acceptodds
Under review as a conference paper at ICLR 2027

CASIN: Curvature-Aware SImplicial Networks for Time-Dependent PDEs on Unknown Riemannian Manifolds

20 views

Abstract

Partial differential equations (PDEs) are central to mathematical descriptions of physical, biological, and environmental processes, providing a means of representing how systems evolve across multiple interacting variables. In practice, however, observations of such systems are frequently sparse, irregularly distributed, and collected over domains whose geometry is not known in advance. This presents a fundamental challenge in applications such as biomedical imaging, where disease-related patterns develop over curved anatomical surfaces, and environmental monitoring, where observations may be obtained from weather stations distributed unevenly across complex terrain. Although these non-Euclidean settings occur throughout fields ranging from wildfire modeling to healthcare, computational approaches for solving PDEs on unknown geometric domains have received substantially less attention than methods developed for conventional Euclidean spaces. We develop a learning-based framework for approximating the dynamics of time-dependent PDEs directly from observations sampled on unknown Riemannian manifolds. Rather than imposing a predefined coordinate system or requiring explicit knowledge of the underlying geometry, the proposed approach constructs a simplicial representation of the observed domain and uses discrete exterior calculus to approximate the differential operators governing the dynamics. Simplicial convolution provides the principal mechanism for propagating information across this geometric representation, enabling the resulting neural solver to account explicitly for local curvature and higher-order geometric structure. The framework is evaluated on nonlinear reaction–diffusion systems exhibiting complex pattern formation over both synthetic manifolds and real tooth surfaces. In comparisons involving eight state-of-the-art neural PDE solvers across nine experimental configurations, the proposed method produces the smallest error in seven cases. It reduces RMSE substantially relative to the strongest competing method that explicitly incorporates curvature and achieves nearly an order-of-magnitude improvement over approaches that do not account for the underlying geometry. Experiments using real dental scans further show that the proposed approach achieves the best performance for all three reaction–diffusion systems. Finally, we assess its generality on a nonlinear advection equation describing fluid transport on manifolds, demonstrating that the framework extends beyond reaction–diffusion dynamics to other classes of time-dependent PDEs on complex non-Euclidean domains.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science

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_bkUgOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes CASIN for learning the evolution of time-dependent PDEs on unknown Riemannian manifolds from irregularly distributed observations. The authors motivate the work by noting that most existing numerical PDE solvers focus on Euclidean domains, while modeling PDE dynamics on manifolds with incomplete geometric information remains challenging.

CASIN first reconstructs local geometry and a simplicial complex from point clouds. It then incorporates Gaussian curvature, mean curvature, and geometric features of edges and triangles into a simplicial neural network. Hodge-Laplacian-based convolution is used for spatial information propagation, followed by an MLP decoder that predicts the state increment over one time step. The model is trained using multi-step trajectory prediction errors. It learns a discrete-time evolution map rather than explicitly identifying the governing PDE operators.

The method is evaluated on several reaction–diffusion systems and shallow-water dynamics over synthetic manifolds and real tooth geometries. The paper reports improvements over existing neural PDE methods and includes component ablations, geometry reconstruction tests, and computational analysis.

Critical strengths
  1. A potentially useful integration of curvature and simplicial structures

CASIN integrates local geometric information with higher order spatical connectivity within a unified neural prediction framework. By incorporating curvature, edge features, and triangle relationship, the model can exploit geometric information that is not directly available to conventional node wise predictors. Although the individual components are estabilished techniques, their conbiation may be useful for learning PDE dynamics on irregular manifolds.

  1. Experiments on multiple geometries and dynamical systems

The experimental evaluation covers several reaction diffusion systems and shallow water dynamics on synthetic manifolds and real tooth geometries. Table 1 reports that CASIN achieves the lowest mean RMSE in seven of nine reaction diffusion settings. The component ablation in Table 2 also suggests that edge level propagation and curvature features contribute to improved prediction accuracy.

Critical weaknesses
  1. The methods mainly combines estabilished components, while its independent methodological contribution remains unclear

CASIN integrates point cloud reconstruction, curvature estimation, Hodge Laplacians, simplicial convolution, and residual state prediction. While this combination has practical value, the paper does not sufficiently establish whether the authors introduce a new mathematical mechanism, network architecture, or learning principle beyond these existing components.

Using manifold geometry to constrain dynamical learning is also not entirely new. Related works such as 'Learning vector fields of differential equations on manifolds with curvature-adaptive geometrically constrained SINDy', etc., have explored geometric constraints and curvature information in dynamics modeling. These works mainly concern vector fields on state manifolds, whereas CASIN studies PDEs defined over spatial manifolds. Nevertheless, discussing their relationship and differences would help clarify the additional contribution of the proposed approach.

More importantly, the current ablation study does not fully establish the necessity of the simplicial convolution architecture itself. Table 2 incrementally adds L0, L1 and curvature features. However, the L0-only variant still performs graph based message passing. It is not a baseline without simplicial convolution. Consequently, the experiments do not sufficiently distinguish the benefits of higher order simplicial propagation from those of curvature features combined with conventional spatial message passing.

I suggest comparing CASIN with variants that replace the simplicial convolution layers with standard MLP and GNN layers while keeping the same input features, decoder, and training settings. The MLP variant would help assess the overall role of spatial propagation, while a conventional GNN with the same neighborhood information would provide a more direct test of the additional benefit of simplicial convolution.

  1. The temporal modeling design, geometric reconstruction assumptions, and practical relevance require further justification

First, eq 7 uses a residual model to predict the state increment over a fixed time step rather than directly learning a continuous time vector field. This formulation absorbs the finite step evolution into a learned discrete mapping, but the paper does not explain why this design is preferable to learning the instantaneous vector field and using a reliable numerical integrator for state propagation. These approaches may differ in prediction accuracy, computational cost, noise robustness, time step flexibility, and long horizon stability. Demonstrating that the current increment predictor achieves low prediction errors does not establish the necessity of this particular temporal modeling choice.

Second, eq 8 introduces a multi-step trajectory prediction objective. Although the appendix specifies Adam, learning rate scheduling, and gradient clipping, the actual rollout length, gradient computation strategy, and use of truncated backpropagation or other stabilization techniques are not sufficiently described. These details are important for understanding the training cost and numerical stability of the method.

