acceptodds
Under review as a conference paper at ICLR 2027

PINNsur: Physics-Informed Neural Networks for PDEs on Curved Surfaces

Abstract

Partial differential equations (PDEs) on surfaces are fundamental to scientific computing and geometry processing. A popular approach to solving PDEs on surfaces is the finite element method (FEM), where the surface is divided into discrete geometric elements (usually triangles). Recently, physics-informed neural networks (PINNs) have emerged as a continuous, mesh-free alternative that does not suffer from FEM’s sensitivity to mesh quality or geometric discretization errors. We present PINNsur, a simple framework for using PINNs on curved surfaces: we train a neural field to approximate the surface’s normals, and then we express surface differential operators using their projection from onto the surface. Since every orientable smooth manifold has well-defined normals, our method is suitable for all such surfaces, regardless of curvature or topology, enabling many geometry processing applications. Moreover, despite their empirical success in solving PDEs in flat Euclidean domains, PINNs lack convergence guarantees. This gap restricts their adoption as reliable solvers compared to established methods like FEM, where convergence to the true solution is well understood and theoretically grounded. Surface PDEs are particularly challenging to solve convergently, as one must not only deal with the convergence of the function approximation, but also with the convergence of the geometric approximation of the surface itself. In this work, we empirically show that our formulation to solve surface PDEs results in a convergent behavior with respect to the number of tunable weights in the network, and analyze the behavior of other PINNs using this criterion.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
PINNs, Surface PDEs, Poisson, Helmholtz, Convergence

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zMbqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The submission presents a PINN approach for solving elliptic PDEs on manifold surfaces. The key idea is to approximate a smooth normal field on a given geometry and use it to construct projection-based differential operators for PINN computation. The approach is clear and simple, avoids the need for the surface parameterization used by [Williamson & Mitra 2025], and achieves improved solution accuracy. The paper also investigates network convergence with respect to the number of trainable parameters.

Critical strengths

The main strength, in my view, is the use of a learned normal field to formulate surface-based PINNs. While the underlying differential formulation is classical, this integration provides a simple and parameterization-free formulation for constructing surface differential operators within the PINN framework.

The second strength is its applicability to a broader range of surface topologies compared with [Williamson & Mitra 2025].

Critical weaknesses
  1. The current method is somewhat limited in its applicability. When the boundary conditions g,hg,h are changed, the solution network needs to be retrained, taking considerable time and limiting its practical use. Taking g,hg,h as network inputs would significantly strengthen the proposed method. Furthermore, the normal network could potentially be extended to handle varying geometry as input. In this way, the coupled networks could support a much broader range of problem instances without per-instance retraining.

  2. The paper uses the cotangent Laplacian implementation to criticize the sensitivity of linear FEM to mesh quality. However, this issue has been extensively studied, with techniques such as intrinsic Delaunay triangulation being well established in geometry processing. Furthermore, mesh refinement and mesh-quality improvement are relatively inexpensive operations compared with network training. Therefore, I do not see a clear practical benefit of the proposed PINN approach over mesh-based FEM in the current setting. In addition, traditional meshless approaches can operate directly on point clouds. Overall, I find the argument for the benefits over traditional numerical approaches relatively weak. Addressing Point 1, particularly by avoiding per-instance retraining for different PDE inputs, would make the motivation substantially stronger.

  3. Two closely related spectral-based works are worth citing and discussing: (1) Δ-PINNs: Physics-Informed Neural Networks on Complex Geometries (https://doi.org/10.1016/j.engappai.2023.107324); and (2) Numerical Methods for PDEs over Manifolds Using Spectral Physics-Informed Neural Networks (https://arxiv.org/abs/2302.05322).

  4. The Degrees of Freedom experiment (Fig. 9) provides an asymmetric comparison with FEM. The FEM baseline uses linear elements and increases its degrees of freedom only through mesh refinement, whereas increasing the number of PINN parameters also increases the representation capacity of the neural approximation. A more meaningful comparison should also consider higher-order FEM or p/hp-refinement, which can achieve substantially faster convergence.

  5. Boundary conditions are treated as soft constraints, and the errors near the boundary appear significantly larger than those in the interior in Fig. 9. Furthermore, in many examples, the largest errors remain relatively high, making the results less competitive from a scientific-computing perspective. Although the paper attempts to establish convergence with respect to the network capacity, it remains unclear how large the network needs to be to achieve a desired accuracy in practice, for example, an average error of 10−510^{-5}.

