- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes LiGNO, a neural operator for PDEs on unstructured meshes and point clouds with varying geometries. The nodal data are mapped onto a regular latent grid, and the part the grid misses is kept separately at the nodes. The grid representation is then processed by a learnable multiresolution (lifting-based) backbone and mapped back to the nodes. No mesh connectivity is needed. The method is tested on several 2D and 3D benchmarks with varying geometries, plus one long autoregressive rollout task, against GINO, Transolver and PCNO.
Critical strengths
- The problem is relevant: operator learning on irregular, varying geometries without mesh connectivity.
- The point-to-grid front end is simple and cheap.
- The benchmarks are diverse in terms of geometry, and the long-horizon rollout results (Table 2) are striking.
Critical weaknesses
- In terms of novelty, as far as I can tell, the whole multilevel backbone is taken from LiNO, a recent arXiv preprint that the paper builds on but never compares against. Nothing in the method seems to depend on this particular backbone, so the actual contribution comes down to the front end: splatting, filling the empty cells, the geometry channels, and the nodal residual. That can be a fine contribution, but then I would expect to see either that this front end helps several grid backbones compared to plain interpolation or a GINO-style encoder, or that it works specifically well with the lifting backbone. Right now neither is shown. At a minimum, I would ask the authors to swap the backbone for FNO and WNO on the same latent grid (and, as a reference, to run them with plain interpolation), so that the effect of the front end can be separated from that of the backbone.
- The main claim is not really supported. The abstract says that no information is discarded before the operator acts. Prop. 2 is true for any choice of , since is just defined as the leftover (the proof says as much). More importantly, eq. 19 shows that only goes through a pointwise linear map and is added back at each node. It never passes through the operator and never interacts with neighbouring nodes, so in practice it is a skip connection. There are also no ablations at all. I would like to see at least: , the same backbone on a plainly interpolated grid, the harmonic fill against a simpler fill, and the model without the mask/density channels. Without these it is hard to say where the gains come from.
- Claims vs. results. The abstract stresses "any lattice resolution" and varying node counts, but there is no experiment across mesh or lattice resolutions. The appendix shows it is also the slowest model to train on DrivAer and Airfoil, with the highest memory use on DrivAer; this is not discussed, and inference times are missing. Finally, the Airfoil rollout uses one fixed mesh, so it says little about varying geometry, and MeshGraphNets is not included even though the dataset comes from that paper.
Questions for authors
See Weaknesses.
3 comments
W3: Claims vs. results (resolution, cost, AIRFOIL).
We thank the reviewer for this constructive and insightful feedback.
Mesh resolution. To evaluate LiGNO's generalization across mesh resolutions, we conduct additional experiments on Darcy flow. Specifically, we train LiGNO on meshes whose node sets were thinned by farthest-point sampling to a fraction uniformly drawn from . All models are tested on the same held-out problems at full resolution, so the model is tested at a denser resolution than every trained sample. Below, we present cross-discretization results in the mean relative error in % computed over the test set.
| Trained on / Tested on | 35% | 50% | 70% | 100% (Full Res) |
|---|---|---|---|---|
| 35-70% | 1.18 | 1.05 | 0.98 | 1.22 |
The results demonstrate that LiGNO maintains consistent accuracy across different test resolutions, despite being trained exclusively on subsampled meshes. In particular, the model generalizes effectively to full-resolution meshes without requiring retraining or additional fine-tuning. This suggests that training across a range of mesh resolutions enables LiGNO to learn representations that are robust to changes in discretization density.
Lattice resolution. Varying the latent lattice on Darcy flow:
Here we present the effect of latent lattice resolution on relative error (%), training time, and peak memory on Darcy flow.
| Lattice | 32² | 48² | 64² | 80² |
|---|---|---|---|---|
| Error (%) | 1.04 | 0.80 | 0.76 | 0.66 |
| Train time (h) | 0.27 | 0.31 | 0.39 | 0.49 |
| Peak memory (GB) | 1.50 | 2.25 | 3.29 | 4.62 |
Training cost and memory. The cost of LiGNO is governed by the lattice size and by the Jacobi sweeps retained for back-propagation: on Darcy flow, memory grows from 2.6 to 6.6 GB as goes from 0 to 40 and from 1.5 to 4.6 GB as the lattice goes from 32² to 80²; on airfoil-flap (240² lattice), the harmonic extension alone accounts for 12.1 → 25.4 GB and 3.4 → 5.5 h. We will update these trade-offs in the revised manuscript. Further on AIRFOIL, where the mesh is fixed, all geometry-dependent quantities (splat weights, occupancy and density channels, and the extended coordinate fields) should be computed once rather than at every step. With these changes, we expect a significant reduction in airfoil flow training time, this experiment, we will conduct before the final revision. Regarding inference time, the per-sample inference time is below few seconds for every model.
