{"id":"4aa5a0df-942f-4abb-8aed-3529d754ec05","arxiv_id":"2507.22370","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A physics-informed neural network reproduces the Runge-Kutta solution of the complex 1D acoustic wave equation in ducts with temperature gradients to about 1e-5 relative error.","lead":"The authors train a neural network to solve the equation for sound waves in a pipe where temperature and flow change along its length, and show that it matches a standard numerical solver. This offers a mesh-free way to compute acoustic fields relevant to gas turbine combustors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The derivation of Eq. (27) relies on the isentropic mean-state relation Eq. (7), which is inconsistent with the paper's own mean-flow profiles; the reported O(10^-5) matches against bvp4c therefore establish only that the network solves Eq. (27), not that Eq. (27) describes the acoustic field.","rationale":"The reader identified the physical accuracy of Eq. (27) as the weakest assumption; my reading agrees and sharpens it. The paper's internal validation against bvp4c cannot discriminate between the neural network and the ODE, because both solve the same equation. The load-bearing question is whether Eq. (27) is the correct acoustic model. I found a specific, checkable error in the derivation: Eq. (7) is valid only for an isentropic mean state, but the paper's mean flow has a strong temperature gradient and nearly constant pressure, so the relation underestimates the mean density gradient by about a factor of 50 at the outlet. This term enters the linearized continuity equation and is not a negligible Mach-order effect. If the proposed test confirms the discrepancy, the central claim is undermined; the paper would then be, at best, a demonstration that a PINN can solve a particular ODE, not a prediction of the acoustic field. The verdict should remain conditional, but the condition should include re-deriving or verifying Eq. (27) against the original linearized Euler equations. The reader's proposal to check against an exact analytical solution is related but not sufficient: even an exact solution of Eq. (27) would not expose the Eq. (7) error. The specific test proposed here isolates the failure.","tokens_in":15339,"tokens_out":24592,"duration_ms":261871,"concrete_test":"Recompute the mean flow for the linear temperature case at x = 1 and evaluate both sides of Eq. (7). Then solve the original linearized equations (4)-(5) with the exact mean density gradient from p = ρRT (without the isentropic mean-state assumption) using bvp4c, and compare that solution with the solution of Eq. (27) for the same boundary conditions (bp0 = 1 Pa, bpL = -1 Pa). If the two acoustic-pressure solutions differ by more than a few percent at 500 Hz, Eq. (27) is not the correct physical model, and the neural network results in Tables 2-5 do not validate the acoustic physics.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Equation (27) is the central object: the neural network learns its solution, and bvp4c is used as the 'true' reference. The derivation of Eq. (27) is therefore load-bearing. That derivation uses Eq. (7), dρ/dx = ρ/(γp) dp/dx, taken from Appendix A's assumption p/ρ^γ = constant. For the mean state solved in Section 2, this is false: the temperature varies by a factor of two while pressure changes by only a few percent, so ρ is dominated by T, not p. Using the paper's own equations (36)-(38) for the linear profile at x = 1, the exact density gradient from p = ρRT is dρ/dx ≈ 0.46 kg/m^4, while the right-hand side of Eq. (7) is ≈ 0.009 kg/m^4, roughly a factor of 50 smaller. This relation enters the linearized continuity equation and changes the coefficient of the bu dp/dx term in Eq. (8); the missing temperature-gradient part of dρ/dx is not a higher-order Mach-number correction. Consequently, ζ1-ζ3 in Eq. (27) may not represent the acoustic field in a duct with a strong temperature gradient. The validation in Tables 2-5 is circular: bvp4c solves Eq. (27), so the O(10^-5) relative errors establish only that the neural network solves Eq. (27) accurately; they do not test whether Eq. (27) is the correct physical model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives a second-order ordinary differential equation, Eq. (27), for the complex acoustic pressure in a one-dimensional duct carrying a mean flow with an axial temperature gradient, under assumptions of inviscid perfect-gas flow and isentropic acoustics. The equation is solved with a feedforward neural network: a trial solution enforces the boundary conditions, the complex ODE is split into real and imaginary parts, and a residual-based loss is minimized with L-BFGS. Results are reported for two temperature profiles (linear and sinusoidal) and four frequencies (500?2000 Hz), with the neural-network solution compared against the MATLAB bvp4c solver; relative errors are on the order of 10^-5 to 10^-7 for acoustic pressure and particle velocity. The paper also