- 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
- 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.
- 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
- 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.
- 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
- 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?
- 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?
- 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?
- 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.
- 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.
- 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?
- 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
A to reviewer comment 1 (continued).
Full ablation (shallow water).
| Variant | Curvature | Torus | Distorted torus |
|---|---|---|---|
| MLP (no propagation) | ✓ | 0.1742 ± 0.0183 | 0.1895 ± 0.0214 |
| GNN | ✓ | 0.0705 ± 0.0197 | 0.1188 ± 0.0201 |
| – | 0.0689 ± 0.003 | 0.0887 ± 0.006 | |
| + curvature | ✓ | 0.0589 ± 0.003 | 0.0567 ± 0.004 |
| lower () | ✓ | 0.0555 ± 0.004 | 0.0535 ± 0.004 |
| upper () | ✓ | 0.0547 ± 0.004 | 0.0526 ± 0.004 |
| ✓ | 0.0559 ± 0.003 | 0.0530 ± 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 (*: ; **: ; Holm-corrected).
The ablation highlights the following.
1. Graph-only model with the same curvature features. + 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 . Adding the lower term (edge-level, ), the upper term (triangle connectivity, ), 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 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 ( + 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 Hodge terms alone add little. Either the edge-level term or the triangle-level term 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 and lowers the error by a further 22% on the torus and 18% on the distorted torus.
- Relative size. Of the total reduction from 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 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, , 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 and ) 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 -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 ( + 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 , 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.
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.
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.
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 L1, boundary maps B1,B2, 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 with κ2=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) 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.
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.