Third, geometric reconstruction depends on the quality and distribution of the observed point cloud. The paper describes local tangent-space estimation, surface fitting, and triangulation, but the practical choices of neighborhood size, sampling density, curvature adaptive sampling, and tolerance to off manifold measurement noise need further clarification.

Finally, the dental application has potential practical relevance, but its actual significance remains insufficiently established. For a tooth surface that has already been scanned, reconstructing the manifold from point clouds mainly transforms existing geometric information into a representation suitable for the proposed network. This is a useful engineering step, but its advantage over directly using an available surface mesh should be clarified. More importantly, the dental experiments combine real tooth geometries with simulated PDE trajectories. The method predicts physical state evolution on a fixed surface rather than changes in tooth geometry caused by dental decay. The current experiments therefore do not establish the ability to predict actual patient-specific disease progression or reconstruct unknown tooth geometry from partial observations. The authors should clarify whether the dental application is intended primarily as a mathematical benchmark on realistic geometries or as a validated model for a specific clinical prediction task.

Questions for authors
  1. Research gap and contribution positioning

Could the authors more precisely define the research gap addressed by CASIN? Rather than broadly emphasizing the limited development of PDE solvers on non-Euclidean domains, the specific challenge appears to involve learning time-dependent PDE dynamics from finite, irregularly distributed observations on manifolds with incomplete geometric information. How does CASIN differ from existing manifold-based PDE solvers and geometrically constrained dynamical learning methods? What are the specific new algorithmic or theoretical contributions?

  1. Independent contribution of the simplicial architecture

CASIN combines several established methodological components. Could the authors clarify whether the primary novelty lies in a new simplicial operator, a new mechanism for integrating curvature and topology, or the combination of existing techniques? In particular, why is the proposed simplicial convolution necessary when the same geometric features and neighborhood information could also be incorporated into a conventional GNN?

  1. Gradient computation and numerical stability in multi-step training

Equation (8) uses a multi-step trajectory prediction loss. Although the appendix specifies Adam as the optimizer, could the authors clarify how gradients are computed through the temporal rollout? For example, is reverse-mode automatic differentiation used to backpropagate through the entire unrolled computation graph? Is truncated backpropagation or another temporal-window strategy employed? What measures are taken to address vanishing or exploding gradients and memory growth during long-horizon training? How sensitive are the results to these choices?

  1. Why use one-step increment prediction instead of continuous-time vector-field learning?

According to Eq. (7), CASIN learns a state increment over a fixed time step. This formulation incorporates the effect of finite-step dynamics into the learned discrete map rather than explicitly approximating the instantaneous time derivative. Why do the authors choose this approach instead of retaining the same geometric encoder and simplicial network to learn a continuous-time vector field and then integrating it using a reliable numerical solver? These two formulations may behave differently under changes in time step, long-horizon prediction, numerical stability, and computational cost. A clearer justification of this design choice would help establish its methodological value.

  1. Additional ablation experiments

I suggest two additional controlled comparisons to isolate the contributions of the proposed architecture.

a. Simplicial convolution ablation: Replace the simplicial convolution layers with standard MLP and conventional GNN layers while keeping the same curvature features, prediction objective, and training protocol. A variant using (L0+curv.) would also help determine whether edge-level and triangle connectivity provide additional benefits once curvature information is already available.

b. Temporal prediction ablation: Compare the current increment-prediction CASIN with a vector-field-prediction CASIN using a numerical integrator. Both variants should use comparable spatial networks, training data, and model capacity. Relevant metrics include prediction accuracy, training and inference time, noise robustness, time-step flexibility, and long-horizon stability.

These comparisons would help clarify the independent benefits of the simplicial architecture and the temporal prediction formulation.

  1. Practical assumptions in point-cloud reconstruction

Could the authors provide more details about the point-cloud reconstruction procedure? How many neighboring points are used to estimate each local tangent space? Are the samples approximately uniform over the manifold, or is their density adapted to local curvature? Do the experiments assume that all sampled points lie exactly on the manifold? If measurement noise perturbs some points away from the true surface, how does this affect geometric reconstruction and subsequent PDE prediction accuracy?

  1. Practical significance of the dental application

The dental scans already provide substantial information about the complete tooth surface. Could the authors clarify what additional practical benefit is obtained by reconstructing the manifold and learning simulated PDE dynamics on it? The current framework appears to predict state evolution on a fixed manifold rather than changes in tooth geometry during disease progression. Is the dental experiment intended mainly as a benchmark using realistic complex geometries, or does the proposed model aim to support prediction of actual dental disease progression? Furthermore, can the framework infer unknown tooth geometry from partial surface observations? If this is outside its current scope, the distinction between geometric reconstruction, physical-state prediction, and evolving-geometry modeling should be made clearer.

14 comments

We thank the reviewer for the careful reading and constructive feedback. We respond to each point below. These changes will also be reflected in an updated draft of our paper.

Comment 1. Research Gap and Contribution Position

A: Thank you for this comment, which has led us to state our contributions more transparently. We do not claim a new mathematical mechanism. Each component has precedent: local surface reconstruction, estimation of geometric quantities, and simplicial convolution. Our contribution is not any single part: the way CASIN's ingredients are coupled lets the network represent dynamics driven by surface geometry, which existing simplicial and operator-learning methods capture less well.

Why geometry-driven dynamics. We test reaction–diffusion and shallow-water systems because surface geometry shapes their solutions in ways the same PDEs do not show on flat domains. In nature, domain geometry decides reaction–diffusion patterns such as animal coat markings, e.g., spots on the body becoming stripes on the narrower tail (Gray–Scott, Turing) [Murray, 2002]. Fisher–KPP produces an invading front, which we use as a phenomenological model of lesion spread on tooth surfaces. In shallow-water flow over uneven terrain, the surface geometry speeds up, slows down or redirects the flow.

Evidence that the coupling matters.

