{"id":"99ddd6ce-789c-4dad-87d6-a0dd8fa14d9d","arxiv_id":"2505.20300","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"FMEnets couples three physics-informed networks, in MLP or KAN form, to predict velocity, concentration, and temperature in non-ideal plug flow reactors and to infer activation energy from sparse data.","lead":"FMEnets is a machine learning framework that links three neural networks to jointly solve flow, species transport, and heat balance equations for plug flow reactors. It can infer reaction activation energy from sparse outlet data with around 2 percent error, but only on simulated test cases so far.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The FEM weak form in Section 3.5 does not match the strong form in Eq. (4); if the FEniCS code follows the written equations, the ground truth is corrupted and all reported errors are against the wrong reference.","rationale":"The single most load-bearing condition for the paper's central claim is that the simulated FEM solutions used as ground truth actually solve the stated reactor model. The reader identified this as the weakest assumption, focused on physical fidelity to a real reactor. I sharpen this to an internal consistency problem: the weak form written in Section 3.5 does not correspond to the strong form in Eq. (4). Because the paper reports all forward and inverse errors against this FEM reference, an error in the FEM formulation would invalidate every accuracy number, including the headline '<2.5%' activation-energy error. This is not a disagreement with consensus or a request for experimental data; it is a checkable mathematical inconsistency in the numerical method section. The concern is load-bearing because the ablation study and inverse results all inherit the same reference. I do not move the verdict because the reader already rated the paper CONDITIONAL; my concern reinforces that conditionality rather than changing it. If the concrete test shows the FEM reference is insensitive to the weak-form issue, the remaining concerns—no code, missing error bars, overclaimed priority—still support a conditional recommendation. If the test shows sensitivity, the paper would need major revision or rejection. The multi-residence-time surrogate approximation in Section 3.4 is a secondary worry, but the weak-form inconsistency is more fundamental and more immediately checkable.","tokens_in":18662,"tokens_out":14425,"duration_ms":148985,"concrete_test":"Re-derive the weak form of Eq. (4) on the axisymmetric domain with measure r dr dz and re-run the FEniCS simulation for Case 1 with the corrected bilinear form; then recompute the FME-PINNs L2 errors in Table 1 against the new FEM solution. If the corrected concentration or temperature profiles differ from the published FEM profiles by more than about 1%, or if any Table 1 entry shifts materially, the ground truth is sensitive to the weak-form error and the paper's accuracy claims must be re-evaluated. As a secondary check, inspect the released FEniCS code (if provided) to determine whether the implemented variational form matches Eq. (16) or the corrected form.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim rests on comparing FMEnets predictions with FEM solutions generated by the FEniCS code described in Section 3.5. As written, however, the FEM weak form is inconsistent with the strong form. Equation (4) has diffusion term -(1/Pe)[(1/r)∂C/∂r + ∂²C/∂r² + ∂²C/∂z²], so the (1/r)∂C/∂r term is part of the Laplacian with coefficient -1/Pe. Equation (16) writes the weak form as ∫(u·∇C)φ dΩ - (1/Pe)∫ΔCφ dΩ + ∫(1/r)∂C/∂r φ dΩ + ∫ rφ dΩ = 0, with ΔC apparently the Cartesian Laplacian ∂²C/∂r² + ∂²C/∂z². The (1/r)∂C/∂r term thus appears with the wrong coefficient and sign. After 'integration by parts', the term ∫(1/r)∂C_h/∂r φ is kept unchanged (Equations 17-18), which is not the weak form of the axisymmetric Laplacian under either the Cartesian measure dΩ=dr dz or the cylindrical measure r dr dz. If the FEniCS implementation follows this weak form, the FEM reference is not a solution of the PDE that FMEnets is trained on; the reported L2 errors in Tables 1-3, and the inferred Ea errors, are then comparisons against a corrupted solution. If the code is correct, the manuscript must provide the correct derivation, since the written method is not reproducible. Either way, the validation of the central <2.5% claim is not currently supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes FMEnets, a modular physics-informed machine-learning framework for steady-state non-adiabatic plug-flow reactors. Three sequentially trained subnetworks enforce the Navier-Stokes, species material-balance, and energy-balance equations in an axisymmetric (r,z) domain; each subnetwork can be an MLP (FME-PINNs) or a Chebyshev KAN (FME-KANs). In forward mode the model is trained with PDE residuals, boundary conditions, and sparse inlet/outlet concentration data; in inverse mode it additionally uses sparse simulated multi-residence-time outlet measurements to infer the activation energy Ea. Results are reported for three reaction systems against FEniCS finite-element solutions, together with noise-robustness experiments and ablations. The central quantitative claim is that relative L2 errors for state variables are mostly a few percent and that inferred Ea errors remain below 2.5%.","tokens_in":19013,"tokens_out":7584,"duration_ms":72690,"significance":"If the results hold, the sequential decomposition into flow, material, and energy subnetworks is a practically useful surrogate for reactor CFD and a plausible template for kinetic-parameter inference from sparse measurements. The ablation study (Table 4) gives useful evidence that the multi-network decomposition, rather than the PINN or KAN representation alone, is responsible for accuracy, and the noise-robustness comparison in Figure 8 is a valuable datapoint for practitioners. A notable strength is that the paper explicitly acknowledges the proof-of-concept nature of using simulated data instead of experiment. However, every reported accuracy number depends on the FEniCS reference solutions reproducing the written governing equations, and the manuscript provides no code or data; the validation is therefore only as strong as the written numerical formulation.","major_comments":[{"comment":"The weak form written in Eqs. (16)-(18) is not the weak form of the strong form in Eq. (4). In Eq. (4) the radial term (1/r)∂C/∂r appears only as part of the axisymmetric Laplacian multiplied by -1/Pe, whereas Eq. (16) adds a separate term ∫Ω (1/r)(∂C/∂r)φ dΩ with coefficient +1 alongside -(1/Pe)∫Ω ΔC φ dΩ. If ΔC denotes the Cartesian Laplacian ∂²C/∂r²+∂²C/∂z², the diffusive operator in the weak form becomes -(1/Pe)(∂²C/∂r²+∂²C/∂z²)+(1/r)∂C/∂r, which is not the operator in Eq. (4); if ΔC denotes the cylindrical Laplacian, the term (1/r)∂C/∂r is counted twice with inconsistent coefficients. The integration by parts in Eq. (17) leaves this term untouched, so the form is not a valid variational formulation under either the Cartesian measure dΩ=dr dz or the cylindrical measure r dr dz. Because every L2 error in Tables 1-3 and every Ea error is evaluated against FEniCS solutions obtained from Eq. (18), the central accuracy claim is unsupported unless the FEniCS implementation differs from the written form; the authors must either provide the corrected weak form or release the actual code so the reference solver can be verified.","section":"Section 3.5, Eqs. (16)-(18)"},{"comment":"The Chebyshev recurrence is stated as T_n(ζ)=2ζT_{n-1}(ζ)+T_{n-2}(ζ), but the standard three-term recurrence is T_n(ζ)=2ζT_{n-1}(ζ)-T_{n-2}(ζ). With the plus sign, T_2(ζ)=2ζ²+1, which is not a Chebyshev polynomial. Since FME-KANs are a central contribution and the paper claims that cKANs are numerically stable, this written definition must be corrected or the implementation clarified for the KAN results to be reproducible.","section":"Section 3.1.2, Chebyshev KAN recurrence"}],"minor_comments":[{"comment":"The phrase 'using only inlet and outlet information' overstates the forward setup: the forward model also enforces PDE residuals and boundary conditions throughout the domain, with inlet/outlet data as an additional constraint; please rephrase to avoid implying a data-only input-output map.","section":"Abstract and Section 1"},{"comment":"The residual eα(x_i,θ)=|αhat(x_i)-α(x_i,θ)| is a sensible definition for boundary and data losses, but for PDE residuals the target αhat is zero; this convention should be stated explicitly, since Eq. (8) currently reads as if every loss term has a nonzero target field.","section":"Section 3.1.1, Eq. (8)"},{"comment":"The claim that this is the 'first attempt to recover kinetic parameters in reactive flow systems using the physics-informed machine learning model' is too strong given the existing inverse-PINN literature on reacting flows cited in Section 1.1; please moderate the novelty claim.","section":"Section 3.4"},{"comment":"The caption 'single-component ablation study of the three-component reaction system' is confusing; the table reports ablations on the three-component system, not a single-component system, and should be reworded.","section":"Table 4 caption"},{"comment":"The symbol r is used both for the radial coordinate and for the reaction-rate term in Eq. (16), making the last integral ambiguous; use a distinct symbol such as R_i for the reaction contribution.","section":"Section 3.5, Eq. (16)"}],"recommendation":"major_revision","confidential_remarks":"The