{"id":"d509d14a-bf58-4f4b-935e-746c3e69cdea","arxiv_id":"2608.00856","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A physics-informed neural network reproduces finite-difference solutions for wall-reactive solute dispersion in three canonical shear flows and yields differentiable wall-uptake diagnostics.","lead":"The paper applies physics-informed neural networks to simulate two-dimensional solute dispersion with absorbing walls in Couette, Poiseuille, and Couette-Poiseuille shear flows, matching a finite-difference benchmark at Péclet number 10. The trained networks are then used to extract wall-uptake and dispersion diagnostics that show how shear and asymmetric wall reactivity organize reactive removal.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Shared domain truncation and unverified ADI benchmark: the PINN–ADI agreement may reflect a common artificial-boundary error rather than true accuracy.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing gap: the ADI benchmark is used as ground truth without a grid-convergence study, and both PINN and ADI share the same x-truncation with zero-gradient conditions. If the plume reaches the artificial boundaries at Pe=10 over t in [0,1], then agreement between PINN and ADI merely shows that two numerical methods solve the same truncated problem consistently; it does not validate the physical claim that the PINN reconstructs the dispersion field for the underlying infinite-domain shear-flow problem. This is the single most load-bearing concern because every downstream diagnostic (D_a, M0, nu2, U_tot, wall-flux centroids) is derived from the validated concentration field; if the field carries a common truncation error, the physical conclusions about shear-controlled uptake organization are also compromised. The paper otherwise has strengths: it uses smooth Gaussian sources to avoid Dirac-delta issues, validates multiple integrated diagnostics, and reports training-loss stability. But those strengths do not compensate for the missing independent baseline. I therefore recommend no change to the reader's CONDITIONAL verdict, with the condition being that the authors must demonstrate domain-independence (or conservatively quantify truncation error) and report an ADI grid-convergence study before the central claim can be accepted.","tokens_in":34016,"tokens_out":8498,"duration_ms":82566,"concrete_test":"Run the same ADI scheme for the Couette-flow point-source case at Pe=10 with (beta1,beta2)=(0,0) and (1,1) on larger domains, e.g., x in [-10,10] and [-20,20], keeping grid spacing comparable (Nx=201 and Nx=401). Compare the y-averaged concentration, axial variance nu2(t), apparent dispersion coefficient D_a(t), and cumulative wall uptake U_tot(T) at t=1 against the reported [-5,5] results. If the changes in these diagnostics exceed the PINN-ADI discrepancies in Tables 1-3, the shared truncation error is the dominant source of the apparent agreement, and the central claim is not established. As a simpler check, inspect the [-5,5] ADI solution at t=1: if C/C_peak at x=±5 is above about 1%, the Section 2.6 'sufficiently large' assertion is violated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the ADI finite-difference solution as ground truth for the physical dispersion problem. But Section 2.6 truncates the infinite streamwise domain to x in [-5,5] and imposes zero-gradient conditions, asserting that the cloud 'remains away from the artificial axial boundaries' over t in [0,1]. That assertion is not quantitatively supported, and it is doubtful at Pe=10. For the Couette point-source case, a Taylor–Aris estimate gives D_eff on the order of 10 in dimensionless units, so the streamwise standard deviation at t=1 is roughly sqrt(2 D_eff t) ~ 3-4. The concentration at x=±5 is then a nontrivial fraction of the peak, so the zero-gradient boundary will reflect mass and suppress variance growth. Since the PINN loss (Section 3.4) and the ADI benchmark (Section 3.8) impose exactly the same truncated-domain condition, their agreement in Tables 1-3 and Figures 2-7 is consistency between two solvers solving the same artificially bounded problem, not validation against the physical dispersion problem claimed in the abstract. The missing grid-convergence study for ADI (Nx=101, Ny=51, dt=1e-4, no refinement) compounds this: even if the domain were large enough, the benchmark itself is unverified. The manuscript even concedes in Section 5 that the dispersion coefficient is 'prone to local approximation errors,' further weakening the headline diagnostic. Unless