{"id":"1e5a5239-112e-42b1-ab2a-3cc55e538d3c","arxiv_id":"2506.08622","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A PINN-based mapping generates custom structured grids for irregular domains, enabling PDE solving with FDM and PICNN on rectangular grids.","lead":"This paper uses physics-informed neural networks to generate structured grids on irregular domains by mapping them to rectangles. The approach lets standard finite difference and convolutional neural network solvers handle complex-shaped domains without traditional mesh generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (13) is not the correct transformed Laplacian for the PINN-generated map, so the FDM/PICNN results in Figs. 4--6 solve a different PDE than claimed.","rationale":"The reader's weakest assumption concerned whether the soft Jacobian penalty guarantees global invertibility of the learned map. That is legitimate but not the most load-bearing problem. The fundamental issue is that the paper's own transformation formula, Eq. (13), is not the Laplacian in general curvilinear coordinates; it omits the metric coefficients and cross term and has incorrect first-order terms. Because the entire FDM and PICNN pipeline in Section 2 relies on Eq. (13), the numerical demonstrations in Figs. 4--6 do not show that the method solves the original PDEs on irregular domains. No amount of global invertibility fixes the missing rho^{-2} factor. This is an internal mathematical inconsistency, not a disagreement with prior art, so it moves the verdict from CONDITIONAL to REJECT: the central claim is unsupported as written. A corrected transformation and a quantitative comparison on a nontrivial map would be required before the claim can be assessed.","tokens_in":51,"tokens_out":9316,"duration_ms":238080,"concrete_test":"Use the known polar map x=rho cos theta, y=rho sin theta with rho in [1,2], theta in [0,2pi]. Solve the steady heat equation with Dirichlet data T(1,theta)=cos theta and T(2,theta)=0 on a uniform computational (rho,theta) grid using finite differences and applying Eq. (13) exactly. Compare to the analytic solution T(rho,theta)=((4-rho^2)/(3rho)) cos theta. If Eq. (13) were the correct Laplacian transform, these would match; the missing rho^{-2} factor on the theta derivative produces an O(1) discrepancy for theta-dependent data. This isolates the transformation error without relying on the PINN training.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim requires that solving Eq. (18) with Eq. (13) in computational coordinates is equivalent to solving the original Laplace equation in the physical domain. This equivalence is false for the general non-affine maps produced by the PINN. Applying the chain-rule operators in Eqs. (9)--(12) to T_xx + T_yy yields T_xx + T_yy = [(x_eta^2 + y_eta^2)/J^2] T_xixi - 2[(x_xi x_eta + y_xi y_eta)/J^2] T_xieta + [(x_xi^2 + y_xi^2)/J^2] T_etaeta + (xi_xx + xi_yy) T_xi + (eta_xx + eta_yy) T_eta, where J is the determinant defined in Eq. (14). Expanding Eq. (13) gives T_xixi + T_etaeta + (J_xi/J) T_xi + (J_eta/J) T_eta. These agree only when the metric coefficients are all 1 and J is constant, i.e., for affine or identity maps, not for the curved irregular-domain grids in Figs. 3 and 7--10. For a polar map (x=rho cos theta, y=rho sin theta), J=rho and Eq. (13) becomes T_rhorho + T_rho/rho + T_thetatheta, while the true Laplacian is T_rhorho + T_rho/rho + T_thetatheta/rho^2. Thus the FDM solution of Eq. (18) and the PICNN comparison in Figs. 4--6 do not solve the intended heat or Navier-Stokes equation. This is an internal inconsistency with Eqs. (9)--(12), independent of the sampled-Jacobian invertibility issue identified by the reader.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a physics-informed neural network (PINN) approach for generating structured grids in irregular physical domains by learning a mapping from a rectangular computational domain. The network is trained with a loss combining Laplace smoothing, boundary matching, and a soft Jacobian-determinant positivity penalty. The authors then use the mapping to transform PDEs to the computational domain, solving a heat conduction problem by finite differences and steady Navier-Stokes equations by physics-informed convolutional neural networks, with comparisons against OpenFOAM. The paper further demonstrates customized grids with denser boundary layers, denser/sparser regions around specified points, and merged subdomains.","tokens_in":6361,"tokens_out":8173,"duration_ms":93261,"significance":"The central idea---using a single PINN training to produce a structured grid and then applying