  • Advantage grows with geometric complexity. On the standard torus, whose curvature varies smoothly, CurvGT is competitive and slightly better on Fisher–KPP and Gray–Scott (Table 1). On the more irregular distorted torus and Stanford bunny, CASIN is best in all six settings, with 25–67% lower error. On real teeth (Table 3), CASIN is best on all three systems, by a clear margin on Gray–Scott (0.021 vs. 0.047). For shallow water, CASIN's error is 20–30% lower than CurvGT's on all three manifolds (Table 4).
  • Better dynamics from comparable geometry. CASIN's recovered metric is slightly less accurate than CurvGT's (median relative error 4.9% vs. 4.5%; Appendix C.5), yet its predicted dynamics are more accurate. This suggests the gain comes from how CASIN uses the geometry, not from a better reconstruction.
  • Against a higher-order, topology-aware baseline. We added a new comparison of CASIN with DEC-HOGNN [Liao et al., 2025], suggested by Reviewer zzdN (details below).

New comparison with DEC-HOGNN. DEC-HOGNN is a strong, recent baseline whose inductive bias is itself higher-order and topology-aware. It operates on the DEC operators (Hodge Laplacian L1L_1, boundary maps B1,B2B_1, B_2, lower/upper edge adjacency). To favor it, we use a steady elliptic problem with no time stepping, the problem class it was designed for: the screened-Poisson solution operator (L+κ2M)x=Ms↦x(\mathcal{L}+\kappa^2 M) x = M s \mapsto x with κ2=0.05\kappa^2 = 0.05. Its eigenpairs are determined by the surface geometry, so it tests whether a model has learned the geometry. We consider both cochain degrees: degree 0 (scalar field on nodes, L=L0w\mathcal{L} = L_0^{w}) and degree 1 (1-form on edges, the metric Hodge Laplacian), on the flat, regular and distorted torus. Models are parameter-matched and trained on identical data (160/40/100 train/val/test) with Adam + MSE and early stopping on validation error. We report nRMSE.

ManifoldCochain degreeCASINDEC-HOGNN
Flat torusL0L_00.06900.1593
Flat torusL1L_10.07340.1487
TorusL0L_00.13420.1742
TorusL1L_10.06130.0674
Distorted torusL0L_00.17460.1958
Distorted torusL1L_10.05410.0703

CASIN outperforms DEC-HOGNN in all 6 settings, even in this regime favorable to DEC-HOGNN. These results support our hypothesis that CASIN's curvature-aware design helps it represent the surface geometry that governs geometry-driven dynamics.

A to reviewer comment 1 (continued).

Full ablation (shallow water).

VariantCurvatureTorusDistorted torus
MLP (no propagation)✓0.1742 ± 0.01830.1895 ± 0.0214
GNN✓0.0705 ± 0.01970.1188 ± 0.0201
L0L_0–0.0689 ± 0.0030.0887 ± 0.006
L0L_0 + curvature✓0.0589 ± 0.0030.0567 ± 0.004
L0+L1L_0 + L_1 lower (B1⊤B1B_1^\top B_1)✓0.0555 ± 0.0040.0535 ± 0.004
L0+L1L_0 + L_1 upper (B2B2⊤B_2 B_2^\top)✓0.0547 ± 0.0040.0526 ± 0.004
L0+L1L_0 + L_1✓0.0559 ± 0.0030.0530 ± 0.005
CASIN (full)✓0.0435 ± 0.002*0.0432 ± 0.003**

Normalized RMSE, mean ± std over 5 seeds. Stars on CASIN (full) mark that it is significantly lower than every other variant on that surface (*: p<0.05p < 0.05; **: p<0.01p < 0.01; Holm-corrected).

The ablation highlights the following.

1. Graph-only model with the same curvature features. L0L_0 + curvature is a graph-only model (vertex Laplacian, pairwise messages) with the same curvature inputs as CASIN. Curvature reduces its error by 15% on the torus and 36% on the distorted torus, so curvature features help even without higher-order structure, and more on the irregular surface.

2. Lower vs. upper terms of L1L_1. Adding the lower term (edge-level, B1⊤B1B_1^\top B_1), the upper term (triangle connectivity, B2B2⊤B_2 B_2^\top), or both gives a small further reduction (0.0589 → 0.0547–0.0559 on the torus), about one standard deviation, and the three variants are statistically indistinguishable. Hence, on their own, the L1L_1 terms do not separate edge-level from triangle-level gains.

The ablation also shows how much curvature and simplicial structure each help.

  • Curvature. Adding curvature to the graph Laplacian (L0→L0L_0 \to L_0 + curvature) lowers the error by 15% on the torus (0.0689 → 0.0589) and by 36% on the distorted torus (0.0887 → 0.0567), significant on both surfaces. Its effect is 3.2× larger on the more irregular surface, and it removes the gap between the two: without curvature the distorted torus is 29% harder than the torus; with it the two are within 4%.
  • Simplicial structure. The L1L_1 Hodge terms alone add little. Either the edge-level term B1⊤B1B_1^\top B_1 or the triangle-level term B2B2⊤B_2 B_2^\top lowers the error by a further 6–7%, the two differ by less than 2%, and neither step is significant. The gain comes from the full model, which adds simplex features and vertex–edge message passing through B1B_1 and lowers the error by a further 22% on the torus and 18% on the distorted torus.
  • Relative size. Of the total reduction from L0L_0 to full CASIN (37% on the torus, 51% on the distorted torus), curvature accounts for 39% and simplicial structure for 61% on the torus, and for 70% and 30% on the distorted torus. The simplicial contribution is nearly constant in absolute terms (0.0154 and 0.0135), while the curvature contribution grows with geometric irregularity.

These shares are sequential, with curvature added first, and the final step combines simplex features (edge and triangle) and message passing. All comparisons are paired over 5 seeds with Holm–Bonferroni correction.

We have added this table to the paper and revised our claim accordingly: the advantage of higher-order structure arises from coupling vertex, edge and triangle representations, not from the L1L_1 diffusion terms alone.

Together, these findings suggest that CASIN's advantage grows as the geometry becomes more complex, supporting its use for geometry-dependent dynamics on realistic curved surfaces.

References

  • J. D. Murray. Mathematical Biology I: An Introduction (3rd ed.). Springer, 2002.
  • Liao et al. Boundary-Value PDEs Meet Higher-Order Differential Topology-aware GNNs. NeurIPS 2025.

Comment 2 Using manifold geometry to constrain dynamical learning is not entirely new; e.g., curvature-adaptive geometrically constrained SINDy on manifolds explores geometric constraints and curvature information. These works concern vector fields on state manifolds, whereas CASIN studies PDEs over spatial manifolds. Discussing the relationship and differences would clarify the contribution.