  6. Normal network usage. ∇Ωu\nabla_\Omega u (Eq. 6) depends on nn. Therefore, when autodifferentiating the surface gradient, derivatives of nn with respect to the spatial coordinates are involved as well. The supplementary code also does not stop the gradient through nn (I checked test_convergence.py and python files under /src, didn't find stop-gradient). However, line 194 states that the normals can be “used without any autodifferentiation,” which seems inconsistent with both the formulation and the implementation. I would expect auto differentiation to be necessary for providing the differential of nn; otherwise, the normal network seems unnecessary in most examples where the normals can be retrieved from the mesh faces.

Questions for authors

Given the limited time available for the rebuttal, I would particularly welcome a response to Points 1, 2, and 6, even if only partially.

One suggestion: as the paper only considers two elliptic PDEs, the title could be made more precise rather than suggesting a broader scope of general PDEs.

4 comments

Response (1/2)

We would like to thank the reviewer for their thoughtful comments and feedback. Below, we try to respond to each question in order (questions are written in bold)

The current method … per-instance retraining.

Trained PINNsur models can be reused for new source terms using reduced-basis PINNs (GPT-PINN; Chen & Koohy, 2023). Since the Laplace–Beltrami operator is linear, we use the last-hidden-layer features of several trained models as a basis and solve for their coefficients with a single linear least-squares fit, without training. Using two models (150×3) trained for the Poisson equation on spot (Fig. 3f) with sources sin(πx)sin(\pi x) and sin(4πx)sin(4\pi x), we solve for the unseen sources sin(2πx)sin(2\pi x) and sin(3πx)sin(3\pi x):

Source termsL2 error using least square fitL2 error after training 
f=sin(2πx)f = sin(2\pi x)9.574e-036.711e-04
f=sin(3πx)f = sin(3\pi x)3.052e-022.542e-03

The least-squares fit obtains solutions for new source terms in a few seconds instead of minutes of training, at the cost of roughly an order of magnitude in accuracy.

In general, we agree that per-instance retraining is a limitation of our method. But, it is shared by many PINNs like standard PINNs (Raissi et al., 2019), variational formulations such as Deep Ritz (Yu et al., 2018) and hp-VPINNs (Kharazmi et al., 2021), and recent surface PDE methods (Hu et al., 2024; Williamson & Mitra, 2025). In our case, the normal network depends only on the geometry and can be reused across all source terms and boundary conditions on a surface.

Removing per-instance retraining is precisely the goal of neural operators like DeepONet (Lu et al., 2021), Fourier Neural Operator (Li et al., 2021), and physics-informed variants (Wang et al., 2021; Li et al., 2024). Extending PINNsur in this direction is a promising direction for future work. We will state this limitation explicitly and discuss neural operators as future work in the revision.

The paper uses the cotangent Laplacian … substantially stronger.

We agree that, given a valid manifold mesh, element quality can be improved cheaply (e.g., with intrinsic Delaunay triangulations), and that FEM is then faster and typically more accurate than our method, as stated in our Limitations. We will revise the paper to avoid overstating mesh-quality sensitivity.

Our setting is one where such a mesh is not readily available. PINNsur only requires point samples and normals, so it applies to any explicit representation that supports sampling, such as triangle meshes (including non-manifold, self-intersecting, or multi-component ones) and point clouds (Fig. 5), with no need for connectivity, manifoldness, or watertightness. Converting raw scans into analysis-ready meshes remains hard (Berger et al., 2017; Huang et al., 2024; Sulzer et al., 2024), and even curated collections are often unsuitable. For example, in Thingi10K, about 45% of models self-intersect and 29% have multiple components (Zhou & Jacobson, 2016).

We agree that classical meshless methods also operate on point clouds, and we do not claim to outperform them. Our work instead contributes to PINNs, an active area of scientific machine learning (Raissi et al., 2019; Karniadakis et al., 2021), where a central open problem is the lack of reliable convergence. By showing convergent behavior for surface PDEs as network capacity increases, our work advances the state of the art on this problem, a contribution in its own right.

Response (2/2)

Two closely related spectral-based works …. 