describes a transfer-learning variant for particle velocity and a uniform-flow limit check.","tokens_in":15694,"tokens_out":28635,"duration_ms":297571,"significance":"If the governing equation were correct, the paper would make a useful contribution by demonstrating a physics-informed neural-network approach for a complex-valued duct-acoustics problem with non-uniform mean flow and temperature, and the reported O(10^-5) errors would be a strong numerical result. The trial-solution construction for enforcing boundary conditions, the real/imaginary splitting of a complex ODE, and the use of automatic differentiation for particle velocity are all sensible methodological elements, and the uniform-flow reduction in Eq. (62) is a good consistency check. However, the central physical derivation is flawed: the mean-state relation used to derive the governing equation is inconsistent with the mean-flow model solved in the paper. Because all numerical results are solutions of this same equation, the significance of the paper is not currently established. The transfer-learning claim also lacks the promised quantitative demonstration.","major_comments":[{"comment":"The relation dρ/dx = (ρ/(γp)) dp/dx is derived in Appendix A from the assumption p/ρ^γ = constant for the mean state, but the mean state computed in Section 2 via Eqs. (33)-(38) does not satisfy that assumption. For the linear profile with T0=1600 K, TL=800 K, p0=1e5 Pa, and M0=0.2, p/ρ^γ changes by a factor of roughly 2.7 between x=0 and x=1. Consequently Eq. (7) is false: at x=1 the exact density gradient from p=ρRT is approximately 0.44 kg/m^4, whereas the right-hand side of Eq. (7) is approximately 0.009 kg/m^4, a discrepancy of about a factor of 50. More fundamentally, substituting Eq. (14) into Eq. (7) gives α = M^2 α, which forces α=0 for M≠1; the two equations are mutually inconsistent. Since Eq. (7) is used to obtain Eq. (8), the coefficients ζ1-ζ3 in Eq. (27) do not represent the linearized acoustics of the stated mean flow. In particular, the u' ∂ρ/∂x term in Eq. (4) is under-represented by roughly a factor of 1/M^2 ≈ 25 for M=0.2. This is a load-bearing error because every reported result is a solution of Eq. (27).","section":"Section 2, Eq. (7)"},{"comment":"The validation is circular with respect to the physical model. The 'true solution' is the bvp4c solution of Eq. (27), which is the very equation whose residual defines the neural-network loss. The O(10^-5) errors in Tables 2-5 therefore demonstrate only that the neural network solves Eq. (27) accurately; they do not provide evidence that Eq. (27) is the correct equation for the acoustic field. The uniform-flow reduction in Section 4.3 is a necessary consistency check but cannot detect errors that vanish when α=0 and dM/dx=0. Please validate the nonuniform results against an independent solution of the linearized Euler equations (for example, direct numerical integration of Eqs. (4)-(5) without the isentropic mean-state assumption) or against an exact analytical solution for a temperature gradient.","section":"Section 4.1, Tables 2-5"},{"comment":"All reported relative errors are from a single training run, with no error bars, no statistics over multiple initializations, no convergence curves, and no statement of the number of L-BFGS iterations or stopping criterion. The central quantitative claim of O(10^-5) maximum relative error is therefore not established as robust. Since the reported errors span 3e-7 to 8e-5, a few additional runs with different random seeds and a report of mean±standard deviation would be needed to support the accuracy claim.","section":"Section 4.1, Tables 2-5"},{"comment":"The transfer-learning variant for particle velocity is described in detail, and the text states that 'both methods yield similar results,' but no quantitative comparison, figure, or table is provided for this method. Since the abstract and the conclusion list transfer learning as a demonstrated contribution, this claim is currently unsupported. Please include the promised quantitative comparison or remove the transfer-learning claim from the abstract and conclusions.","section":"Section 4.2, Eqs. (59)-(61)"}],"minor_comments":[{"comment":"The time-harmonic convention is inconsistent: the paper states ρ', p', u' ∝ e^{-jωt}, but Eqs. (8) and (9) appear to have been derived with the opposite sign convention (e^{+jωt}). Please make the sign convention consistent throughout, or explicitly state the convention used in the derivation.","section":"Section 2, Eqs. (8)-(9)"},{"comment":"Equation (B.11) contains a typo: the first term on the right-hand side should be (1/c) du/dx rather than (1/c) dM/dx.","section":"Appendix B.4, Eq. (B.11)"},{"comment":"The statement that a single neural network architecture predicts both acoustic pressure and particle velocity is misleading: the particle velocity reported in Figs. 7 and 8 is evaluated algebraically from the predicted pressure using Eq. (57), not predicted by the same network. The transfer-learning