classical or CNN-based solvers on a rectangle---is attractive and, if made correct, could be a useful alternative to elliptic grid generation. The loss-function formulation is flexible and the mesh examples in Figs. 3 and 7--10 are visually plausible. However, the paper as written does not establish the validity of the PDE-solving pipeline: the transformed Laplacian in Eq. (13) is incorrect for general non-affine maps, and the numerical comparisons are only qualitative. With a corrected transformation and quantitative validation, the method could be a worthwhile contribution.","major_comments":[{"comment":"Equation (13) is not the Laplacian in the computational coordinates for a general mapping. Starting from Eqs. (9)--(12), the sum ∂²T/∂x² + ∂²T/∂y² contains the metric coefficients g^{ξξ}=(x_η²+y_η²)/J², g^{ξη}=−(x_ξ x_η+y_ξ y_η)/J², and g^{ηη}=(x_ξ²+y_ξ²)/J²; the correct expression is ∇²T = (1/J)[∂_ξ(J g^{ξξ} T_ξ + J g^{ξη} T_η) + ∂_η(J g^{ξη} T_ξ + J g^{ηη} T_η)]. Equation (13) is recovered only if g^{ξξ}=g^{ηη}=1 and g^{ξη}=0, i.e., for an isometric mapping. For a polar map (x=ρ cosθ, y=ρ sinθ), Eq. (13) gives T_ρρ + T_ρ/ρ + T_θθ, whereas the correct Laplacian is T_ρρ + T_ρ/ρ + T_θθ/ρ². Consequently Eq. (18) and the FDM/PICNN results in Figs. 4--6 solve a different PDE from the intended heat and Navier--Stokes problems; this is an internal inconsistency with Eqs. (9)--(12) and invalidates the central claim.","section":"§2.2.1, Eq. (13)"},{"comment":"The Jacobian constraint is enforced only as a soft penalty at N_f sampled collocation points. This does not guarantee det J > 0 over the whole computational domain, and the transformation in Eq. (18) divides by J; a sign change or near-zero J between sample points would break both the grid validity and the PDE solve. The authors should report the minimum Jacobian over a dense validation grid or enforce the constraint in a way that provides a global guarantee.","section":"§2.1, Eq. (6)"},{"comment":"The comparisons with OpenFOAM are purely qualitative; no pointwise error, L2 norm, or convergence study is reported. Since the transformed equation is currently wrong, the apparent agreement cannot be taken as validation. After correcting Eq. (13), the authors should provide quantitative errors against OpenFOAM for the heat conduction and Navier--Stokes examples and a grid-convergence study for the FDM solver.","section":"§2.2.2, §2.3, Figs. 4--6"}],"minor_comments":[{"comment":"The notation uθ(ξ,η) in the boundary loss is inconsistent with the mapping notation (x,y)=NN(ξ,η); please clarify the meaning of uθ.","section":"§2.1, Eq. (5)"},{"comment":"The summation in Eq. (6) is missing a lower limit, and the term inside the max should be written unambiguously as max(0, −(x_ξ y_η − x_η y_ξ)).","section":"§2.1, Eq. (6)"},{"comment":"The sentence 'The term Leq means solving a diffusion equation in computational space' is imprecise; Eq. (4) is a Laplace smoothing term, not a time-dependent diffusion equation.","section":"§2.1"},{"comment":"The expressions for P and Q are singular at (ξ0,η0); please specify how the loss is evaluated at the densification point.","section":"§3.3, Eqs. (29)--(30)"},{"comment":"The manuscript does not report the MLP architecture, activation function, optimizer, learning rate, or training time; without these details the grid-generation and PDE-solving experiments are not reproducible.","section":"§2.1, §2.2"},{"comment":"The title and abstract contain grammatical errors ('PDEs solving', 'PDES'); please copyedit.","section":"Title and Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising grid-generation component, but the PDE transformation error in Eq. (13) is central and affects all reported solver results. I would only consider it for publication after the authors replace Eq. (13) with the correct metric-tensor form and redo the numerical experiments with quantitative validation. Given the scope of the correction, major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is using a PINN to solve the elliptic grid-generation equations in computational space, with extra loss terms to enforce custom grid properties like boundary-layer densification, point clustering, and mesh merging. Those demonstrations (Figs. 7–10) are visually plausible and extend the earlier work of Gao et al. in a real way. If the paper only claimed to generate structured grids this way, it would be a modest but useful contribution.