We will add the citations and discuss them in the related work section upon revision.

The Degrees of Freedom … faster convergence.

As suggested, we report the errors comparing with quadratic elements (higher-order FEM) for the example shown in Figure 9 below:

Degrees of FreedomRelative L2 Error (ours)Relative L2 Error (linear FEM)Relative L2 Error (quadratic FEM)
29110.03970.11120.1566
78510.01950.03810.0325
151910.01510.01920.0121
249310.01130.01160.0058
370710.00700.00770.0031
516110.00600.00550.0019
685510.00730.00410.0012

This experiment verifies that quadratic FEM achieves faster convergence compared to linear FEM and our method. We will add this experiment upon revision.

Also, we would like to clarify that Fig. 9 is not intended to show that PINNsur outperforms FEM and our Limitations section explicitly notes that our method does not always beat FEM in absolute error or runtime. Its purpose is to show that PINNsur exhibits convergent behavior, i.e., the error decreases systematically as the degrees of freedom grow, which not all prior PINN baselines do (Fig. 9). We use linear FEM as the reference because it is the standard discretization in geometry processing and its convergence on surfaces is well established (Dziuk, 1988). We will clarify this in the revision.

Boundary conditions … error of 1e-5.

Boundary errors. We agree that soft boundary constraints lead to larger errors near the boundary, a known limitation of PINNs. Hard constraints, which impose Dirichlet conditions exactly via a distance-function ansatz (Lagaris et al., 1998; Sukumar & Srivastava, 2022), would remove this error, but they are considerably harder in our setting. On curved surfaces, one requires distance to the boundary curve that must itself be computed (e.g., via the heat method, Fig. 1b), and Neumann conditions require more complex constructions. We therefore use soft constraints, which apply uniformly to both boundary condition types on arbitrary surfaces. Incorporating hard constraints to reduce boundary error is a promising direction that we will discuss as future work.

Accuracy. We agree that our errors on some surfaces are not competitive with FEM on meshes, and we do not claim otherwise (addressed in the limitation section). Our goal is to establish convergent behavior with respect to network capacity, which prior PINNs lack.

Required network size. We observe that the network size needed for a given accuracy depends on the problem and is different across surfaces, source terms, and boundary conditions (slopes between [−0.3 and −1.5] in Fig. 4). This is analogous to FEM, where the convergence rate is known a priori but the constant depends on the unknown solution and the geometry, so the number of nodes required for a target accuracy is likewise not known beforehand. Our convergence study enables the same procedure by fitting e≈C⋅Wse \approx C·W^s on a few network sizes gives the number of weights needed for a target error, W≈W0(e0/e)1/∣s∣W \approx W_0(e_0/e)^{1/|s|}.

Normal network usage … from the mesh faces.

We thank the reviewer for finding this mistake. The statement on line 194 that the normal network is “used without any autodifferentiation” is an error in the writing but our formulation and implementation is correct because without the derivatives of nn, the curvature term of the Laplace–Beltrami operator would be missing and the solver would converge to an incorrect solution. We will remove this statement during revision.

Reviewer_zMbqOpenReview ↗

Thank you for the detailed rebuttal

Thank you for the detailed rebuttal in both threads. I appreciate your honest responses to my questions and the additional experiments, even when the results (e.g., comparisons with FEM) do not favor your method. I particularly appreciate that you treat the rebuttal as genuine scientific communication.

Regarding the experiment addressing per-instance retraining, the unseen source terms are not very different from the training ones, yet the accuracy drops by approximately one order of magnitude. Therefore, although the experiment demonstrates some generalization capability, its practical usefulness remains questionable. However, I understand that you do not aim to address this issue in this work.

I understand your position that the goal is not to compete with FEM (and that you will revise the FEM comparison and related text), but rather to contribute to the convergence behavior of surface PINN. Some comparisons do show that other methods have difficulty converging. However, I still have concerns about whether the convergence study represents a sufficiently strong scientific contribution, especially given that the achieved accuracy remains limited for practical use, the convergence rate barely matches that of linear FEM, and there is no guarantee that increasing network capacity will lead to lower errors in practice (as many convergence plots are not monotonically decreasing).

I will discuss these concerns with the other reviewers during the discussion phase.