variant uses a separate network with its own parameters.","section":"Conclusion, point (1)"},{"comment":"The text states that the uniform-flow case was validated with a maximum relative error of 0.01% but explicitly excludes the results 'to avoid redundancy.' Either include this comparison in a figure or table, or remove the numerical value, since an unreported validation result cannot be assessed.","section":"Section 4.3"},{"comment":"The description of Eq. (33) as a 'first-order approximation of the Maclaurin series' is imprecise. With the steady mass-conservation relation ρu = ρ0u0, the integrated momentum equation p + ρ0u0u = constant is exact, not an approximation.","section":"Section 2, Eq. (33)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the neural-network methodology is reasonable. The main obstacle is the incorrect derivation of the governing equation, which invalidates the physical interpretation of the results. This is fixable by re-deriving the linearized equations with the correct mean-density gradient and re-running the numerical experiments; the methodological framework itself does not need to be abandoned. I would not recommend rejection, but the revision must address the Eq. (7) inconsistency and provide independent validation of the corrected equation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me cut to the chase: this is a clean PINN demo for a complex-valued 1D duct ODE, but the physical grounding of the ODE itself looks shaky. The network achieves relative errors around 1e-5 against bvp4c on that ODE, and the real/imaginary split and trial solution are competently done. That part is fine. But the derivation of Eq. (27) uses Eq. (7), which assumes the mean flow is isentropic (p/ρ^γ constant). The paper's own mean flow—temperature dropping from 1600K to 800K while pressure changes by only a few percent—violates that. Using their Eqs. (36)-(38), the density gradient from p=ρRT is about 0.45 kg/m^4, while the RHS of Eq. (7) is about 0.009 kg/m^4. The missing temperature-gradient term is not a small Mach-order correction; it's the dominant term. So ζ1-ζ3 in Eq. (27) likely don't describe the acoustic field in this duct. The validation against bvp4c simply confirms the network solves Eq. (27); it says nothing about the physics.\n\nThere are other soft spots. The paper never compares with the existing analytical solutions for temperature-gradient ducts (e.g., Sujith et al. 1995, Karthik et al. 2000), which would have caught the discrepancy. There are no error bars or multiple-seed runs—tables are single-run relative errors. The transfer-learning method is described at length but no results are shown; the reported particle velocity comes from the algebraic relation, not from the separate network. And the claim of a 'maximum relative error of O(10^-5)' covers only four frequencies and two profiles.\n\nIf you ignore the physical derivation, the paper demonstrates that a standard PINN can solve a second-order ODE with complex, x-dependent coefficients to high accuracy. That is useful as a numerical exercise, but it's not new: the same techniques are in Lagaris, Raissi, etc. The derived equation follows Li & Morgans, minus the careful treatment of mean-flow entropy gradients, which is likely the source of the error.\n\nI would not cite this in its current form, and I wouldn't send it to a serious referee until the derivation is fixed or the claim is scaled back to 'neural solver for a given ODE.' If an editor did send it out, any competent reviewer should flag Eq. (7). This is more of a cautionary example than a publishable result.","headline":"The PINN machinery works, but the governing equation it solves appears to be derived from an isentropic mean flow that the paper's own temperature profile violates—so the 1e-5 matches only show the network solves that equation, not the physical problem.","tokens_in":16212,"tokens_out":5847,"would_cite":false,"duration_ms":64330,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["76Q05","68T07"],"pacs":["43.20.Ma"],"model":"deepseek-v4-flash","headline":"This paper derives a frequency-domain acoustic equation for ducts with varying mean flow and temperature, and shows that a physics-informed neural network solves it for both acoustic pressure and particle velocity up to 2000 Hz with…","keywords":["physics-informed neural networks","duct acoustics","complex-valued acoustic field","temperature gradient","mean flow","transfer learning","automatic differentiation","one-dimensional wave equation"],"falsifier":"Run the same neural-network formulation against an exact analytical solution for a duct with a steep temperature gradient at higher Mach number, where |M $\\alpha$| is no longer small compared to k, and compare both the network and bvp4c against the exact solution; if the approximation is invalid, the deviation from the exact solution will exceed the reported O($10^{-5}$) even though the network and bvp4c agree with each