A: We thank the reviewer for pointing to GSINDy on manifolds [He & Li, 2026]. As noted in the reviewer's comment, the two settings differ in what the manifold represents. GSINDy on manifolds identifies ODEs whose state evolves on a manifold. CASIN identifies PDEs posed on a curved surface, where the surface is the spatial domain and its geometry shapes the solution dynamics directly. The manifold in our setting is the PDE's physical domain, not the state space of an ODE.

The two approaches are still related. CASIN's discrete flow map can be written as a high-dimensional ODE system, U˙=L(h,θ;U,K,H)\dot{U} = L(h,\theta ; U, K, H), evolving on a manifold of PDE solutions, and a SINDy-type method could in principle learn from that representation. However, the PDE formulation keeps information that the ODE view loses; the ODE view may not be detailed enough to capture the subtle influence of the surface geometry. That is, CASIN learns the spatial structure of the dynamics through differential operators, and it learns how local geometry (e.g., the curvatures KK and HH) affects local pattern formation. To reiterate, GSINDy uses local curvature to constrain the learned ODE vector field to the tangent space of the state manifold. In CASIN, curvature instead enters the governing surface PDE itself and drives the solution dynamics.

We have added GSINDy [He & Li, 2026] to the Related Work section.

  • H. He and Y. Li. Learning vector fields of differential equations on manifolds with curvature-adaptive geometrically constrained SINDy. Communications in Nonlinear Science and Numerical Simulation, 2025.

Comment 3 Independent contribution of the simplicial architecture and ablation.

A: We agree that an L0L_0-only variant still performs graph message passing, so the original ablation did not isolate the simplicial architecture completely. We have added two baselines that do: an MLP with no propagation, and a conventional GNN (a graph convolutional network) on the same neighbourhood (see the ablation table in our response to Q1, Part 2). We point out the following three observations.

  • Propagation is important. The MLP has an error of 0.1742 (torus) and 0.1895 (distorted torus); full CASIN lowers this by 75% and 77%.
  • CASIN beats curvature features with conventional message passing. The curvature-fed graph Laplacian (L0L_0 + curvature) tests this, and full CASIN is 26% and 24% lower (0.0589 → 0.0435 and 0.0567 → 0.0432). Against the conventional GNN it is 38% and 64% lower.
  • The gain comes from the full model, which adds simplex features (edge and triangle) and vertex–edge message passing through B1B_1, lowering the error by a further 22% and 18%. The benefit of the simplicial architecture therefore lies in coupling information across simplex dimensions, beyond what curvature features and node-level message passing provide.
Reviewer_az9vOpenReview ↗
Rating
2
Confidence
2
Informativeness
3(+1)
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The paper proposes CASIN, a data-driven model for forecasting time-dependent PDE dynamics on unknown surfaces represented by point clouds. It reconstructs a simplicial complex, estimates local Gaussian and mean curvature, and combines these features with node–edge message passing using Hodge Laplacians. The model learns a residual time-stepping map from observed trajectories without requiring the governing equations. Its main contribution is integrating curvature features with simplicial learning for dynamics prediction on irregular curved domains. Validation includes reaction–diffusion systems on synthetic geometries and real dental surfaces, alongside shallow-water experiments, with competitive results against geometry-aware baselines. However, insufficiently specified evaluation protocols and incomplete ablations limit confidence in the model’s generalization and the distinct benefit of its higher-order simplicial structure.

Critical strengths
  1. CASIN combines estimated curvature with node–edge interactions and triangle connectivity to predict dynamics on surfaces represented by point clouds. This provides a plausible inductive bias for irregular curved domains without requiring a supplied mesh or governing equations, making the approach potentially useful for data-driven scientific modeling.
  2. CASIN achieves the lowest reported RMSE in seven of nine synthetic reaction–diffusion settings and all three dental settings, including comparisons with CurvGT and GLNO. Additional shallow-water experiments extend the evaluation beyond reaction–diffusion dynamics. These results provide meaningful empirical evidence that the proposed combination merits investigation.
Critical weaknesses
  1. The paper does not clearly report the number of independent trajectories, their train/test separation, prediction horizons, or whether evaluation measures generalization to unseen initial conditions and geometries versus prediction within previously observed trajectories. Mentioning the dental dataset’s official split does not fully specify the experimental protocol. These omissions make it difficult to determine what the reported errors demonstrate and substantially limit confidence in the claimed forecasting capability.
  2. The cumulative ablation compares curvature, but omits a graph-only model with the same curvature and geometric features. It also does not separate the lower and upper terms of L_1, which would help distinguish gains from edge-level representations from those attributable to triangle connectivity. Consequently, the results do not establish whether higher-order simplicial structure provides a meaningful advantage beyond geometric feature augmentation and additional model capacity, leaving the paper’s main novelty insufficiently substantiated.
Questions for authors
  1. Does CASIN outperform a graph-only variant equipped with the same curvature and geometric features under comparable parameter and training budgets? Can you distinguish the contribution of triangle coupling, B_2B_2^\top, from that of edge-level representations using only B_1^\top B_1? These comparisons would directly affect my assessment of the claimed advantage of higher-order simplicial structure.
  2. What does the manifold represent physically—a fixed reference surface, a bottom surface, or an evolving free surface? Please define d and u^{\perp} in Eq. (13), clarify the mass-flux notation in Eq. (12), and explain how the reference simulation and network represent tangential velocity. These details are necessary to assess the scope and validity of the fluid-dynamics claims.
  3. How many independent trajectories are used for each PDE and geometry, and how are training, validation, and test sets separated? Please specify whether test initial conditions and geometries are unseen, whether predictions are fully autoregressive, and the training and evaluation horizons. For dental experiments, is the reported performance evaluated on held-out patients?
5 comments

We thank the reviewer for the careful reading and constructive feedback. We respond to each point below. We have also appropriately updated the main draft with the changes and will upload it soon.

Q1. The paper does not clearly report the number of independent trajectories, the train/test separation, the prediction horizons, or whether evaluation measures generalization to unseen initial conditions and geometries. Mentioning the dental dataset's official split does not fully specify the protocol.

A: We agree that this is a mishap on our side: some experimental details were missing. We provide the complete code at https://anonymous.4open.science/r/CASIN-DB06/README.md, as well as details on reproducibility. We further add the following clarifications of the experimental protocol.