\n\nThe problem is the PDE-solving half. Eq. (13) is not the Laplacian in general curvilinear coordinates. For a map (x,y)→(ξ,η), the correct transformation contains metric coefficients g^{ij} and cross terms. Eq. (13) treats the metric as if it were a single scalar |J| with unit weights, which is only valid when the map is a conformal similarity (uniform scale, no distortion). The grids in Figs. 3 and 7–10 are far from that, so the heat and Navier–Stokes solutions in Figs. 4–6 are not solutions of the stated PDEs. This is not a subtle numerical issue; it is a geometric identity. The polar example in the stress-test makes it concrete: Eq. (13) gives T_θθ where the true Laplacian has T_θθ/ρ².\n\nThe reader's concerns about missing quantitative error norms, no code, and the soft Jacobian penalty are valid but secondary. The transformation error is load-bearing; it invalidates the central claim of solving PDEs on irregular domains via FDM/PICNN.\n\nThat said, the grid-generation part stands on its own and deserves engagement. The paper should be sent to peer review, because the error is instructive and the grid-generation technique may be worth salvaging after the PDE transformation is corrected or the PDE-solving claims are removed. As written, though, I would not cite it and the internal inconsistency with Eqs. (9)–(12) makes it hard to call the paper sound on its own terms.","headline":"The grid-generation idea is promising, but the PDE-transformation formula in Eq. (13) is wrong, so the FDM/PICNN results solve a different equation than claimed.","tokens_in":6860,"tokens_out":4080,"would_cite":false,"duration_ms":43965,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a PINN can learn an irregular-to-rectangular domain mapping, enabling finite-difference and convolutional neural network solvers on complex geometries.","keywords":["physics-informed neural networks","irregular domains","structured grids","domain mapping","finite difference method","physics-informed convolutional neural networks","Jacobian determinant","partial differential equations"],"falsifier":"Take the trained PINN mapping and evaluate $\\det(J)=\\frac{\\partial x}{\\partial \\xi}\\frac{\\partial y}{\\partial \\eta}-\\frac{\\partial x}{\\partial \\eta}\\frac{\\partial y}{\\partial \\xi}$ on a dense grid much finer than the training samples; finding any interior point with $\\det(J)\\le 0$ would show the coordinate transformation folds and the mapped PDE (Eq. 13) is invalid there.","tokens_in":5727,"feed_emoji":"🗺️","tokens_out":7431,"duration_ms":72763,"temperature":0.7,"pith_summary":"This paper proposes that a physics-informed neural network can learn the coordinate mapping between an irregular physical domain and a rectangular computational domain, replacing traditional elliptic grid generation. The network takes coordinates in the rectangle as input and outputs matching coordinates in the irregular domain, trained so boundaries align, the mapping is smooth, and cell Jacobians stay positive. Once the mapping is learned, PDEs stated on the irregular domain are rewritten in the rectangular coordinate system, so finite-difference and physics-informed convolutional neural network solvers can be used instead of unstructured-mesh solvers. The authors demonstrate the idea on heat conduction and steady Navier-Stokes examples, comparing against OpenFOAM finite-volume results. If the approach holds, one PINN training run yields both a reusable structured mesh and the transformation needed to solve a whole class of PDEs on complicated geometries.","feed_headline":"One PINN maps irregular domains onto rectangular grids","feed_subtitle":"The learned mapping lets finite-difference and CNN-based solvers handle complex geometries without unstructured meshes.","key_machinery":"The load-bearing object is the learned coordinate map $G$, implemented as an MLP that takes computational-space coordinates $(\\xi,\\eta)$ and outputs physical-space coordinates $(x,y)$. Its quality is controlled by three loss terms: a Laplacian regularizer for smoothness, a boundary-matching term for conformity to the irregular boundary, and a Jacobian-determinant penalty enforcing $\\det(J)>0$ to keep grid cells from folding. This map replaces the elliptic equations of classical body-fitted grid generation; requirements such as boundary-layer densification or local refinement are encoded directly as modifications to the loss rather than as new PDE systems.","core_discovery":"The central claim is that a single PINN, trained with a composite loss, produces a valid regular-to-irregular mapping $G:\\Omega_c\\to\\Omega_p$ and its functional inverse, so that a structured grid on an irregular domain is obtained by pulling back the regular grid through $G$. The training loss combines a Laplacian smoothing term that keeps grid lines