- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes a hybrid data-driven approach to solve PDE problems that combines finite-volume computations and flux updates with a GNN backbone that provides intermediate features, from which the states at the interacting surfaces and, subsequently, the final updates are computed. By following conservative flux updates, the model is consistent with the underlying equations, provided the state values at the surfaces are consistent.
Results indicate the effectiveness of the model against some classical baselines for PDE problems. Additionally, the authors experiment with pretraining a model and fine-tuning it on some higher-resolution trajectories, achieving great results when testing on such finer meshes.
Critical strengths
- Although not completely new, the main idea of the paper is sound, as it is based on a well-established classical methodology for solving PDEs combined with neural methods for intermediate calculations. How to calculate these intermediate quantities is often a difficult design choice and strongly influences the final results. Hence, the problem is relevant.
- The results on fine-tuning to finer meshes are very interesting and good.
Critical weaknesses
There are several critical issues in this paper:
- The methodology is not completely new and, given the currently writing of the paper, it is difficult to really isolate the different to existing methods that combine finite-volume with GNNs or other neural networks. For example, there are [1,2]. The authors should provide more comparisons with these methods, both from a methodological and experimental point of view. Especially, I am interested in the comparison with [1], which does not seem substantially different but has way worse results.
- Furthermore, the methods that the authors compare to FV-GNO are very few, with many different methods being developed after the introduction of the FNO and U-Net. For example, GINO, which the authors cite, is a more comparable baseline for irregular meshes. Since the model loss is also physics-inspired, other such methods informed by the equations (for example, [3]) should be compared. Furthermore, MeshGraphNetworks and their derivatives are also cited but not really evaluated, and have shown great performance lately. For example, see [4].
- It is not clear what is the method used to calculate the reference solution, as it simply referred to as DG. Furthermore, I am not convinced by the time comparison with this method, as they report "no GPU time". It is still interesting to see the time comparison with the reference solution, in any way it was calculated.
- Almost every section is too succint and feels like a summary of what it should contain. The authors sometimes provide very specific details on not-so-relevant choices, and then almost no detail on the main architecture. For example, what GNN backbone is used? What is in Eq. 2? (Almost no detail on the latter equations is given. What does it mean that "Additional irregular-mesh neighbors provide context but do not add physical exchange faces."? The terminology is also not common for a paper in the field of AI. The "efficient graph execution" paragraph is very difficult to get the head around. Overall, the paper is very difficult to understand with no detail.
- I am not convinced by the claim related to "The reference solver may take multiple internal steps between two observations, whereas one learned update spans their full interval " (lines 239-240). Looking at Section 3.2, the employed GNN also stacks multiple layers to calculate an update.
- The theoretical analysis is very hard to follow: short sentences with no interpretation, every detail is left to the appendix, the main propositions introduce a lot of terms without further explanations, and they have no clear interpretation of the methods.
Overall, the paper is really hard to follow and does not provide much information as it is.
[1] Masanobu and Mitsune, "Graph Neural PDE Solvers with Conservation and Similarity-Equivariance", ICML 2024.
[2] Li et al., "Predicting unsteady incompressible fluid dynamics with finite volume informed neural network", 2024.
[3] Li et al., "Physics-Informed Neural Operator for Learning Partial Differential Equations", 2021
[4] Hoang et al., "Improving Long-Range Interactions in Graph Neural Simulators via Hamiltonian Dynamics", ICLR 2026.
Questions for authors
See the weaknesses above. Furthermore:
- In the caption of Table 1: what does it mean that "five-state models score after context"? What are five-state models?
- Please add the computational times of the reference solution, and comment on possible advantages of the model.
- What is a direct nodal predictor (line 439)? Is it a simple GNN or something closer to a MeshGraphNet?
8 comments
I wanted to offer a somewhat different read of this paper, in case it’s useful for the discussion. I agree the original submission was dense, but I think the revised PDF may address several of the concerns in your review.
GINO, MeshGraphNet, and Kochkov’s learned interpolation have been added (Tables 1–2). On the regular grids, FV-GNO’s margins over all baselines are well beyond the seed-to-seed variation (Table 7). On FLUXGNN specifically, the gap is large (roughly 60–80% lower error on the homogeneous tasks), and FLUXGNN was unstable on topographic shallow water. The revision also names the reference solver (Trixi DG, §5.1), reports DG2 CPU timings (Table 3), and spells out the architecture and the update operator (§3.2–3.3).
I don’t think the paper is without issues. The Euler margin over GNOT is modest, and as Reviewer YaA9 noted, the learned-flux and MPNN ablations perform comparably. Presentation could still be improved. To me, though, these feel like questions about scope and writing rather than reasons for rejection. Since we both see the core idea as sound, I lean toward thinking the contribution clears the bar.
I’d value your perspective on whether the revision changes your view, and I’m happy to be corrected if I’ve missed something.
Thank you for the authors' responses and the new additions to the manuscript, which improve on the previous version. In particular, the methodological description is now clearer, and the additional experimental comparisons are appreciated.
- The comparison with FluxGNN is now clearer, although I am concerned about the results obtained. Given the fact that FV-GNO decodes the physical states from the graph features, and there are 10 layers of GNN, is the comparison fair? What is the computational burden compared to FluxGNN? Are the two methods compared fairly? I am also surprised by the results, especially those on the original task from the FluxGNN paper, where FV-GNO performs more than 1000x better. Please provide the setups and a comparison of the number of parameters and inference cost for both the FluxGNN task and the ones from FV-GNO. If they are very different, it would be beneficial to provide a parameter-matched comparison. This is not a blocking issue, but it is important to understand whether the models were evaluated equally or whether the difference in performance is solely due to the better architecture.