the truncation error is shown to be negligible, the reported agreement cannot establish the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a physics-informed neural network (PINN) for two-dimensional advection–diffusion transport of a reactive solute in Couette, Poiseuille, and Couette–Poiseuille shear flows between absorbing walls. The PINN is trained to satisfy the convection–diffusion equation, Robin wall conditions, and localized Gaussian initial sources on a finite streamwise domain, and the results are compared with an alternating-direction implicit (ADI) finite-difference benchmark at Pe=10 for non-reactive, symmetric, and asymmetric wall reactivities. The authors further exploit the differentiability of the network to extract diagnostics such as the apparent axial dispersion coefficient, total surviving mass, cumulative wall uptake, wall-dominance index, and flux-weighted uptake centroids. The central claim is that the PINN reconstructs the spatiotemporal concentration field and that the extracted wall-resolved diagnostics are validated by close agreement with the ADI solution.","tokens_in":34424,"tokens_out":7570,"duration_ms":67058,"significance":"If the central claim is established, the paper offers a useful differentiable surrogate for boundary-coupled reactive transport in canonical shear flows, and the attempt to validate not only concentrations but also derived integral diagnostics is commendable. The ADI comparison is an independent numerical benchmark rather than a circular check, and the range of flow and reaction regimes considered is appropriate for a methods paper. However, the validation rests on two unverified assumptions: that the finite streamwise domain is large enough not to perturb the plume, and that the ADI grid is converged. In addition, the quantitative error reporting is internally inconsistent. These are load-bearing gaps because both the PINN and the ADI solver impose the same artificial truncation, so agreement between them does not by itself establish accuracy against the infinite-domain physical problem.","major_comments":[{"comment":"The truncation of the infinite streamwise domain to x∈[-5,5] with zero-gradient conditions (Eq. 2.14) is not shown to be harmless. For the Couette case at Pe=10, the mean-centred shear u^{(C)}=y/2 gives a Taylor–Aris estimate D_eff≈1+Pe^2/15≈7.7, so the streamwise standard deviation at t=1 is σ≈sqrt(2D_eff)≈3.9, implying that the concentration at x=±5 is not negligible. The assertion in §2.6 that the cloud 'remains away from the artificial axial boundaries' is therefore not quantitatively credible, and the sentence in §3.7 stating that the concentration 'may stay non-zero at the artificial longitudinal boundaries throughout the simulation' further undercuts it. Because the PINN loss (§3.4) and the ADI benchmark (§3.8) impose exactly the same zero-gradient condition, the agreement reported in Tables 1–3 and Figures 2–7 may only demonstrate consistency between two solvers on the same artificially bounded problem. Please add a domain-size study (e.g., L_x=20 or 40) or compare against an analytical infinite-domain solution to show that truncation does not affect the reported fields and diagnostics.","section":"§2.6, §3.7"},{"comment":"The ADI finite-difference solution is used as the reference, but no grid-convergence study is reported: the manuscript fixes Nx=101, Ny=51, and Δt=10^-4 and does not refine the grid. The reader cannot tell whether the small PINN–ADI differences in Tables 2 and 3 reflect approximation errors of the PINN or discretization errors of the benchmark. Please add a convergence test with at least two refined grids and, for at least one reactive case, a comparison with a published analytical solution (e.g., the reactive dispersion solutions cited in the introduction, such as Jiang et al. 2022). Without this, the ADI benchmark is not independently verified as ground truth.","section":"§3.8"},{"comment":"The quantitative validation is internally inconsistent. For identical non-reactive line-like Couette cases at t=1, Table 1 reports an L2 error of 0.0415 while Table 2 reports 0.0161; for line-like Couette–Poiseuille at t=1 the values are 0.0721 and 0.0199, respectively. Tables 2 and 3 do not define whether the listed numbers are relative or absolute, L1 or L2, and Table 1 gives no definition of the norm either. This makes it impossible to assess the accuracy claims quantitatively. Please define the error metrics precisely, ensure consistency across all tables, and correct the duplicate entry for the Couette–Poiseuille