  • Trajectories. For each PDE and manifold we generate independent trajectories that differ in their initial condition. Training uses 3 trajectories with 3 different initial conditions. Evaluation uses disjoint trajectories that the model never saw during training.
  • Prediction horizon. Each trajectory has T=40T = 40 recorded steps. Training is performed for a fixed horizon step and the model never sees beyond this step during the test. Predictions, at train and test time, are fully from the initial horizon time frame only.
  • Replicates. Each run is repeated over 5 seeds, and at every seed we resample the point cloud, the simplicial complex, and the initial conditions to make sure we have independent replicates. Significance is assessed with paired tests. Reproducibility details are in Appendix D.

Q2. The cumulative ablation omits a graph-only model with the same curvature and geometric features, and does not separate the lower and upper terms of L1L_1. It therefore does not establish whether higher-order simplicial structure helps beyond geometric feature augmentation and additional capacity.

A: We ran a more detailed ablation on the shallow-water system. All eight variants share the same decoder, optimizer, epochs and data, and parameter counts are matched. They differ in how information propagates and, where noted, in their geometric inputs (the Curvature column, and the edge and triangle features of the full model).

VariantCurvatureTorusDistorted torus
MLP (no propagation)✓0.1742 ± 0.01830.1895 ± 0.0214
GNN✓0.0705 ± 0.01970.1188 ± 0.0201
L0L_0–0.0689 ± 0.0030.0887 ± 0.006
L0L_0 + curvature✓0.0589 ± 0.0030.0567 ± 0.004
L0+L1L_0 + L_1 lower (B1⊤B1B_1^\top B_1)✓0.0555 ± 0.0040.0535 ± 0.004
L0+L1L_0 + L_1 upper (B2B2⊤B_2 B_2^\top)✓0.0547 ± 0.0040.0526 ± 0.004
L0+L1L_0 + L_1✓0.0559 ± 0.0030.0530 ± 0.005
CASIN (full)✓0.0435 ± 0.002*0.0432 ± 0.003**

Normalized RMSE on held-out trajectories with unseen initial conditions, mean ± std over 5 runs. Stars mark that CASIN (full) is significantly lower than every other variant on that surface (*: p<0.05p < 0.05; **: p<0.01p < 0.01; Holm-corrected).

We highlight the following observations.

Graph-only model with the same curvature features. L0L_0 + curvature is a graph-only model (vertex Laplacian, pairwise messages) with the same curvature inputs as CASIN. Curvature reduces its error by 15% on the torus and 36% on the distorted torus, so curvature features help even without higher-order structure, and more on the irregular surface. The GNN row is a standard GCN on the same graph with the same curvature inputs. Some of its rollouts diverge, which inflates its spread, so we use L0L_0 + curvature as the graph-only reference.

Lower vs. upper terms of L1L_1. Adding the lower term (edge-level, B1⊤B1B_1^\top B_1), the upper term (triangle connectivity, B2B2⊤B_2 B_2^\top), or both gives a small further reduction (0.0589 → 0.0547–0.0559 on the torus), about one standard deviation, and the three variants are statistically indistinguishable. Hence, on their own, the L1L_1 terms do not separate edge-level from triangle-level gains.

Benefit beyond feature augmentation and capacity. The largest gain beyond curvature comes from the full model. It adds edge and triangle geometric features (Eqs. 4–5) and passes messages between vertices and edges through B1B_1. This lowers the error from 0.0559 to 0.0435 (−22%) on the torus and from 0.0530 to 0.0432 (−18%) on the distorted torus, a difference of 2–4 standard deviations. Both variants have matched capacity and the same curvature features, so the gain comes from coupling geometric information across simplex dimensions, not from curvature augmentation or model size.

We have added this table to the paper and revised our claim accordingly: the advantage of higher-order structure arises from coupling vertex, edge and triangle representations, not from the L1L_1 diffusion terms alone.

Q3. Does CASIN outperform a graph-only variant equipped with the same curvature and geometric features under comparable parameter and training budgets? Can you distinguish the contribution of triangle coupling (B2B2⊤B_2 B_2^\top) from that of edge-level representations using only B1⊤B1B_1^\top B_1?

A: Yes. Under matched parameter and training budgets, CASIN outperforms the graph-only variants on held-out initial conditions (table in our response to Q2, Part 1). Relative to the curvature-fed graph Laplacian (L0L_0 + curvature), CASIN reduces the error by 26% on the torus (0.0589 → 0.0435) and 24% on the distorted torus (0.0567 → 0.0432); relative to the conventional GNN, by 38% and 64%.

On the second question, we ran the edge-level term B1⊤B1B_1^\top B_1 and the triangle-coupling term B2B2⊤B_2 B_2^\top as separate variants. Each improves on L0L_0 + curvature by about 6–7% (torus: 0.0555 and 0.0547; distorted torus: 0.0535 and 0.0526), but the two differ by less than 2%, and combining them gives no further gain (0.0559 and 0.0530). Hence, the L1L_1 terms alone do not separate edge-level from triangle-level contributions. The main advantage comes from the full model, which adds edge and triangle geometric features (Eqs. 4–5) and vertex–edge message passing through B1B_1: it lowers the error by a further 18–22% over L0+L1L_0 + L_1. The benefit of higher-order structure thus lies in coupling geometric information across simplex dimensions, not in either L1L_1 term in isolation.

Q4. What does the manifold represent physically: a fixed reference surface, a bottom surface, or an evolving free surface? Please define dd and u⊥u^{\perp} in Eq. (13), clarify the mass-flux notation in Eq. (12), and explain how the reference simulation and the network represent tangential velocity.

A: The manifold is a fixed reference surface on which the shallow-water state evolves, i.e., it is neither the bottom surface nor the free surface. The shallow-water system lives on the surface manifold (for example, thin water flow on the surface of the Earth). The state is (h,uθ,uψ)(h, u^\theta, u^\psi), and the simulation integrates the system with an RK4 scheme.

u⊥=n×uu^{\perp} = n \times u is the cross product of the normal vector nn and the flow velocity vector uu. It is the Coriolis and vorticity force term due to, for example, the Earth's rotation. We simplify it by setting the Coriolis constant to 1.

The mass-flux equation, Eq. (12), should be corrected to

∂th+∇M(hu)=0.\partial_t h + \nabla_M (hu) = 0.