main technical risk is the discrepancy between the strong form in Eq. (4) and the written FEM weak form in Eqs. (16)-(18). This is a load-bearing issue because all reported errors, including the 2.5% Ea claim, are measured against FEM solutions generated with that formulation. The issue is potentially fixable if the code itself is correct, but without code release the current manuscript does not support its central validation. I would encourage the editor to request the FEniCS implementation or a corrected variational derivation as part of the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe interesting bit here is the sequential three-network design: NN1 solves Navier–Stokes, NN2 the species balances, NN3 the energy balance, with the earlier networks feeding the later ones. The ablation actually shows the architecture matters—removing it sends errors to 177% on Ca. The inverse setup, inferring activation energy from multi-residence-time outlet data, is a genuinely useful scenario for pharma process development, and the KAN noise-robustness comparison is a nice touch. The paper is also honest that this is a simulated proof-of-concept.\n\nBut there is a load-bearing flaw that a reader's report missed. Section 3.5 writes the FEM weak form of the material balance in a way that does not match the strong form in Eq. (4). The strong form has the cylindrical Laplacian term -(1/Pe)[(1/r)C_r + C_rr + C_zz]. The weak form in Eqs. (16)–(18) keeps the (1/r)C_r term as a separate term with the wrong sign and coefficient (+1 instead of -1/Pe) and no compensating measure term. That is not the correct weak form of the axisymmetric Laplacian under either dr dz or r dr dz measure. If the FEniCS code literally follows the written equations, then the FEM 'reference' is solving a different PDE, and every reported L2 error—including the <2.5% Ea claim—is a comparison against a corrupted solution. If the code is actually correct, then the paper's description of the discretization is wrong and the method is not reproducible. Either way, the validation as written does not support the central claim.\n\nThe other issues are minor by comparison: no code or data, no error bars on the main tables, an over-broad 'first' claim, and the computational-efficiency statement is just per-iteration wall time, not total cost or a comparison against FEM.\n\nMy read: the architecture is worth a serious look, and the inverse problem framing is valuable. But the weak-form inconsistency must be resolved before any of the error numbers can be trusted. I would send it to peer review, with the expectation that the authors either provide the correct weak form and verify the FEniCS code against it, or release the code so reviewers can check. If the FEM reference is indeed wrong, the empirical validation collapses, but the forward/inverse framework may still be salvageable.","headline":"A useful sequential multi-network architecture for reactor design, undermined by a written FEM weak form that is inconsistent with the strong form, so the reported error numbers lack a trustworthy reference.","tokens_in":19568,"tokens_out":6571,"would_cite":false,"duration_ms":61104,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","80A32"],"pacs":[],"model":"deepseek-v4-flash","headline":"Three coupled physics networks predict reactor fields from inlet-outlet data and infer activation energy to within 2.5 percent.","keywords":["physics-informed neural networks","Kolmogorov-Arnold Networks","sparse data","stiff equations","conservation laws","inverse multi-physics problems","plug flow reactor design","activation energy inference"],"falsifier":"Measure outlet concentrations and temperatures from a physical isothermal-bath plug flow reactor whose activation energy is independently known, run the inverse FMEnets on those data, and compare the inferred activation energy to the known value; if the deviation exceeds the reported roughly 2.5 percent, or if retesting against a substantially finer finite element mesh changes the reported errors, the claimed accuracy is an artifact of training and testing on the same simulated solver.","tokens_in":18492,"feed_emoji":"⚗️","tokens_out":8756,"duration_ms":76209,"temperature":0.7,"pith_summary":"FMEnets is a physics-informed machine learning framework that splits a non-ideal plug flow reactor's governing physics across three sequentially trained sub-networks: one solves the Navier–Stokes equations for the velocity field, one enforces the material balance for species concentrations, and one enforces the energy balance for temperature. The paper's claim is that this decoupled sequencing is what lets the model predict velocity, pressure, concentration, and temperature from inlet and outlet data alone, and, in inverse mode, recover