We thank the reviewer for their comment. We would like to note that the primary contribution of this work is the formulation of the surface operator by learning the normal field directly. Unlike prior surface PINNs, which require a spherical parametrization (restricting them to genus-zero surfaces) or an explicit level-set function (restricting them to closed surfaces with a known level set), our formulation needs only point samples and normals. Our formulation therefore handles surfaces of arbitrary topology, open or closed, with Dirichlet and Neumann boundary conditions, and operates directly on meshes and point clouds. This makes it directly applicable to many geometry processing tasks traditionally solved with FEM (details in Sec. 4.3) like minimal surface computation (Fig. 8a), harmonic interpolation (Fig. 8b), heat diffusion and heat-based geodesic distances (Fig. 1b), and mean curvature flow (Fig. 1a). The convergence study complements this formulation by showing that its accuracy improves systematically with network capacity.

Reviewer_iZPcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of approximating solutions to PDEs on curved surfaces using coordinate-based neural networks, focusing on the Poisson and Helmholtz equations, with applications to geometry processing. First, it fits a neural normal field to sampled surface point–normal pairs. This network remains fixed while a PINN learns the PDE solution with soft boundary constraints. The paper also studies empirical error decay with network degrees of freedom, drawing an analogy with FEM refinement.

Critical strengths
  • Learning normals instead of a neural SDF reduces geometry derivatives from second to first order, simplifying differentiation during PDE training. The experiments involving a heat step and mean curvature flow are interesting.

  • The paper compares PINNsur with FEM, including on degraded meshes (Figs. 4, 5, 9). Figures and surface renderings are very helpful.

Critical weaknesses
  • The reported 2 h 40 min of normal-network pretraining is an important limitation. Also, comparing learned normals with gradients from neural SDFs would be helpful. M-plicits [1], which reports average surface-fitting times of 4.07–8.53 min, could be a baseline. Finally, have the authors tried a shallow SIREN with, say, 2–3 hidden layers of 256 neurons and ω0<30\omega_0<30 for the normal field? If so, how did it affect training time and PDE accuracy?

  • The convergence analysis is somewhat weak, as it is unclear whether the observed error decay reflects approximation order or training effects.

Questions for authors
  • In Sec. 3.4, please specify ω0\omega_0 for the normal and PDE networks, including any layer-dependent settings.

  • NISE [3] also addresses mean curvature flow through neural level-set evolution. Should it be a baseline for the smoothing experiment?

  • Have you explored the problem of surface reconstruction from oriented point clouds, in the spirit of Poisson surface reconstruction [4]?

  • Williamson and Mitra [2] demonstrate heat and mean curvature flow. Please specify how their framework was adapted for the PDE comparison in Sec. 4.2 (Fig. 7).

  • I think Eq. (12) should use “−” instead of “+”. For Helmholtz, should it also include a k2∫ΩudAk^2\int_\Omega u dA term?

3 comments

Response (1/2)

We would like to thank the reviewer for their thoughtful comments and feedback. Below, we try to respond to each question (questions are written in bold)

The reported 2 h 40 min of normal-network pretraining is an important limitation. Also, comparing learned normals with gradients from neural SDFs would be helpful. M-plicits [1], which reports average surface-fitting times of 4.07–8.53 min, could be a baseline. Finally, have the authors tried a shallow SIREN with, say, 2–3 hidden layers of 256 neurons and ω0<30\omega_0<30 for the normal field? If so, how did it affect training time and PDE accuracy?

Comparing with M-plicits. We trained M-plicits on the closed spot model (Fig. 3f) and replaced our normal network with the normalized SDF gradient n=∇f/∣∇f∣n = \nabla f / \\|\nabla f\\|, keeping the PDE network and training setup unchanged:

PDE network sizeRelative L2 error (M-plicits normals)Relative L2 error (our normal network)
30 x 30.37930.5504
50 x 30.12440.1227
70 x 30.02590.0599
90 x 30.03390.0332
110 x 30.02160.0483
130 x 30.02010.0279
150 x 30.01460.0216

With both normal fields, the errors are comparable and decrease as the PDE network grows. PINNsur is thus agnostic to the source of the normal field and converges, consistent with Fig. 13, where the PDE error stops changing once the normal network is expressive enough.