other.","tokens_in":15152,"feed_emoji":"🔊","tokens_out":6297,"duration_ms":63409,"temperature":0.7,"pith_summary":"This paper derives the frequency-domain equation for sound propagation in a one-dimensional duct whose mean temperature and velocity vary along its length, then solves it with a feedforward neural network instead of a classical solver. The network predicts both acoustic pressure and particle velocity up to 2000 Hz, matching the MATLAB boundary-value Runge-Kutta solver bvp4c with maximum relative errors on the order of $10^{-5}$ for both linear and sinusoidal temperature profiles. The work matters because duct acoustics in gas-turbine combustors involves exactly this kind of heterogeneous, mean-flow-carrying medium; a network that reproduces the complex-valued acoustic field without training data is a step toward higher-dimensional physics-constrained formulations. The paper also shows that a separate network trained by transfer learning recovers particle velocity from the momentum equation alone, and that imposing a temperature gradient changes both the spatial growth and the resonance of the acoustic amplitude.","feed_headline":"Physics-constrained network matches duct-acoustic solver to 1e-5","feed_subtitle":"Pressure and particle velocity in ducts with temperature gradients and mean flow match Runge-Kutta to 1e-5.","key_machinery":"The load-bearing object is the derived second-order governing equation for complex acoustic pressure, whose coefficients zeta1, zeta2, and zeta3 are given in Eqs. (28)-(30). These coefficients encode the mean-flow Mach number M, its gradient dM/dx, the logarithmic density gradient alpha, and the second logarithmic derivative beta, all computed from an inlet condition and a prescribed axial temperature profile under a first-order Maclaurin approximation of the steady momentum equation. Two pieces of machinery turn this ODE into a trainable loss: the trial-solution method, which enforces the boundary conditions exactly by construction, and a real/imaginary split of the complex equation that makes the loss function real-valued. The network itself is a seven-layer feedforward net with sine activations and 90 neurons per hidden layer; automatic differentiation supplies the first and second derivatives needed to evaluate the residual, and a transfer-learned second network recovers particle velocity from the momentum equation.","core_discovery":"The paper claims that the acoustic field in a one-dimensional uniform duct with axially varying mean flow and temperature obeys a single second-order linear ordinary differential equation with complex coefficients, Eq. (27), and that a feedforward neural network can solve this equation to high accuracy without any simulation or experimental data. The authors split the complex ODE into two real-valued equations, embed the boundary conditions directly into a trial solution, and minimize the residual loss at 10,000 collocation points using the L-BFGS optimizer. Against the bvp4c solution of the same equation, the network predicts acoustic pressure with relative errors between $10^{-7}$ and $10^{-5}$ across 500-2000 Hz for both temperature profiles; particle velocity, whose magnitude is several orders smaller, is recovered from the pressure field with similar accuracy. They further claim that the temperature gradient increases acoustic amplitude along the duct and shifts resonance frequencies, and that the formulation demonstrates the utility of automatic differentiation and transfer learning for acoustic applications.","pith_inferences":["Because the reported errors are measured against the same derived equation solved by bvp4c, the paper demonstrates solver agreement rather than physical-model validity; an exact analytical solution or an experiment would be needed to test the equation itself.","The small-gradient assumption |M alpha| << k bounds the valid regime; a testable extension is to increase Mach number or temperature-gradient steepness until the errors degrade, mapping the method's operational envelope.","The trial-solution structure ties each network to one domain and one set of boundary conditions, a limitation the authors state; conditioning the network on frequency as an additional input would let a single model sweep the acoustic spectrum without retraining.","The success on two temperature profiles suggests the network could serve as a fast surrogate inside combustor thermoacoustic instability solvers, where repeated acoustic-field evaluations at varying temperature states are needed."],"forward_implications":["One trained network predicts both acoustic pressure and particle velocity even though their magnitudes differ by several orders, so quantities like acoustic intensity, impedance, and transfer functions can be computed directly from the network.","Because the network is a differentiable function of position, automatic differentiation can extract other acoustic state variables without re-solving the governing equation.","A transfer-learned second network recovers particle velocity from