AIRFOIL Flow We agree with the reviewer that AIRFOIL uses a single fixed mesh, so it tests long-horizon temporal prediction rather than geometry generalization. As suggested by the reviewer, we trained MeshGraphNets (15 message-passing steps, latent size 128, mesh edges with relative-position features, node-type input) under the same training and rollout protocol of Table 2.
Relative -error in % (mean std over 200 test trajectories):
| Field | Model | 1 | 10 | 20 | 50 | 100 | 200 | 400 |
|---|---|---|---|---|---|---|---|---|
| MeshGraphNets | 1.51 ± 0.66 | 2.23 ± 0.71 | 3.07 ± 1.12 | 4.41 ± 2.14 | 6.07 ± 4.46 | 9.40 ± 8.30 | 16.56 ± 12.49 | |
| LiGNO | 1.76 ± 0.78 | 1.65 ± 0.63 | 1.73 ± 0.60 | 2.02 ± 0.75 | 2.87 ± 1.59 | 5.34 ± 4.29 | 10.33 ± 9.29 | |
| MeshGraphNets | 7.29 ± 1.70 | 7.84 ± 2.63 | 9.44 ± 3.40 | 10.43 ± 4.14 | 12.26 ± 5.58 | 18.43 ± 8.66 | 28.71 ± 12.43 | |
| LiGNO | 6.96 ± 1.96 | 5.990 ± 2.28 | 6.11 ± 2.70 | 5.98 ± 2.44 | 6.39 ± 2.59 | 9.15 ± 4.32 | 15.48 ± 9.46 | |
| MeshGraphNets | 0.62 ± 0.50 | 0.83 ± 0.58 | 1.16 ± 0.85 | 1.40 ± 0.97 | 1.88 ± 1.62 | 3.19 ± 3.21 | 5.69 ± 5.38 | |
| LiGNO | 0.54 ± 0.30 | 0.48 ± 0.29 | 0.54 ± 0.31 | 0.67 ± 0.41 | 0.84 ± 0.70 | 1.49 ± 1.49 | 2.82 ± 2.80 |
LiGNO outperforms MeshGraphNets on all three fields (, , and ) by 1.4–2.2× from 10 steps onwards, despite not using mesh connectivity. MeshGraphNets remains more accurate on at a single step. We will include these results in Table 2 in our revised manuscript.
We thank the reviewer again, these experiments have substantially clarified where the accuracy of LiGNO comes from, and we will incorporate them into the revised manuscript. We would also be happy to provide any additional clarification during the discussion period.
Regarding MeshGraphNets on AIRFOIL. As suggested, we added MeshGraphNets to the AIRFOIL benchmark. To make the comparison as fair as possible, we retrained all models with the protocol of MeshGraphNets (Pfaff et al., 2021): state-increment targets, training-noise injection, optimizer steps, and single-precision rollout, with the inflow nodes prescribed. Under this matched protocol, the baselines improve substantially relative to Table 2 of our manuscript, in which they were not trained with these stabilizers. The updated table below is presented below:
Both tables report the relative error in % (mean ± standard deviation over the 200 test trajectories), averaged over the first steps of a free autoregressive rollout from . Bold: best model per column.