Also, M-plicits do not extend to surfaces with boundary. A signed distance function requires a well-defined inside and outside, so it cannot represent open surfaces or provide the boundary conormal needed for Neumann conditions. Our normal networks handle any orientable surface, with Dirichlet or Neumann boundaries. We will add this comparison to the revision.

Time to train the normal network. The reported 2h 40m corresponds to a fixed budget of 10^7 iterations, used uniformly for the normal network in all experiments, rather than to the time required for convergence. We deliberately chose a very large budget so that the normal network is fully converged in every case and is not the limiting factor to study convergence with respect to the size of the PDE network.

Surfaces with smooth curvature (like the ellipsoid, heightfield, duck, and high genus ball (Fig. 7)) tend to converge much faster than surfaces with sharp details like the lion model (Fig. 6). The table below shows time to convergence for different models:

ModelTime to convergence for the normal network
Ellipsoid3s
Heightfield5s
Duck2m 8s
High genus ball10m
Lion2h 10m

Experiment with different ω0\omega_0. Regarding the Siren network, we use ω0=1\omega_0 = 1 for all layers of the normal network, and ω0=30\omega_0 = 30 for the first layer and ω0=1\omega_0 = 1 for the hidden layers of the PDE network.

Following the reviewer’s suggestion, we trained a shallow SIREN normal network (2 hidden layers of 256 neurons each) with different values of ω0≤30\omega_0 \leq 30 and used it to solve the Poisson equation on the spot model (Fig. 3f):

ω0\omega_0Number of iterations until convergenceTraining time for the Normal networkPDE error using 150x3 PDE network
14461909m 40s0.052
10585701m 16s0.047
3055588s0.040

The following are the errors and the time to solve the Poisson equation on the intricate lion model (Fig. 6) 

ω0\omega_0Number of iterations until convergenceTraining time for the Normal networkPDE error using 150x3 PDE network
157476272h 6m0.467
1010073523m 5s0.579
301694713m 43s0.494

Increasing ω0\omega_0 speeds up normal-network training substantially, while the PDE error remains comparable. This is consistent with Fig. 13 (Appendix B.1), where we show that once the normal network approximates the normals sufficiently well, the PDE error no longer depends on the normal network’s parameters. 

We thank the reviewer for this suggestion, which removes the limitation, and we will add this experiment to the revision and change our default ω0\omega_0 to 30 for the first layer of the normal network.

Response (2/2)

The convergence analysis is somewhat weak, as it is unclear whether the observed error decay reflects approximation order or training effects.

All networks are trained with identical parameters (same optimizer, learning-rate schedule, collocation points, and 150k iterations), and each is trained to saturation (Appendix C, Fig. 15). The reported errors therefore reflect converged networks rather than differences in training. Under this fixed protocol, the error decreases with network size across all surfaces, PDEs, boundary conditions, and test functions (Fig. 4, Appendix E, F), and depends on the number of weights rather than on how they are split between width and depth (Fig. 12). In contrast, baseline PINNs trained with the same refinement strategy do not converge (Fig. 9), indicating that the convergent behavior stems from our formulation rather than from the training setup. 

In Sec. 3.4, please specify ω0\omega_0 for the normal and PDE networks, including any layer-dependent settings.

For the normal network, we use ω0=1\omega_0 = 1 for both the first and hidden layers. For the PDE network, we use ω0=30\omega_0 = 30 for the first layer and ω0=1\omega_0 = 1 for the hidden layers.

Considering the experiment in the first response, we will revise our normal network to use ω0=30\omega_0 = 30 for the first layer of the normal network as well. We will add these details in Sec. 3.4 during revision. 

NISE [3] also addresses mean curvature flow through neural level-set evolution. Should it be a baseline for the smoothing experiment?

We believe that NISE and our mean curvature flow example (Figure 1a) address different problems. NISE represents the evolving surface implicitly, as the zero level set of a space-time network fθ(x,t)f_θ(x, t), trained on the level-set equation in R3R^3 × time. Our method, by contrast, is designed for static PDEs with no time component. Our mean curvature flow example, therefore, uses PINNsur as a drop-in Laplace–Beltrami solver within the classic implicit-fairing scheme (Desbrun et al., 1999). Time is discretized outside the network, and each step reduces to a static elliptic PDE (Eq. 15) solved for the coordinate functions directly on the current surface. Our goal in Section 4.3 is to show that PINNsur can be used in standard geometry-processing pipelines and not to propose a new curvature-flow method.