- On a related note, the newly added baselines do not have a description of the training budget and their setups. Given the difficulty of these tasks, and the possible complexity and dimensionality of the models, this is an important comparison to make. The revised manuscript should also include their setups for these tasks to understand if it is correct and where the gap lies.
- The comparison with FVGN is very short, and the authors do not provide further results on this methodology. I would have liked an experimental comparison with this method as well, but I understand now that time is limited, so this is not necessary now and can be skipped. However, the similarity between the models was dismissed with a very short explanation, which is not satisfactory to me. Both methods predict interface-related quantities from graph features and use them in a finite-volume-inspired numerical update. The authors should explain why decoding two physical interface states instead of velocity, pressure, and flux compensation constitutes a substantive methodological advance.
There is also an important question that has not been solved from the initial W1: the authors claim that a key contribution is the fact that predicting the two-side interface states and combining them in a prescribed flux provides an advantage over the fully learned one. However, Ablations in Table 14 show that, in general, this advantage is not clear and the two possibilities are comparable. This is also particularly important for novelty and the comparison with previous methods and to understand where the major contribution lies (either in the learned two-side interface state, the prescribed flux, or the finite-volume style update).
On a final note, could the authors highlight the changes in the manuscript in a different color? Given the paper's density and the substantial new information, I would like to clearly see how the paper has changed and which parts have been adapted to provide a complete evaluation.
Methodological Distinctions, Additional Experiments, and Presentation Improvements
Thank you for the suggestions. We removed less essential details and clarified the method and theory. W1-W6 follow your six weaknesses; Q1-Q3 denote your additional questions.
W1. Differences from FLUXGNN and FVGN. FLUXGNN evolves conservative latent states with linear encoding and left-inverse decoding (DE=I); our task adaptation retains this structure. FV-GNO decodes effective left/right physical states from nonlinear graph features, then applies prescribed fluxes, sources, coefficients, and geometry. Observed transitions train these states over the saved interval. Regular-grid balance is imposed by shared physical fluxes in the final assembly.
Sharp-front capture motivates the two-sided design. Rusanov uses Φ=2Fn(U−)+Fn(U+)−2α(U+−U−), where Fn=F⋅n and α estimates wave speed. Separate states retain jumps for transport and dissipation; shared fluxes enter neighboring updates with opposite signs. Sharp-region results:
Topographic SWE requires pressure-flux/bed-source balance; small depths can amplify velocity errors. This may help explain FLUXGNN's large errors on this task.
On the FLUXGNN authors' released 1D convection-diffusion data, we use their training setup, replacing only the model. FV-GNO achieves MSE 1.578×10−8. The original FLUXGNN paper reports test MSE 3.080×10−5.
FVGN learns face velocity, pressure, and flux compensation with spatial integration and face supervision. We retain prescribed flux/source calculations with state-only supervision.
W2. Additional baselines and loss. We train with normalized state MSE, with physics built into the forward update. We added GINO and an independent MeshGraphNet (full-rollout Macro-RMSE):
All use original splits, one input frame, the same boundary data, and full horizons. The revised Related Work also clarifies our relation to PINO and IGNS.
W3 / Q2. Reference solver and cost. Reference trajectories use Trixi shock-capturing DG: degree four for Burgers, three otherwise (Appendix B.1). DG2 is a separate degree-two comparator. Its implementation lacks a GPU backend, explaining "no GPU time." DG2 uses an eight-thread Xeon Gold 6459C/Float64; FV-GNO uses RTX 5090/FP32. The table reports full-test Macro-RMSE and warmed rollout times:
Under matched GPU settings, FV-GNO is more accurate and 2.1-9.7 times faster than the tested FV2 schemes.
W4 / Q3. Architecture, assembly, and nodal controls. The GNO processor has ten width-120 blocks: ziℓ+1=σ(LN[Wℓziℓ+∣Ni∣−1∑j∈NiKℓ(ξijℓ)⊙Hℓzjℓ]). σ is GELU; LN is layer normalization; Wℓ,Hℓ are linear maps; Ni is the context neighborhood. Kernels use endpoint features, their differences, and geometry. The decoder outputs (qf−,qf+); Eq. (2)'s P is Pi=−ViΔt∑f∈∂iAfsifΦf+ΔtSi. Af,Vi,sif are face measure, volume, and orientation; Φf,Si are flux and source. Context edges pass features; physical faces exchange flux. Execution optimizations cache geometry and combine equivalent operations while preserving the update.
Nodal controls use our processor to predict increments; projection enforces global conservation during training. MeshGraphNet is evaluated separately as an external baseline.
W5. Layers versus physical time steps. Layers transform features at one physical time; assembly advances by Δt. Solver substeps advance physical states through intermediate times. Runtime includes all network layers.
W6. Interpretation of theory. Paired assembly preserves local balance for any reconstruction. Proposition 4.1 shows a representation benefit of known nonlinear computation over a specified capacity-limited direct class. Under closed-domain assumptions, Eq. (6) gives an error floor when an invariant excludes the target. Together, these results motivate retaining known physics and choosing constraints compatible with observations.
Q1. Five-state inputs. Five-state models use frames 0-4 as context, scoring 75 subsequent frames (46 for topography). Single-frame models score 79 (50).
Thank you. We would be happy to discuss any further questions.