an unknown activation energy from sparse multi-residence-time measurements with relative errors under 2.5 percent. A sympathetic reader cares because conventional reactor design relies on ideal-reactor assumptions and costly finite element solves, whereas FMEnets promises a fast surrogate that also extracts kinetic parameters from the sparse, noisy data typical of pharmaceutical process development.","feed_headline":"Three linked neural nets solve reactor design within 2.5% error","feed_subtitle":"Trained on inlet-outlet data, the network trio also recovers unknown activation energy from sparse measurements.","key_machinery":"The load-bearing object is the FMEnets architecture itself: three neural networks chained in sequence, in which NN1 solves the axisymmetric steady Navier–Stokes equations for the velocity field, NN2 uses that velocity to enforce the material balance for each species concentration, and NN3 uses both to enforce the energy balance for temperature. Training runs in two stages, with NN1 trained alone for 30,000 iterations and then frozen while NN2 and NN3 train, so coupling variables propagate downstream only after the flow is learned. Each sub-network may be a multilayer perceptron (FME-PINNs) or a Chebyshev Kolmogorov–Arnold network (FME-KANs), and the equations are non-dimensionalized through Reynolds, mass-transfer Péclet, and thermal Péclet numbers. Supporting mechanisms include exact Dirichlet boundary conditions via approximate distance functions, weight normalization, and residual-based attention weights that balance per-point loss contributions; the ablation study singles out the FME structure and the attention weights as the components whose removal most damages accuracy.","core_discovery":"The central claim, stated on the paper's own terms, is that coupling the Navier–Stokes, material balance, and energy balance equations through three interconnected sub-networks with independent optimizers is the decisive design choice: a single network trained on all residuals at once fails, while the sequential FME structure succeeds. In forward mode, trained on only inlet and outlet information, FME-PINNs reproduce velocity and pressure to within about 1 percent relative error, species concentrations to within 0.99–7.95 percent, and temperature to within 1.7 percent across a two-component, a three-component sequential, and a six-component parallel reaction system. In inverse mode, outlet data from quarter-, half-, and full-length reactors let the same architecture infer unknown activation energies with relative errors under 2.5 percent while simultaneously predicting concentration and temperature fields. The paper further claims that the FME-KAN variant matches FME-PINN accuracy while tolerating injected measurement noise better, holding state-variable errors below about 6 percent even at 10 percent noise. The ablation study is offered as proof of the architecture's necessity: removing the FME structure entirely raises concentration errors to hundreds of percent, and removing residual-based attention weighting also degrades accuracy substantially.","pith_inferences":["Editorial inference: every accuracy number is measured against finite element solutions of the same PDE model, so the reported errors certify how faithfully the network fits that model, not how faithfully the model describes a physical reactor; agreement with real experimental data is the untested next step.","Editorial inference: the frozen-network sequencing implies each block could be swapped independently, say replacing the flow sub-network with an empirical pressure-drop correlation or a different surrogate, without retraining the full stack, a modularity the paper gestures at but does not demonstrate.","Editorial inference: using quarter-, half-, and full-length reactor outlets as surrogate interior data is a general recipe that should transfer to other tubular equipment, such as fixed-bed or monolithic reactors, whenever a residence-time series can be sampled.","Editorial inference: a direct test of the 2.5 percent inverse-error claim would push beyond the tested configurations, with fewer measurement locations, higher noise, or a six-component inverse problem, to see where the error begins to climb."],"forward_implications":["Forward reactor design problems could be solved from inlet and outlet measurements alone, replacing repeated finite element solves with a trained surrogate for rapid iteration and optimization.","Unknown kinetic parameters, particularly activation energy, could be estimated from sparse multi-residence-time outlet measurements, addressing reactions whose kinetics are not yet established, a common situation in pharmaceutical process development.","Process data carrying up to 10 percent measurement