Thank you for the comment on that. The momentum equation, Eq. (13), for the flow velocity vector reads

∂tu+(u⋅∇M)u+fn×u+g∇Md=0,\partial_t u + (u \cdot \nabla_M) u + f n \times u + g \nabla_M d = 0,

where d=h+bd = h + b is the height of the free surface and bb is the bottom topography. Here we set f=1f = 1.

Experiment Details How many independent trajectories are used for each PDE and geometry, and how are the training, validation, and test sets separated? Are test initial conditions and geometries unseen? Are predictions fully autoregressive, and what are the training and evaluation horizons? For the dental experiments, is performance evaluated on held-out patients?

A: Thank you for asking these clarifying questions.

  • Trajectories. For each (PDE, geometry) pair we generate independent trajectories that differ in their initial condition. Training uses 3 trajectories (distinct ICs) on a single sampled geometry; evaluation uses disjoint trajectories that the model never saw during training.
  • Horizons and autoregression. Each trajectory has T=40T = 40 recorded steps. Training supervises an autoregressive rollout up to a fixed horizon Htrain=24H_{\text{train}} = 24; the model never sees frames beyond HtrainH_{\text{train}} during training. All predictions, at train and test time, are fully autoregressive from the initial frame only.
  • Geometries. For the synthetic manifolds the surface is held fixed and each seed draws a new point-cloud sample of it, so the sampling is unseen while the surface itself is not, and the PDE parameters are identical between training and test. On dental, the unseen geometries are the held-out dental patients.
  • Robustness and time extrapolation. Robustness to coarser or denser point clouds and to off-manifold perturbation is assessed separately in a zero-shot setting, and time extrapolation is measured by rolling the unseen-initial-condition trajectories past the training horizon HtrainH_{\text{train}}.
  • Replicates. Each run is repeated over 5 seeds, and every seed resamples the point cloud, the simplicial complex, and the initial conditions, so the seeds are independent replicates, not merely re-initializations of the same data. Significance is assessed with paired tests over aligned seeds.
Reviewer_zzdNOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CASIN, a curvature-aware simplicial neural network for learning time-dependent PDE dynamics on unknown Riemannian manifolds represented by irregularly sampled point clouds. The method reconstructs a simplicial complex from observations, estimates local geometric quantities including Gaussian and mean curvature, and incorporates them into node representations. It further employs Hodge-Laplacian-based simplicial convolutions to capture interactions across different topological dimensions, followed by a residual flow-map decoder for temporal prediction. The authors evaluate CASIN on several reaction–diffusion systems and shallow-water dynamics over synthetic manifolds and dental surfaces. Experimental results demonstrate competitive predictive accuracy against eight neural PDE solvers, with CASIN achieving the lowest RMSE in seven of nine synthetic benchmark settings.

Critical strengths
  1. Practical relevance. The paper addresses PDE learning on irregularly sampled curved surfaces, a meaningful problem with potential applications in scientific computing and biomedical modeling. The dental surface experiments provide a concrete application scenario beyond standard synthetic geometries.
  2. Empirical performance. CASIN achieves favorable prediction accuracy across multiple nonlinear reaction–diffusion systems and geometrically diverse manifolds, particularly on distorted surfaces. The comparison with eight baselines provides useful empirical evidence of its effectiveness.
  3. Component-wise evaluation. The ablation studies separately examine the contributions of higher-order simplicial interactions and curvature features, helping establish the empirical benefits of the combined architecture though already verified in earlier works.
Critical weaknesses
  1. The writing and figure presentation needs considerable refinement. The typos such as M=xj1NM = {x_j}_{1}^N in line 198 and inconsistent symbol usage (e.g. definitions of M={x_i}1^N and P={x_i}{j=1}^N in line 198 and 199) , repetitive label "C" in line 884; it is curvgt not curvegt in figure 2. The improper presentation of equations: unreasonable splitting. e.g. eq. (1). The abstract can be more concise.

  2. The proposed methodology appears to be trivial composition of already developed methods. It seems to merely leverage conventional surface reconstruction methods to fetch vertex features on a undetermined manifold. Beyond surface reconstruction, recent works like [1] use regression method to estimate the geometric quantities of manifolds like metric and curvature tensor. Therefore it lacks enough novel contribution on offering new node feature. Also, simplical convolution has been discussed for long in higher-order graph learning, and learning on simplices has been widely studied in ai4pde works like [2,3] where either relations between higher-order topology and its underlying physics is studied, or a more comprehensive learning framework is built up. However, turning back to this paper, it does not give informative insights beyond the already scope since combining representations of different topological levels has been sophisticated.

Most critically, though claimed learning on an uncertain manifold, all methods appears to require obtaining a certain mesh or point cloud via surface reconstruction methods. It does not offer any insights on dealing with such uncertainty but opt for a hand-made rule to interpret such uncertainty. In this sense, it cannot be told apart from works on determined geometry.

Ref. [1] Curvature Estimation on Data Manifolds via Diffusion-augmented Sampling. Jason Wang et al. NeurIPS 2025

[2] Boundary-Value PDEs Meet Higher-Order Differential Topology-aware GNNs. Liao et al. NeurIPS 2025.

[3] Topological Neural Operators. Lennart Bastian et al. arxiv.2606.09806

Questions for authors

I suggest the authors to further refine the presentation and justify their contributions.

5 comments

We thank the reviewer for the careful reading and constructive feedback. We respond to each point below. We have also corrected the inconsistencies in the main draft and will update a revised version soon.

Q. The methodology appears to be a trivial composition of existing methods (surface reconstruction to obtain vertex features; simplicial convolution). Recent works estimate geometric quantities by regression [1], and learning on simplices is well studied in AI4PDE [2, 3]. The paper gives limited insight beyond this scope.

A: We thank the reviewer for this question and for the references! Addressing it helped us crystallize CASIN's novelty. We agree that each component has precedent: local surface reconstruction, estimation of geometric quantities, and simplicial convolution. However, our contribution is not any single part, and we respectfully disagree that the method is a trivial composition. The way CASIN's ingredients are coupled lets the network represent dynamics driven by surface geometry, which existing simplicial and operator-learning methods capture less well.