point-like case in Table 1.","section":"§4.1, Tables 1–3"},{"comment":"The time-dependent dispersion coefficient D_a(t) (Eq. 4.5) is presented as a central diagnostic, yet the manuscript's own conclusion states that it is 'prone to local approximation errors.' The validation figures 11–15 show the discrepancy E_D(t) but provide no quantitative summary, so the reader cannot judge whether the errors are a few percent or substantial. Please report the maximum and mean of E_D(t) over t for the cases shown and discuss how this uncertainty propagates into the physical conclusions in Sections 4.7–4.10.","section":"§4.3, §5"}],"minor_comments":[{"comment":"The x-axes of the local wall-flux profiles span -10 to 10, whereas the computational domain stated in §3.7 is x∈[-5,5]; please reconcile the axes with the domain actually used.","section":"Figure 23"},{"comment":"In the Couette–Poiseuille point-like row, the error at t=0.5 is identical to the error at t=0.1 (0.0170186), which appears to be a typographical duplication and should be corrected.","section":"Table 1"},{"comment":"The code availability statement says that code is available 'upon reasonable request'; for a computational methods paper, depositing the code in an open repository would substantially improve reproducibility.","section":"§3.7, Data availability"},{"comment":"The sentence in §2.6 asserting that the cloud 'remains away from the artificial axial boundaries' should be replaced by a quantitative statement once the truncation check requested in the major comments is performed.","section":"§2.6"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the central idea is potentially useful, but the validation claims need to be tightened: the domain-truncation effect and the convergence of the ADI benchmark are load-bearing and currently unquantified. I would not reject on novelty grounds, but I would require the requested convergence and truncation studies before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid but incremental PINN application, and the main validation claim is undercut by a domain-truncation problem. What is new: it treats the initial release as a smooth Gaussian (mollified delta) and uses the differentiable network to extract wall-resolved diagnostics - uptake centroids, spreads, dominance index - that go beyond just matching concentration fields. The formulation is clean, and the reported PINN-ADI errors are mostly a few percent across three shear profiles and symmetric/asymmetric reactivities.\n\nThe load-bearing soft spot is the finite window. Both solvers run on x in [-5,5] with zero-gradient axial conditions. At Pe=10 and t=1, a Taylor-Aris estimate for Couette gives an r.m.s. plume width around 3, so the concentration at x=±5 is a nontrivial fraction of the peak. The zero-gradient boundary reflects mass and suppresses variance growth. The paper asserts the cloud stays away from the boundaries but gives no quantitative check. Therefore the PINN-ADI agreement shows the two methods solve the same artificially bounded problem; it does not establish accuracy for the infinite-domain dispersion problem stated in the abstract. This is a genuine flaw, not a nitpick.\n\nThere are two further issues. No grid-convergence study is reported for the ADI benchmark (Nx=101, Ny=51, dt=1e-4), and no comparison is made to the analytical reactive-dispersion solutions the paper cites, so the benchmark is unverified. The error tables are also inconsistent: the non-reactive Poiseuille line-like case has L2 error 0.024 at t=0.001 in Table 1 but 0.0017 in Table 2, and one Couette-Poiseuille point-like entry repeats exactly at t=0.1 and t=0.5. That makes the quantitative claims hard to trust as reported.\n\nCredit where due: the authors are honest about limitations - Section 5 concedes the dispersion coefficient is prone to local approximation errors - and the supplementary material is thorough. The paper is not a waste of time; it is a competent draft for the computational-transport audience. I would not cite it in its current form, and the revision bar is high. A serious editor could send it to review, but the authors should first show the diagnostics are unchanged when Lx is doubled, add an ADI convergence study, and fix the error tables.","headline":"Competent but incremental PINN application; the PINN-ADI agreement likely reflects a shared artificial-boundary error rather than true accuracy, so the validation needs a domain-window check and a convergence study before the central claim can