Have you explored the problem of surface reconstruction from oriented point clouds, in the spirit of Poisson surface reconstruction [4]?

We thank the reviewer for the suggestion and we will add this application to future work. We currently did not consider Poisson surface reconstruction, as it solves a volumetric problem. It seeks an indicator function X:R3→R\mathcal{X} : R^3 \to R whose gradient matches a vector field VV obtained by splatting the oriented normals. PINNsur instead solves PDEs on a given surface, projecting derivatives onto its tangent plane. This does not carry over to reconstruction since X\mathcal{X} is constant on its level set, hence its surface derivatives vanish and therefore all relevant information lies in the normal direction, which our projection removes.

Williamson and Mitra [2] demonstrate heat and mean curvature flow. Please specify how their framework was adapted for the PDE comparison in Sec. 4.2 (Fig. 7).

Williamson & Mitra (2025) obtain surface normals from the Jacobian of a learned surface map. For the comparison in Section 4.2 (Fig. 7), we reimplemented their method where we train their surface map and compute normals from its Jacobian. We then train our normal network on these Jacobian-based normals, keeping all of its parameters unchanged, and the PDE network uses it to solve the surface PDE. The normal information from Williamson & Mitra is thus the only difference from our method and everything else is identical. The results show that the surface map introduces errors that lead to noisy normals which results in a higher error for the PDE solution, whereas learning the normals directly yields a lower error. Moreover, since Williamson & Mitra parameterize the surface over a sphere, their method is restricted to genus-zero surfaces, while ours applies to any orientable surface (Fig. 7). We will add these details in Sec. 4.2 upon revision.

I think Eq. (12) should use “−” instead of “+”. For Helmholtz, should it also include a k2∫udAk^2\int udA term?

We thank the reviewer for pointing out the typo in Eq. 12. We will fix it upon revision.  No, for the Helmholtz equation, a unique solution exists for any data provided k2k^2 is not an eigenvalue of −ΔΩ-\Delta_\Omega. We will add this detail in Section 3.3.

Thank you for the detailed response. I’m happy with your answers.

Reviewer_9GVuOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work represents a surface PDE solution with a neural field and then uses a learned network to approximate normals to construct the differential operators on the surface. This way, we can use surface samples and normal information without the need for parameterization or level set descriptions. This work considers Poisson and Helmholtz equations with different boundary conditions, and have demonstrated extensive results on how it behaves on applications such as on harmonic interpolation, heat-based distances and surface evolution. The approach is very practical. It is clearly presented and relatively easy to follow.

Critical strengths
  1. It shows a practical and interesting way of combining a normal prediction network and the PDE network, to avoid global parameterization.
  2. The method showed extensive support for Dirichlet and Neumann conditions.
  3. The paper provided a wide range of experiments on different equations, and showed a variety of applications.
  4. The ablation of SIREN and Tanh is informative and interesting.
Critical weaknesses
  1. My main concern about the contribution is its positioning against existing surface PINNs. Tang and Fu (2021) already use ambient projections and surface PDE residuals, and Hu et al. (2024) already solve PDEs on nonplanar surfaces. The paper cites these works, but the conclusion that "Prior PINNs either cannot handle surfaces or are limited to genus zero" is too broad. I would frame the advance around learning the normal field, reducing geometric prerequisites, and studying the resulting solver.
  2. I am more convinced by the successful examples than by the stronger interpretation of mesh independence. In particular, the claim about avoiding geometric discretization error needs qualification when normals are fitted to a mesh. Refining the solution network while keeping the learned geometry fixed shows an approximation trend for that setup. However, it does not establish that the geometric operator is becoming more accurate. I believe that “mesh-free” does not mean “free of geometry error”.
  3. The surface operator also uses derivatives of the normal field, so fitting its directions alone does not establish operator accuracy. The additional exact vs learned normal experiments are helpful, and I am not asking for a general proof. What would strengthen the practical case is a clearer account of when learned normals improve the trade-off, including the already-disclosed geometry preprocessing cost and which parts can be reused across PDE solves.
Questions for authors
  1. Could you state more precisely what is gained over existing extrinsic surface PINNs such as Tang and Fu and Hu et al., and explain which capabilities follow from learning the normal field? In particular, which results best isolate the benefit of that replacement from the solution-network architecture and training objective?

