acceptodds
Under review as a conference paper at ICLR 2027

Operator learning for PDEs on unstructured meshes with varying geometries via second-generation wavelets

15 views

Abstract

Learning solution operators for partial differential equations on unstructured meshes with varying geometry remains a fundamental challenge for neural operators. Existing approaches either transform scattered data onto a regular grid as a preprocessing step, discarding what the grid cannot represent, or encode data on the mesh graph, where cost and receptive field are tied to a constructed neighborhood. We propose LiGNO, a lifting-based geometric neural operator built on a second-generation wavelet in which we treat the point cloud itself as the finest level of a multiresolution hierarchy. A geometric lifting step decomposes nodal features into a coarse field on a fixed lattice and a nodal residual carrying the fine-scale detail. Together, they reconstruct the nodal data exactly, at any lattice resolution, so no information is discarded before the operator acts. A multi-level parameterized second-generation wavelet transform then recursively decomposes coarse and directional detail coefficients, enabling the operator to model physical interactions at multiple scales. We evaluate LiGNO against state-of-the-art methods on Darcy flow, the Poisson equation, and aerodynamic flows on DrivAer and airfoil geometries, with deformed domains and varying numbers of nodal points.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Neural operator, unstructured mesh, second generation wavelet, multi-scale PDEs

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_TBCJOpenReview ↗
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 AA, since dd is just defined as the leftover (the proof says as much). More importantly, eq. 19 shows that dd only goes through a pointwise linear map WdetW_{det} 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: Wdet=0W_{det}=0, 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

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\ell_2 errors (%) on the test set.

Point to grid transferDarcy flowAirfoil-flap
L03.405.53
L11.255.21
L21.372.85
L30.691.81
L40.631.74
L4S0.691.79
GL1.23#
L3F1.523.83
(# denotes experiments that are currently running, results will be reported once completed.)
where the different experiments are defined as follows:
  • L0: Nearest-cell transfer, where each node is assigned to its nearest lattice cell.
  • L1: L0 replaced with linear B-spline interpolation, distributing each node across its 2rd2^{r_d} neighboring cells.
  • L2: L1 + harmonic extension to fill unoccupied cells.
  • L3: L2 + occupancy mm and density log⁡(1+w)\log(1+w) as additional lattice channels, with Wdet⁡=0W_{\det}=0.
  • L4: L3 + the proposed nodal detail branch Wdet⁡dW_{\det}d, representing the complete LiGNO architecture.
  • L4S: L4 with the nodal detail branch replaced by a conventional pointwise skip connection WvWv.
  • GL: The proposed geometric lifting interface replaced with GINO's encoder-decoder, while retaining the LiNO backbone.
  • L3F: L3 interface with the LiNO backbone replaced by a parameter-matched FNO.

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⁡W_{\det} (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.

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 [0.35,0.7][0.35, 0.7]. 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 ℓ2\ell_2 error in % computed over the test set.

Trained on ↓\downarrow / Tested on →\rightarrow35%50%70%100% (Full Res)
35-70%1.181.050.981.22

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 ℓ2\ell_2 error (%), training time, and peak memory on Darcy flow.

Lattice32²48²64²80²
Error (%)1.040.800.760.66
Train time (h)0.270.310.390.49
Peak memory (GB)1.502.253.294.62

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 TT 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 ℓ2\ell_2-error in % (mean ±\pm std over 200 test trajectories):

FieldModel1102050100200400
vxv_xMeshGraphNets1.51 ± 0.662.23 ± 0.713.07 ± 1.124.41 ± 2.146.07 ± 4.469.40 ± 8.3016.56 ± 12.49
vxv_xLiGNO1.76 ± 0.781.65 ± 0.631.73 ± 0.602.02 ± 0.752.87 ± 1.595.34 ± 4.2910.33 ± 9.29
vyv_yMeshGraphNets7.29 ± 1.707.84 ± 2.639.44 ± 3.4010.43 ± 4.1412.26 ± 5.5818.43 ± 8.6628.71 ± 12.43
vyv_yLiGNO6.96 ± 1.965.990 ± 2.286.11 ± 2.705.98 ± 2.446.39 ± 2.599.15 ± 4.3215.48 ± 9.46
ρ\rhoMeshGraphNets0.62 ± 0.500.83 ± 0.581.16 ± 0.851.40 ± 0.971.88 ± 1.623.19 ± 3.215.69 ± 5.38
ρ\rhoLiGNO0.54 ± 0.300.48 ± 0.290.54 ± 0.310.67 ± 0.410.84 ± 0.701.49 ± 1.492.82 ± 2.80

LiGNO outperforms MeshGraphNets on all three fields (vxv_x, vyv_y, and pp) by 1.4–2.2× from 10 steps onwards, despite not using mesh connectivity. MeshGraphNets remains more accurate on vxv_x 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, 1.25×1051.25\times10^{5} 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 ℓ2\ell_2 error in % (mean ± standard deviation over the 200 test trajectories), averaged over the first hh steps of a free autoregressive rollout from u0u^0. Bold: best model per column.

