMSE separately for test points where the compressibility factor Z deviates most from unity (strongly non-ideal states). If the TAIR advantage disappears or reverses on these points, the claim that TAIR guides networks to learn real-fluid departures would be limited to states near the ideal-gas baseline.","supporting_citations":[],"review_version":1}