smooth, a boundary term that pins sampled boundary points to their true physical locations, and a Jacobian term that penalizes non-positive cell determinants, with the aim of ensuring $\\det(J)>0$ everywhere. With the mapping in hand, the paper rewrites differential operators in computational coordinates—notably the Laplacian in the form $\\nabla^2 T = \\frac{1}{|J|}\\left(\\frac{\\partial}{\\partial \\xi}\\left(|J|\\frac{\\partial T}{\\partial \\xi}\\right) + \\frac{\\partial}{\\partial \\eta}\\left(|J|\\frac{\\partial T}{\\partial \\eta}\\right)\\right)$—and solves the transformed equation by FDM on a rectangle, mapping results back to the physical domain. The same PINN-generated mapping is also used as the grid for physics-informed convolutional neural networks on steady Navier-Stokes problems. The authors report that the resulting temperature fields and flow solutions match OpenFOAM reference results.","pith_inferences":["An implication the paper leaves implicit is that the approach could extend to three dimensions by adding a third coordinate and a 3x3 Jacobian positivity penalty, since the loss formulation carries over.","The soft Jacobian penalty suggests a natural stress test: checking the learned map on a much finer grid than used in training would reveal whether non-folding holds between sample points.","The same loss-engineering idea could be pointed at time-dependent or moving-boundary problems, where the map would be re-learned or parameterized by time instead of fixed once."],"forward_implications":["For any PDE on an irregular domain, the learned mapping converts the problem into a rectangular-domain problem, so FDM and spectral-type solvers become applicable without building an unstructured mesh.","Customized structured grids—denser boundary layers, refinement around specified points, and merged subgrids—can be produced by adding or modifying loss terms rather than by formulating new PDE-based grid generators.","The same PINN-generated mapping directly supplies the structured grid required by PICNN, so an AI-based PDE solver can run on complex geometries with no additional mesh-generation step.","Because the mapping is continuous, the resulting grids are reusable across different boundary conditions and source terms for the same geometry."],"supporting_citations":[{"why":"Introduces the PINN loss formulation that the authors adapt from solving PDEs to learning coordinate mappings.","marker":"[3]"},{"why":"Supplies the elliptic-mapping and PICNN baseline whose grids and flow solutions the paper compares against its PINN-generated grids.","marker":"[10]"},{"why":"Provides the standard transformation of differential operators to computational space, used to rewrite the PDE under the learned map.","marker":"[11]"}],"fun_headline_variants":["PINN warps irregular domains onto simple grids","One neural net untangles irregular domains for PDEs","PINN mapping lets solvers work on rectangular grids","Irregular to regular: PINN builds the bridge for PDEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that a soft penalty at finitely many sampled points is enough to guarantee the learned mapping is one-to-one and never folds over the entire domain, including regions between sample points.","fun_headline_variants_meta":{"raw":{"variants":["PINN warps irregular domains onto simple grids","One neural net untangles irregular domains for PDEs","PINN mapping lets solvers work on rectangular grids","Irregular to regular: PINN builds the bridge for PDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1273,"prompt_tokens":891,"completion_tokens":382,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":317}},"tokens_in":507,"tokens_out":382,"duration_ms":4948,"temperature":1.0,"reasoning_tokens":317,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:05:28.434650+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained PINN mapping and evaluate $\\det(J)=\\frac{\\partial x}{\\partial \\xi}\\frac{\\partial y}{\\partial \\eta}-\\frac{\\partial x}{\\partial \\eta}\\frac{\\partial y}{\\partial \\xi}$ on a dense grid much finer than the training samples; finding any interior point with $\\det(J)\\le 0$ would show the coordinate transformation folds and the mapped PDE (Eq. 13) is invalid there.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the elliptic-mapping and PICNN baseline whose grids and flow solutions the paper compares against its PINN-generated grids."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the standard transformation of differential operators to computational space, used to rewrite the PDE under the learned map."}],"review_version":1}