Table A: all models trained with the MeshGraphNets protocol (with training noise).

FieldModelh=1h=1101020205050100100200200400400
vxv_xGINO6.87±3.5320.61±14.7024.84±16.5038.38±32.9063.82±64.50126.10±121.00240.55±198.00
MeshGraphNets1.51±0.662.23±0.713.07±1.124.41±2.146.07±4.469.40±8.3016.56±12.49
Transolver1.54±0.591.82±0.712.30±1.323.68±2.326.46±3.6311.55±6.8517.23±8.63
PCNO3.69±1.135.01±2.296.00±3.6111.38±10.7122.96±22.54div.†^\daggerdiv.†^\dagger
LiGNO1.76±0.781.64±0.621.72±0.551.99±0.562.69±1.294.87±3.859.58±8.72
vyv_yGINO25.29±4.4544.04±2.3150.70±4.5368.64±20.2098.60±54.50161.84±131.00355.32±386.00
MeshGraphNets7.29±1.707.84±2.639.44±3.4010.43±4.1412.26±5.5818.43±8.6628.71±12.43
Transolver6.06±1.806.07±2.877.33±4.5210.35±6.7115.21±10.1923.83±16.8533.19±23.22
PCNO15.09±3.0717.81±6.1219.02±6.5328.17±12.2239.26±27.66div.†^\daggerdiv.†^\dagger
LiGNO6.94±1.945.80±2.235.89±2.615.44±2.085.26±1.687.41±3.5813.13±9.32
ρ\rhoGINO2.72±1.115.68±2.719.14±6.0515.29±11.5021.08±14.0026.82±14.8032.12±16.30
MeshGraphNets0.62±0.500.83±0.581.16±0.851.40±0.971.88±1.623.19±3.215.69±5.38
Transolver0.50±0.320.56±0.310.62±0.240.83±0.411.31±0.972.63±2.223.81±2.76
PCNO1.78±0.592.07±0.812.06±1.174.29±4.578.45±12.24div.†^\daggerdiv.†^\dagger
LiGNO0.54±0.300.48±0.280.54±0.290.65±0.380.80±0.661.39±1.452.65±2.76

†^\daggerPCNO diverges (velocity error above 100%) on 52 and 74 of the 200 trajectories at h=200h=200 and 400400; 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.

FieldModelh=1h=1101020205050100100200200400400
vxv_xGINO3.31±1.285.08±2.035.72±2.377.46±3.7712.10±7.5622.63±14.7049.82±35.00
Transolver5.53±1.6311.17±4.1911.93±4.9112.82±5.7216.33±9.4521.74±14.3025.28±16.40
PCNO3.62±0.997.20±2.2810.63±3.3020.58±7.1137.41±14.4074.54±31.10175.50±89.50
LiGNO1.76±0.781.65±0.631.73±0.602.02±0.752.87±1.595.34±4.2910.33±9.29
vyv_yGINO14.52±3.4618.61±2.9719.37±3.0121.64±3.7927.85±7.6845.11±17.3095.81±45.70
Transolver25.40±4.1043.52±0.9648.39±1.6954.27±4.2257.12±5.3258.97±5.9360.26±6.15
PCNO15.88±3.3934.89±6.7047.98±12.9077.12±33.50122.33±66.10202.23±123.00352.80±245.00
LiGNO6.96±1.965.90±2.286.11±2.705.98±2.446.39±2.599.15±4.3215.48±9.46
ρ\rhoGINO1.38±0.691.96±1.082.32±1.353.23±2.004.48±2.757.22±4.1413.42±7.66
Transolver2.41±1.094.38±2.165.27±2.706.82±3.957.63±4.598.26±4.798.67±4.79
PCNO1.43±0.703.78±1.505.46±2.038.03±2.9110.40±3.6514.02±4.4422.18±6.85
LiGNO0.54±0.300.48±0.290.54±0.310.67±0.410.84±0.701.49±1.492.82±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% (vxv_x), 54% (vyv_y), and 30% (ρ\rho) 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.