Table A: all models trained with the MeshGraphNets protocol (with training noise).
| Field | Model | |||||||
|---|---|---|---|---|---|---|---|---|
| GINO | 6.87±3.53 | 20.61±14.70 | 24.84±16.50 | 38.38±32.90 | 63.82±64.50 | 126.10±121.00 | 240.55±198.00 | |
| MeshGraphNets | 1.51±0.66 | 2.23±0.71 | 3.07±1.12 | 4.41±2.14 | 6.07±4.46 | 9.40±8.30 | 16.56±12.49 | |
| Transolver | 1.54±0.59 | 1.82±0.71 | 2.30±1.32 | 3.68±2.32 | 6.46±3.63 | 11.55±6.85 | 17.23±8.63 | |
| PCNO | 3.69±1.13 | 5.01±2.29 | 6.00±3.61 | 11.38±10.71 | 22.96±22.54 | div. | div. | |
| LiGNO | 1.76±0.78 | 1.64±0.62 | 1.72±0.55 | 1.99±0.56 | 2.69±1.29 | 4.87±3.85 | 9.58±8.72 | |
| GINO | 25.29±4.45 | 44.04±2.31 | 50.70±4.53 | 68.64±20.20 | 98.60±54.50 | 161.84±131.00 | 355.32±386.00 | |
| MeshGraphNets | 7.29±1.70 | 7.84±2.63 | 9.44±3.40 | 10.43±4.14 | 12.26±5.58 | 18.43±8.66 | 28.71±12.43 | |
| Transolver | 6.06±1.80 | 6.07±2.87 | 7.33±4.52 | 10.35±6.71 | 15.21±10.19 | 23.83±16.85 | 33.19±23.22 | |
| PCNO | 15.09±3.07 | 17.81±6.12 | 19.02±6.53 | 28.17±12.22 | 39.26±27.66 | div. | div. | |
| LiGNO | 6.94±1.94 | 5.80±2.23 | 5.89±2.61 | 5.44±2.08 | 5.26±1.68 | 7.41±3.58 | 13.13±9.32 | |
| GINO | 2.72±1.11 | 5.68±2.71 | 9.14±6.05 | 15.29±11.50 | 21.08±14.00 | 26.82±14.80 | 32.12±16.30 | |
| MeshGraphNets | 0.62±0.50 | 0.83±0.58 | 1.16±0.85 | 1.40±0.97 | 1.88±1.62 | 3.19±3.21 | 5.69±5.38 | |
| Transolver | 0.50±0.32 | 0.56±0.31 | 0.62±0.24 | 0.83±0.41 | 1.31±0.97 | 2.63±2.22 | 3.81±2.76 | |
| PCNO | 1.78±0.59 | 2.07±0.81 | 2.06±1.17 | 4.29±4.57 | 8.45±12.24 | div. | div. | |
| LiGNO | 0.54±0.30 | 0.48±0.28 | 0.54±0.29 | 0.65±0.38 | 0.80±0.66 | 1.39±1.45 | 2.65±2.76 |
PCNO diverges (velocity error above 100%) on 52 and 74 of the 200 trajectories at and ; its median velocity errors there are 14.3% and 12.6%. A larger PCNO (71.6M parameters) trained on unwarped coordinates also diverged.
Table B: original manuscript, trained without training noise.
| Field | Model | |||||||
|---|---|---|---|---|---|---|---|---|
| GINO | 3.31±1.28 | 5.08±2.03 | 5.72±2.37 | 7.46±3.77 | 12.10±7.56 | 22.63±14.70 | 49.82±35.00 | |
| Transolver | 5.53±1.63 | 11.17±4.19 | 11.93±4.91 | 12.82±5.72 | 16.33±9.45 | 21.74±14.30 | 25.28±16.40 | |
| PCNO | 3.62±0.99 | 7.20±2.28 | 10.63±3.30 | 20.58±7.11 | 37.41±14.40 | 74.54±31.10 | 175.50±89.50 | |
| LiGNO | 1.76±0.78 | 1.65±0.63 | 1.73±0.60 | 2.02±0.75 | 2.87±1.59 | 5.34±4.29 | 10.33±9.29 | |
| GINO | 14.52±3.46 | 18.61±2.97 | 19.37±3.01 | 21.64±3.79 | 27.85±7.68 | 45.11±17.30 | 95.81±45.70 | |
| Transolver | 25.40±4.10 | 43.52±0.96 | 48.39±1.69 | 54.27±4.22 | 57.12±5.32 | 58.97±5.93 | 60.26±6.15 | |
| PCNO | 15.88±3.39 | 34.89±6.70 | 47.98±12.90 | 77.12±33.50 | 122.33±66.10 | 202.23±123.00 | 352.80±245.00 | |
| LiGNO | 6.96±1.96 | 5.90±2.28 | 6.11±2.70 | 5.98±2.44 | 6.39±2.59 | 9.15±4.32 | 15.48±9.46 | |
| GINO | 1.38±0.69 | 1.96±1.08 | 2.32±1.35 | 3.23±2.00 | 4.48±2.75 | 7.22±4.14 | 13.42±7.66 | |
| Transolver | 2.41±1.09 | 4.38±2.16 | 5.27±2.70 | 6.82±3.95 | 7.63±4.59 | 8.26±4.79 | 8.67±4.79 | |
| PCNO | 1.43±0.70 | 3.78±1.50 | 5.46±2.03 | 8.03±2.91 | 10.40±3.65 | 14.02±4.44 | 22.18±6.85 | |
| LiGNO | 0.54±0.30 | 0.48±0.29 | 0.54±0.31 | 0.67±0.41 | 0.84±0.70 | 1.49±1.49 | 2.82±2.80 |
Transolver improves substantially with training noise, whereas GINO is considerably more accurate without it. Under the noise protocol, GINO's rollout error exceeds that of the previous results from 10 steps onward. One-step errors are comparable across MeshGraphNets, Transolver, and LiGNO. From 10 steps onward, LiGNO is the most accurate model, and its advantage grows with the horizon and after 400 steps, its errors are 42% (), 54% (), and 30% () lower than those of the best baseline. In the final paper, we will report the best of the two protocols for each model.