1. Setting. We study time-dependent PDEs posed intrinsically on a Riemannian manifold represented only by an irregularly sampled point cloud. [2] and [3] learn on a given mesh or cell complex and are evaluated on steady-state problems; [1] estimates curvature to characterize a data manifold, with no dynamics. CASIN is agnostic to the curvature estimator (as Table 6 in the main body shows, accuracy changes little across four estimators), so regression-based estimators such as [1] can be used directly.

2. Most importantly: dynamics driven by reactions and geometry. [2, 3] organize learning around discrete differential operators, which suits transport- and diffusion-dominated problems (Poisson, Darcy, airfoil, elasticity). However, many biological systems, including our benchmarks, are driven by nonlinear reaction terms that act pointwise on the state, and the surface geometry can decide whether, where and how fast structures form or spread [4, 5].

3. Comparison with the reviewer's suggested DEC-HOGNN [2]. DEC-HOGNN is a strong, recent baseline whose inductive bias is itself higher-order and topology-aware. It operates on the DEC operators (Hodge Laplacian L1L_1, boundary maps B1,B2B_1, B_2, lower/upper edge adjacency). To favor it, we use a steady elliptic problem with no time stepping, the problem class it was designed for: the screened-Poisson solution operator (L+κ2M)x=Ms↦x(\mathcal{L}+\kappa^2 M) x = M s \mapsto x with κ2=0.05\kappa^2 = 0.05. Its eigenpairs are determined by the surface geometry, so it tests whether a model has learned the geometry. We consider both cochain degrees: degree 0 (scalar field on nodes, L=L0w\mathcal{L} = L_0^{w}) and degree 1 (1-form on edges, the metric Hodge Laplacian), on the flat, regular and distorted torus. Models are parameter-matched and trained on identical data (160/40/100 train/val/test) with Adam + MSE and early stopping on validation error. We report per-sample relative L2L^2 error on the test set, mean ± std.

ManifoldCochain degreeCASINDEC-HOGNN
Flat torusL0L_00.06900.1593
Flat torusL1L_10.07340.1487
TorusL0L_00.13420.1742
TorusL1L_10.06130.0674
Distorted torusL0L_00.17460.1958
Distorted torusL1L_10.05410.0703

CASIN outperforms DEC-HOGNN in all 6 settings, even in this regime favorable to DEC-HOGNN. These results support our hypothesis that CASIN's curvature-aware design helps it represent the surface geometry that governs geometry-driven dynamics.

4. How geometry controls the dynamics.

(a) Pattern selection (Turing; Gray–Scott around its nontrivial steady state). Linearizing the kinetics gives a dispersion relation that fixes a band of unstable Laplace–Beltrami eigenvalues, the same on every surface. The surface's intrinsic geometry decides which eigenvalues are available and the shapes of the eigenfunctions, i.e., size decides whether a pattern forms (eigenvalues scale like 1/L21/L^2), curvature can influence where it forms first [5], and shape can influence whether spots or stripes form [4]. A small eigenvalue error can create, remove or move a pattern, and the error comes from the geometry representation, not the reaction terms. On teeth, cusp number and position arise during development from reaction–diffusion signaling on curved, growing tissue [7, 8] (development, not caries).

(b) Front propagation (Fisher–KPP). ut=DΔMu+ru(1−u)u_t = D\Delta_M u + r u(1-u) produces an invading front, which we use as a phenomenological model of lesion spread. On a flat domain it moves at c0=2Drc_0 = 2\sqrt{Dr}; on a surface its normal velocity is approximately V≈c0−DκgV \approx c_0 - D\kappa_g [6], with κg≈1/ρ−Kρ/3\kappa_g \approx 1/\rho - K\rho/3 for an expanding front of geodesic radius ρ\rho. Fronts speed up across regions of positive Gaussian curvature (cusps) and slow across saddle-shaped regions (fissures). Predicting lesion spread therefore requires accurate intrinsic geometry (metric and Gaussian curvature), even when the reaction term is known. CASIN supplies KK pointwise as a node feature, which an unweighted graph operator cannot represent; this is consistent with the curvature ablation (Table 2).

5. Findings. On the tooth, CASIN has the lowest error on all three systems vs. the best competitor CurvGT (see Table 3 in the main body), especially in Gray–Scott, where it more than halves the error (0.021 vs. 0.047). These results are consistent with CASIN's advantage coming from its representation of geometry rather than from fitting a particular reaction term. We do not claim these models are mechanistic descriptions of tooth development or caries; we use them as geometry-sensitive surface dynamics on real anatomy, matching our target setting of data whose governing mechanisms are only partially known.

References

  • [1] J. Wang et al. Curvature Estimation on Data Manifolds via Diffusion-augmented Sampling. NeurIPS 2025.
  • [2] Liao et al. Boundary-Value PDEs Meet Higher-Order Differential Topology-aware GNNs. NeurIPS 2025.
  • [3] L. Bastian et al. Topological Neural Operators. arXiv:2606.09806.
  • [4] R. Nishide and S. Ishihara. Weakly nonlinear analysis of Turing pattern dynamics on curved surfaces. Phys. Rev. E 111(2), 2025.
  • [5] G. Vandin, D. Marenduzzo, A. B. Goryachev, and E. Orlandini. Curvature-driven positioning of Turing patterns in phase-separating curved membranes. Soft Matter 12(17):3888–3896, 2016.
  • [6] J. J. Tyson and J. P. Keener. Singular perturbation theory of traveling waves in excitable media (a review). Physica D 32:327–361, 1988.
  • [7] J. Jernvall and I. Thesleff. Reiterative signaling and patterning during mammalian tooth morphogenesis. Mech. Dev. 92:19–29, 2000.
  • [8] I. Salazar-Ciudad and J. Jernvall. A computational model of teeth and the developmental origins of morphological variation. Nature 464:583–586, 2010.

Q Although the paper claims learning on an uncertain manifold, all methods appear to require a mesh or point cloud obtained via surface reconstruction. The paper offers no insight into dealing with this uncertainty and relies on a hand-made rule, so it cannot be told apart from works on determined geometry.

A: We thank the reviewer and agree that our wording was not sufficiently clear. CASIN does not quantify or model uncertainty in the manifold. By "unknown" manifold we mean a manifold that is not given analytically, as a parametrization, or as a mesh with known connectivity: only an irregular point sample is available, and its geometry must be recovered from that sample. We will state this definition explicitly in the introduction.