stand.","tokens_in":34874,"tokens_out":5059,"would_cite":false,"duration_ms":45941,"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 physics-informed neural network trained on the convection-diffusion equation and Robin wall conditions reproduces the spatiotemporal concentration field of wall-reactive solute dispersion in Couette, Poiseuille, and Couette–Poiseuille…","keywords":["physics-informed neural networks","solute dispersion","shear flows","Robin boundary conditions","wall uptake","convection-diffusion equation","alternating-direction implicit method","Péclet number"],"falsifier":"Refine the ADI grid (for example, double $N_x$ and $N_y$ and halve $\\Delta t$) or compare against an independent spectral or high-order solution and an analytical reactive-dispersion solution; if the PINN–ADI pair remains close to each other but both shift when the streamwise window is widened beyond $[-5,5]$ or when the grid is refined, the claim that the PINN reconstructs the true concentration field is refuted.","tokens_in":33841,"feed_emoji":"🌊","tokens_out":11651,"duration_ms":94438,"temperature":0.7,"pith_summary":"The paper asks whether a physics-informed neural network can replace a grid-based solver for two-dimensional dispersion of a solute that reacts at the channel walls. It claims yes: with the convection-diffusion equation and first-order wall-absorption conditions folded into one loss, the network reproduces the concentration field for Couette, Poiseuille, and Couette–Poiseuille flows at Pe=10, matching an alternating-direction implicit finite-difference benchmark. Because the network is differentiable, the same trained field yields transport diagnostics such as the apparent axial dispersion coefficient and cumulative wall-removal rates without discrete differentiation. The physical payoff is a precise picture of how shear determines where along the wall solute is removed and how unequal wall reactivities break transverse symmetry.","feed_headline":"Neural net matches finite-difference solver for wall-reactive plumes","feed_subtitle":"Physics-informed network recovers concentration fields and wall uptake in Couette, Poiseuille, and mixed shear flows.","key_machinery":"The load-bearing object is the composite physics-informed loss $$\\mathcal{L}_{\\rm tot} = \\mathcal{L}_{\\rm PDE} + \\lambda_{\\rm IC}\\mathcal{L}_{\\rm IC} + \\lambda_{\\rm B}(\\mathcal{L}_{x\\rm B}+\\mathcal{L}_{y\\rm B}),$$ minimized by a four-hidden-layer, 64-neuron tanh network. A PINN is a neural network trained to minimize the residual of the governing partial differential equation together with its boundary conditions. The PDE residual is evaluated by automatic differentiation, and the wall residuals enforce first-order uptake at the two plates. The three mean-centred shear profiles $u^{(\\rm C)}=y/2$, $u^{(\\rm Po)}=\\tfrac12(1-y^2)-\\tfrac13$, and $u^{(\\rm CP)}=u^{(\\rm C)}+u^{(\\rm Po)}$ are what distinguish the flow cases. This loss is what lets a mesh-free function approximator enforce the transport physics, and the trained network's differentiability is what turns the solution into a source of flux and moment diagnostics.","core_discovery":"The central discovery is that a single feed-forward network, constrained only by the governing PDE and boundary residuals, can serve as an accurate and differentiable surrogate for wall-reactive solute dispersion. The network output $C_\theta(x,y,t)$ minimizes a composite loss containing the dimensionless convection-diffusion residual $\\partial_t C + \\mathrm{Pe}\\,u^{(s)}(y)\\partial_x C - \\partial_{xx} C - \\partial_{yy} C$, the smooth Gaussian initial source, zero-gradient axial conditions at $x=\\pm 5$, and Robin wall conditions $\\partial_y C-\\beta_2 C=0$ at $y=-1$ and $-\\partial_y C-\\beta_1 C=0$ at $y=1$. Over $t\\in[0,1]$, the learned field stays close to the ADI benchmark for both line-like and point-like sources across non-reactive, symmetric, and asymmetric wall-reactivity regimes, with final cumulative wall-uptake relative errors below about 4%. The differentiable representation then makes wall-resolved diagnostics, such as the apparent dispersion coefficient, cumulative uptake, and flux-weighted centroids, directly computable. The paper's physical claim is that the imposed shear profile controls the streamwise organization of reactive uptake, while unequal wall reactivities control which wall dominates removal.","pith_inferences":["The forward machinery could be inverted: a PINN trained on noisy concentration or flux observations could infer unknown wall reactivities $\\beta_1,\\beta_2$; the paper builds the