We would like to thank the reviewer once again for the thorough review and the many helpful suggestions. We hope the additional experiments and clarifications have addressed the concerns raised.
We thank the reviewer for the careful assessment. We agree with the central technical point about Eq. (19), and we have run every ablation the reviewer requested, the backbone swap, and new experiments across lattice and mesh resolutions. Here we summarize the experiments. All errors below are mean relative ℓ2 errors (%) on the test set.
W1 Novelty: the backbone comes from LiNO, does the front end matter, and does the backbone?
LiNO operates on regular grids only, so it cannot be applied to these benchmarks directly. Its closest applicable form is the LiNO backbone behind a standard point-to-grid transfer, which corresponds to L0 and L1 in the table above. The full LiGNO front end improves on them by 1.9–5.2× on Darcy flow and about 3× on airfoil-flap, and it is also 1.8× more accurate than GINO's learned encoder/decoder with the same backbone for Darcy flow. Further, the backbone matters too, with our interface fixed, FNO is 2.2× worse on both datasets, smooth Darcy flow and shocked airfoil-flap.
W2 The main claim: Proposition 2, and the suggested ablations
The suggested ablations by the reviewer correspond to rows of the summary table above:
Same backbone with simple grid transfer: L0 and L1 use nearest-cell and linear B-spline transfer, respectively, with zero fill and no geometry channels. The full geometric lifting front end substantially improves accuracy on both benchmarks, reducing the errors from 3.40/1.25% to 0.63% on Darcy flow and from 5.53/5.21% to 1.74% on airfoil-flap (W1).
Harmonic extension vs. zero fill: Comparing L1 and L2, harmonic extension has little effect on Darcy flow, where the domain leaves no holes in the lattice, but improves airfoil-flap accuracy by 1.8× (5.21 → 2.85%), where the airfoil and flap create unoccupied lattice regions. With mask and density channels fixed, zero fill achieves 0.69% error on Darcy flow.
Contribution of mask and density channels: Comparing L2 and L3, these channels roughly halve the error on both benchmarks, from 1.37 to 0.69% on Darcy flow and from 2.85 to 1.81% on airfoil-flap. The density channel accounts for most of this improvement (1.28% on Darcy flow with the mask alone), as it captures sub-cell boundary-location information.
Contribution of the detail branch: Removing Wdet (L3) yields comparable performance to the full model (L4): 0.69 vs. 0.63% on Darcy flow and 1.81 vs. 1.74% on airfoil-flap. Replacing the branch with a conventional skip connection (L4S) gives 0.69% and 1.79%, respectively.
The nodal residual branch has a limited effect on accuracy in these benchmarks, while keeping it (L4), removing it (L3), or replacing it with a pointwise skip (L4S) gives comparable errors. This is consistent with the reviewer’s observation that the branch acts as a skip connection. Its role is structural, it keeps the level-zero transform invertible, so no input information is lost before the operator acts. We will revise the manuscript accordingly. The contributions will emphasize the geometric lifting interface (B-spline transfer, harmonic extension, geometry channels) and its combination with the multiresolution backbone.