In this setting, CASIN differs from methods on a determined geometry as follows.

1. Trained on estimated geometry, errors included. Methods on a determined geometry take a mesh and its discrete operators (e.g., the cotangent Laplacian) as exact. CASIN starts from the raw point cloud: the simplicial complex, normals and curvatures are estimated from the samples. The network is trained and evaluated on these estimates, errors included, and across independent reconstructions, since every seed resamples the point cloud and rebuilds the complex. On the tooth, every prediction depends on how faithfully the geometry was recovered from an irregular sample.

2. Metric information enters as features, not operator weights. The geometry fixes the weights of the discrete operator, so a reconstruction error directly biases it and shifts the Laplace–Beltrami spectrum. As discussed in our response to Q2, such a shift can create, remove or move a pattern, or change a front's speed. CASIN's operators encode only connectivity, while curvature and simplex geometry enter as input features that the network learns to weight.

3. The estimator is principled. Curvatures come from local polynomial fitting, whose consistency and convergence as sampling density increases are established [Cazals & Pouget, 2005; Meek & Walton, 2000]. They are computed once, in ambient coordinates, without an intrinsic parametrization. The neighborhood size (k=24k = 24) is the main tuned parameter (we report its effect on curvature recovery in Appendix C).

4. Evidence of robustness to reconstruction error.

  • (a) Recovery accuracy: CASIN's recovered metric has a median relative error of 4.9%, comparable to CurvGT (4.5%) and well below Isomap (14.2%) and diffusion maps (42.6%), and it recovers the correct topology (see Appendix C).
  • (b) Sensitivity to geometric error: perturbing points off the manifold degrades the reconstructed curvature, yet RMSE stays competitive. Accuracy also varies little across curvature estimators whose recovery error differs by an order of magnitude (Table 6).
  • (c) Downstream dynamics: with geometric inputs of accuracy comparable to CurvGT, CASIN has lower error on the tooth for Gray–Scott (0.021 vs. 0.047) and Turing (0.011 vs. 0.013); on Fisher–KPP the two are within error (0.017 vs. 0.018).

5. Scope. Explicitly modeling uncertainty in the reconstruction is beyond this work, but it is a natural extension: for example, training on multiple plausible reconstructions (resampling, or diffusion-based augmentation as in Wang et al., 2025), or supplying a per-node confidence in each geometric quantity as an additional input feature. We will add this to the discussion.

References

  • J. Wang et al. Curvature Estimation on Data Manifolds via Diffusion-augmented Sampling. NeurIPS 2025.
  • F. Cazals and M. Pouget. Estimating differential quantities using polynomial fitting of osculating jets. Computer Aided Geometric Design 22(2):121–146, 2005.
  • D. S. Meek and D. J. Walton. On surface normal and Gaussian curvature approximations given data sampled from a smooth surface. Computer Aided Geometric Design 17(6):521–543, 2000.

Comments on the paper

1 comment

Dear AC and Reviewers,

Thank you very much for the detailed comments, motivating discussion and appreciating the CASIN potential! Below we summarize the main points of our discussion:

1.⁠ What is the key contribution of CASIN and how does it differ from others? (zzdN, bkUg). The key novelty of CASIN is not in its individual components but in its coupling of curvature and per-simplex geometry with simplicial message passing, which allows CASIN to capture nonlinear dynamics driven by reactions and geometry of the unknown Riemannian manifold represented only by an irregularly sampled point cloud. As such, CASIN differs both in its focus and experimental settings from other existing methods and is also agnostic to the to the curvature estimator. We highlighted it with additional experiments and comparison with much stronger baselines suggested by the reviewers (DEC-HOGNN, Appendix C.7, Table 9) and (GSINDy Appendix B).

2.⁠ ⁠Simplicial structure vs. geometric features (az9v, bkUg). We added a parameter-matched ablation on shallow water with an MLP, a GNN, graph-Laplacian variants with and without curvature, separate lower and upper L1L_1 terms and GNN with geometric and CASIN without geometric (Appendix C.6, Table 7).

3.⁠ Increment versus continuous-time prediction, and stability (bkUg). We implemented a vector-field variant on the same backbone with RK4. The increment model is more accurate on the test horizon at about a quarter of the cost, and the two are comparable on the long horizon and revised the scope of stability claims. (Appendix C.7, Table 8 , Appendix D; new limitation in Section 8).

4.⁠ Experimental protocol (az9v, bkUg). We now state the number of trajectories, the train/test separation, the training and evaluation horizons, the fully autoregressive rollout on unseen initial conditions, evaluation on held-out dental patients, and the 5 resampled seeds with paired tests. We also describe the training routine, full backpropagation through the rollout, horizon curriculum, gradient clipping and a divergence safeguard (Appendix D.1, Algorithm 1).

5.⁠ ⁠Meaning of "unknown manifold" and robustness of the reconstruction (zzdN, bkUg). "Unknown" means the surface is available only as an irregular point sample. We do not model reconstruction uncertainty. We added the reconstruction details and a zero-shot robustness study covering off-manifold noise, sparse initial observations and density transfer, alongside manifold-recovery accuracy and sensitivity to the curvature estimator (Appendix C.5–C.7).

6.⁠ ⁠Scope of the dental application (bkUg). The dental experiments use simulated reaction–diffusion trajectories on real tooth geometries, we have clarified and revised the text to reflect this clearly. Section 6.2 is retitled accordingly, Appendix C.4 states the scope and future clinical work, and Table 10 separates demonstrated capabilities from future applications. Importantly, CASIN is one of the very few studies of physics-informal ML for dentistry.

7.⁠ ⁠Shallow-water formulation and presentation (az9v, zzdN). We clarified what the manifold represents, defined dd and u⊥u^{\perp}, and corrected Eq. (12) (Appendix C.1). We also fixed the typos and notation issues noted.

Reviewers zzdN and bkUg have raised their total ratings from 2 to 3. Reviewer az9v increased the presentation and informativeness scores (we have addressed all Reviewer az9v’s questions: the number of independent trajectories, ablation with a graph-only variant and triangle coupling, and the manifold representation).

Regardless of the outcome for our paper, we are very lucky to have interactive discussion with all 3 reviewers and are very grateful for this experience and all the constructive feedback.

The Authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.