differentiable surrogate but does not attempt the inverse problem.","The smooth Gaussian sources are a practical bridge to point releases: weak-form or distributional PINN formulations could extend the same framework to Dirac-delta initial conditions, making classical point-release dispersion benchmarks directly accessible.","The flux-weighted centroid and spread diagnostics are natural design targets for microfluidic or physiological transport, for example selecting a shear profile that delivers solute to a prescribed streamwise uptake zone; this application lies outside the paper.","The Pe=10 results do not automatically extend to advection-dominated regimes; at higher Pe the network faces stiffer residuals, so the claimed mesh-free advantage should be re-tested where advection is stronger."],"forward_implications":["At Pe=10, the trained PINN reproduces the ADI concentration field for line-like and point-like Gaussian sources in Couette, Poiseuille, and Couette–Poiseuille flows, in non-reactive, symmetric, and asymmetric wall-reactivity regimes.","The same differentiable network yields $D_a(t)=\\frac{1}{2}\\frac{d\\nu_2}{dt}$, cumulative wall uptake $U_{\\rm tot}(t)$, and wall-resolved fluxes $J_\\pm(x,t)$ without grid-based derivative approximations.","Unequal wall reactivities break transverse symmetry of the plume: for $(\\beta_1,\\beta_2)=(0.2,2)$ the lower wall dominates removal, and reversing the pair reverses the dominance, as quantified by $\\Phi_-(T)$ and $D_w(t)$.","The shear profile controls the streamwise organization of uptake: Couette shear separates upper- and lower-wall uptake centroids, whereas Poiseuille shear keeps them overlapping, so total removal alone does not characterize reactive dispersion."],"supporting_citations":[{"why":"Introduces the physics-informed neural network idea of embedding PDE residuals in the loss; the paper's composite loss builds on it.","marker":"Raissi et al. 2019"},{"why":"Demonstrates PINNs for forward and inverse advection-dispersion equations, the equation class the paper targets.","marker":"He & Tartakovsky 2021"},{"why":"Provides the alternating-direction implicit integration scheme used as the finite-difference benchmark.","marker":"Douglas 1955"},{"why":"Provides the alternating-direction splitting that the benchmark relies on.","marker":"Peaceman & Rachford 1955"},{"why":"Moment method that underlies the apparent axial dispersion coefficient $D_a(t)$.","marker":"Aris 1956"},{"why":"Supplies the boundary-retention model that the paper adopts as its Robin wall conditions.","marker":"Purnama 1988"}],"fun_headline_variants":["PINN matches benchmark for reactive plume dispersion in shear flows","Physics-informed net solves wall-reactive solute transport in shear","Shear-driven plume dispersion captured by mesh-free neural solver","Neural surrogate predicts wall uptake in Couette-Poiseuille flows","PINN recovers reactive plume dynamics with wall absorption"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ADI finite-difference solution on the chosen grid ($N_x=101$, $N_y=51$, $\\Delta t=10^{-4}$) over the finite window $x\\in[-5,5]$ is an accurate proxy for the true solution, since no grid-convergence study or comparison with an independent analytical reactive-dispersion solution is reported.","fun_headline_variants_meta":{"raw":{"variants":["PINN matches benchmark for reactive plume dispersion in shear flows","Physics-informed net solves wall-reactive solute transport in shear","Shear-driven plume dispersion captured by mesh-free neural solver","Neural surrogate predicts wall uptake in Couette-Poiseuille flows","PINN recovers reactive plume dynamics with wall absorption"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1561,"prompt_tokens":1123,"completion_tokens":438,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":739,"completion_tokens_details":{"reasoning_tokens":356}},"tokens_in":739,"tokens_out":438,"duration_ms":4154,"temperature":1.0,"reasoning_tokens":356,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:16:12.257786+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Refine the ADI grid (for example, double $N_x$ and $N_y$ and halve $\\Delta t$) or compare against an independent spectral or high-order solution and an analytical reactive-dispersion solution; if the PINN–ADI pair remains close to each other but both shift when the streamwise window is widened beyond $[-5,5]$ or when the grid is refined, the claim that the PINN reconstructs the true concentration field is refuted.","supporting_citations":[],"review_version":2}