noise would favor the KAN implementation, which holds state-variable errors below about 6 percent while the MLP variant degrades more sharply.","Because each physical sub-problem is trained and frozen separately, adding new species or reactions means extending the material-balance sub-network while leaving the flow solver untouched."],"supporting_citations":[{"why":"Supplies the physics-informed neural network loss formulation that every sub-network minimizes, the base method of the framework.","marker":"[15]"},{"why":"The prior two-dimensional (r,z) tubular reactor PINN work this paper extends; earlier work assumes an empirical temperature profile instead of solving the energy balance.","marker":"[33]"},{"why":"Provides the Chebyshev Kolmogorov–Arnold network representation and the MLP-versus-KAN comparison underlying FME-KANs.","marker":"[51]"},{"why":"Introduces Kolmogorov–Arnold Networks, the representation family from which the KAN variant is built.","marker":"[53]"},{"why":"Defines the residual-based attention weights for local loss balancing; the ablation study shows removing this component sharply inflates errors.","marker":"[56]"},{"why":"Supplies the approximate distance function method for exact imposition of Dirichlet boundary conditions on the velocity fields.","marker":"[60]"},{"why":"The finite element codebase that generates the ground-truth solutions against which all forward and inverse errors are measured.","marker":"[79]"},{"why":"Provides the sequential training strategy for inverse problems that inspires the two-step FMEnets training schedule.","marker":"[77]"}],"fun_headline_variants":["Three coupled nets cut reactor design error to 2.5%","Neural net trio infers reactor kinetics from sparse data","Physics-aware neural nets solve reactor design and inverse problems","Coupled sub-networks infer kinetics under 2.5% error","Physics-coupled neural nets hit <2.5% error on unknown kinetics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the finite element simulations used as ground truth represent the real reactor: all reported errors are measured against FEniCS solutions of the same PDE system, and Section 3.4 states that simulated data stand in for experimental data in this proof-of-concept, so if the governing equations or discretization diverge from physical reality, every accuracy number in the paper would fail to reflect real reactor behavior.","fun_headline_variants_meta":{"raw":{"variants":["Three coupled nets cut reactor design error to 2.5%","Neural net trio infers reactor kinetics from sparse data","Physics-aware neural nets solve reactor design and inverse problems","Coupled sub-networks infer kinetics under 2.5% error","Physics-coupled neural nets hit <2.5% error on unknown kinetics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001402,"raw_usage":{"total_tokens":5741,"prompt_tokens":1089,"completion_tokens":4652,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":705,"completion_tokens_details":{"reasoning_tokens":4564}},"tokens_in":705,"tokens_out":4652,"duration_ms":36147,"temperature":1.0,"reasoning_tokens":4564,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:41:09.117452+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure outlet concentrations and temperatures from a physical isothermal-bath plug flow reactor whose activation energy is independently known, run the inverse FMEnets on those data, and compare the inferred activation energy to the known value; if the deviation exceeds the reported roughly 2.5 percent, or if retesting against a substantially finer finite element mesh changes the reported errors, the claimed accuracy is an artifact of training and testing on the same simulated solver.","supporting_citations":[{"cited_title":"Optimal temperature trajectory for tubular reactor using physics informed neural networks","cited_arxiv_id":null,"evidence_quote":"The prior two-dimensional (r,z) tubular reactor PINN work this paper extends; earlier work assumes an empirical temperature profile instead of solving the energy balance."},{"cited_title":"Residual-based attention in physics-informed neural networks.Computer Methods in Applied Mechanics and Engineering , 421:116805, 2024","cited_arxiv_id":null,"evidence_quote":"Defines the residual-based attention weights for local loss balancing; the ablation study shows removing this component sharply inflates errors."},{"cited_title":"Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the approximate distance function method for exact imposition of Dirichlet boundary conditions on the velocity fields."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The finite element codebase that generates the ground-truth solutions against which all forward and inverse errors are measured."}],"review_version":1}