the momentum equation alone, on a coarser collocation grid, and can be stored and recalled without retaining the pressure field.","Imposing a temperature gradient changes the acoustic field qualitatively: peak amplitude grows along the duct and resonance frequencies shift, consistent with exact solutions in the literature.","The trial-solution formulation extends naturally to higher dimensions or more complicated physics because it only requires a residual evaluation and boundary-embedded trial functions."],"supporting_citations":[{"why":"Supplies the linearized continuity and momentum equations and the analytical treatment of ducts with arbitrary mean axial temperature gradient and mean flow that the derivation follows.","marker":"25"},{"why":"Provides the baseline duct-acoustics textbook relations, including isentropic speed of sound and plane-wave duct equations used in the derivation.","marker":"2"},{"why":"Introduces the trial-solution method that converts the boundary-value problem into an unconstrained optimization problem.","marker":"32"},{"why":"Establishes the physics-informed neural network framework and the splitting of complex equations into real and imaginary loss terms.","marker":"14"},{"why":"Provides an exact solution for one-dimensional acoustic fields in ducts with an axial temperature gradient, used as a reference for the temperature-gradient effect.","marker":"38"},{"why":"Gives exact solutions to one-dimensional acoustic fields with temperature gradient and mean flow, used as a literature check for observed amplitude growth and resonance shifts.","marker":"26"}],"fun_headline_variants":["Physics-constrained NN hits 1e-7 in duct acoustics","Neural net solves duct acoustics with physics constraints","Duct acoustics solved to 1e-7 via neural networks","Temperature gradient effects on duct sound via NN","Physics-constrained neural nets match duct-acoustic solver"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derived governing equation, Eq. (27), is taken to be the true physics of the duct: specifically, the small-gradient expansion |M $\\alpha$| << k in Eq. (17), the neglect of Mach-number terms beyond $M^{2}$ in Eq. (19), and the first-order Maclaurin approximation of the steady momentum equation in Eq. (33) must all hold for the regime being modeled; if they fail, the neural network could match bvp4c perfectly while both solve the wrong equation.","fun_headline_variants_meta":{"raw":{"variants":["Physics-constrained NN hits 1e-7 in duct acoustics","Neural net solves duct acoustics with physics constraints","Duct acoustics solved to 1e-7 via neural networks","Temperature gradient effects on duct sound via NN","Physics-constrained neural nets match duct-acoustic solver"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000527,"raw_usage":{"total_tokens":2490,"prompt_tokens":836,"completion_tokens":1654,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":1572}},"tokens_in":452,"tokens_out":1654,"duration_ms":15524,"temperature":1.0,"reasoning_tokens":1572,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:45:55.674823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same neural-network formulation against an exact analytical solution for a duct with a steep temperature gradient at higher Mach number, where |M $\\alpha$| is no longer small compared to k, and compare both the network and bvp4c against the exact solution; if the approximation is invalid, the deviation from the exact solution will exceed the reported O($10^{-5}$) even though the network and bvp4c agree with each other.","supporting_citations":[{"cited_title":"Li and A","cited_arxiv_id":null,"evidence_quote":"Supplies the linearized continuity and momentum equations and the analytical treatment of ducts with arbitrary mean axial temperature gradient and mean flow that the derivation follows."},{"cited_title":"Munjal, Acoustics of Ducts and Mufflers (John Wiley and Sons Ltd., West Sussex, UK, 2014), 2nd edition","cited_arxiv_id":null,"evidence_quote":"Provides the baseline duct-acoustics textbook relations, including isentropic speed of sound and plane-wave duct equations used in the derivation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the trial-solution method that converts the boundary-value problem into an unconstrained optimization problem."},{"cited_title":"Raissi, P","cited_arxiv_id":null,"evidence_quote":"Establishes the physics-informed neural network framework and the splitting of complex equations into real and imaginary loss terms."},{"cited_title":"Sujith, G","cited_arxiv_id":null,"evidence_quote":"Provides an exact solution for one-dimensional acoustic fields in ducts with an axial temperature gradient, used as a reference for the temperature-gradient effect."},{"cited_title":"Karthik, B","cited_arxiv_id":null,"evidence_quote":"Gives exact solutions to one-dimensional acoustic fields with temperature gradient and mean flow, used as a literature check for observed amplitude growth and resonance shifts."}],"review_version":1}