  2. For one representative geometry, what is the accuracy and cost trade-off between exact and learned normals when geometry preprocessing and PDE training are both included? It would also help to distinguish a first solve from repeated solves on the same surface. This would make the practical use case clearer without requiring superiority over FEM on every problem.

  3. Please clarify the Helmholtz reference system and solvability assumptions. Appendix A's reaction term appears inconsistent with its stated mass-matrix convention. Are these presentation errors, or do any reference solutions or comparisons need updating?

1 comment

We would like to thank the reviewer for their thoughtful comments and feedback. Below, we try to respond to each question (questions are written in bold)

I am more convinced … not mean “free of geometry error”.

The reviewer is correct to point out that the statement in the abstract (line 026) that PINNs do not suffer from “geometric discretization errors” is inaccurate. It is the only place where we make this claim, and we will correct it upon revision.

We call our method “mesh-free” because our model only requires sampling points (we do not need a mesh, as in FEM). Our model can work directly on point clouds (Fig. 5) or on any explicit surface representation that allows point sampling. We will clarify this point upon revision.

My main concern about the contribution is its positioning against existing surface PINNs ... and studying the resulting solver 

Could you state more precisely … training objective?

Tang & Fu (2021) and Hu et al. (2024) obtain the surface normal field from a level-set representation of the surface and then use it to construct the surface differential operators. These approaches, therefore, require an explicit level-set function, which is known only for specific surfaces. Moreover, a level set of a function on R3R^3 defines a closed, watertight surface, so these approaches do not directly handle surfaces with boundary, which appear in many of our experiments (Fig. 3, 4).

Our framework does not require a level-set representation. It only requires a smooth, normal field, which it learns from normal data, and it is agnostic to where that data comes from. In our response to reviewer iZPc, we show that our method works similarly even when normals are computed using a different method like M-plicits. This reduces the geometric prerequisites of the method and allows it to handle a broad range of surfaces, including closed and open surfaces of arbitrary genus and geometrically complex shapes with different boundary conditions (Figs. 1, 3, 6, and 7).

Regarding our claims about prior PINNs, we agree with the reviewer and will remove the overly broad statement that prior PINN approaches cannot handle surfaces or are limited to genus-zero geometries. We will instead position the main contribution of our work as learning the surface normal field, which removes the need for an explicit level-set formulation, together with a study of the convergence of the resulting surface PDE solver.

The surface operator also uses derivatives of the normal field ... can be reused across PDE solves

For one representative geometry, … FEM on every problem.

In section F of the appendix, we have presented additional experiments where, for the heightfield and ellipsoid, we consider three variants: (i) using the exact surface normals directly, (ii) training an MLP to approximate the exact surface normals, and (iii) training an MLP to approximate normals computed from a triangle mesh. The experiment results are present in the shared code (https://anonymous.4open.science/r/PINNsur_ICLR2027/README.md). 

Comparing variants using exact vs. learned normals, we observe convergence for all examples (see Table 1 in the appendix), with errors of similar magnitude, indicating that learning normals does not introduce error in the PDE solution. More details regarding these experiments are mentioned in Section F of the appendix. 

Trained PINNsur models can be reused for new source terms without retraining, following reduced-basis PINNs (GPT-PINN; Chen & Koohy, 2023). We ran this experiment on the spot model (Fig. 3f), where unseen source terms are solved in seconds, with about an order-of-magnitude lower accuracy than training from scratch. The details and results are in our response to Reviewer zMbq (first point). Removing per-instance retraining entirely is the goal of neural operators, which we discuss as future work in the same response to Reviewer zMbq.

Please clarify the Helmholtz … need updating?

We thank the reviewer for pointing it out. Eq. 19 currently is (−L+k2I)uGT=Mf(-L + k^2I)u_{GT} = Mf but should read (−L+k2M)uGT=Mf(-L + k^2M)u_{GT} = Mf. This is a typo in the paper, and we will correct the equation upon revision. Our implementation (can be verified here: https://anonymous.4open.science/r/PINNsur_ICLR2027/test_computed_parameterization/test_convergence.py#L604) uses the correct formulation, so the empirical results shared need no updating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.