Reviewer_5bYtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LiGNO, a neural operator for PDEs on irregular point sets with varying geometries. The method constructs a regular lattice representation using B-spline splatting and harmonic extension while retaining a nodal residual that permits exact reconstruction of the encoded input. The lattice field is processed using a learnable multiresolution lifting architecture adopted from LiNO, and the result is combined with the transformed nodal residual. Experiments cover variable-domain elliptic problems, aerodynamic surface prediction, and autoregressive airfoil flow prediction.

Critical strengths

The paper is clearly written and well organized. The main architectural stages and their mathematical definitions are presented accessibly.

The evaluation covers several application settings. The reported results are competitive with the selected baselines, including the lowest mean errors on three of the four steady-state benchmarks and lower errors throughout the reported airfoil rollout horizons.

Critical weaknesses
  1. The significance of the incremental contribution over LiNO is insufficiently established.

Section 2.2.4 adopts the learned lifting framework of LiNO, including predict–update decomposition, multilevel processing, and inverse synthesis. The principal addition therefore appears to be the geometric lifting construction and its integration with this backbone. Extending an existing architecture to irregular point sets and varying geometries can be valuable. However, the current experiments do not isolate the additional benefit of the proposed geometric construction over simpler geometry-transfer mechanisms using the same backbone. The paper should clarify the new methodological insight and provide controlled evidence that this extension delivers a substantive advantage.

  1. The architectural choices require stronger theoretical and practical justification.

Preserving information when transferring irregular data to a lattice is a reasonable motivation. However, the paper does not sufficiently explain why its particular construction should improve operator learning compared with other geometry-aware architectures. I would like both a theoretical rationale and empirical evidence supporting the major components: B-spline transfer, harmonic extension, the nodal residual pathway, geometric input channels, and multiresolution coefficient processing. What advantage does each component provide, and how do these advantages combine?

  1. The significance of Propositions 1 and 2 needs clarification.

Proposition 1 establishes adjointness of the unnormalized splat and gather operators. However, the actual analysis pathway also includes normalization and harmonic extension. What useful property does this adjointness establish for the complete architecture?

Proposition 2 establishes exact reconstruction because the nodal detail is defined as the reconstruction residual. This preserves the encoded input, but does not by itself establish improved PDE prediction, stability, or generalization. The authors should explain the theoretical and practical consequences of these propositions, identify any additional assumptions needed for stronger guarantees, and provide evidence that these properties contribute to performance.

  1. Component ablations and hyperparameter sensitivity studies are missing.

The experiments evaluate the complete architecture without isolating its components. Prioritized ablations on representative benchmarks should examine the nodal residual branch, harmonic extension, geometric channels, and multiresolution processing.

Sensitivity to the latent lattice resolution, lifting depth (L), channel width (C), occupancy threshold (\tau), and number of Jacobi sweeps (T) would also help justify the selected configuration. These studies should report accuracy alongside computational cost to establish which choices matter and how robust the reported advantages are.

  1. Whole-model comparisons do not isolate advantages over other geometry-aware methods.

Component-level comparisons would help separate improvements in geometry handling from improvements in the grid-based operator. Where interfaces permit, the authors could compare alternative point-to-grid and grid-to-point mechanisms under a common backbone, and alternative backbones under the same geometric representation. For example, a controlled comparison with GINO-style transfer would help explain what geometric lifting contributes. Such comparisons should use reasonably comparable capacity and training conditions.

  1. The accuracy–resource tradeoffs require more explanation.

Tables 4–6 provide useful information, but parameter counts and memory requirements vary substantially across some comparisons. On Darcy flow, PCNO has 71.58M parameters compared with LiGNO’s 2.40M. On DrivAer, LiGNO and GINO have similar parameter counts—14.21M and 13.67M—but reported peak memory usages of 25.13 GB and 3.80 GB.

These differences do not inherently invalidate the comparisons, but their causes and implications should be discussed. How were baseline configurations selected and tuned? Were memory measurements obtained under comparable batch sizes and measurement conditions? Selected comparisons at comparable capacity or computational budgets would make the architectural advantages easier to assess.

  1. [Bonus]Resolution transfer and super-resolution require clarification.

Does the multiresolution construction provide an advantage when evaluating on finer discretizations or predicting fine-scale outputs from coarse observations? These are distinct settings, and multiresolution processing alone does not establish either capability. In particular, how would the nodal residual pathway operate at new query locations? If such benefits are claimed, what theoretical argument under stated assumptions and what controlled empirical evidence support them? Otherwise, the authors should clarify the scope of the method’s resolution generalization.

Questions for authors

Please see weaknesses

7 comments

We thank the reviewer for the detailed and constructive feedback. We appreciate the feedback and address each of the reviewer's points below. We would also be happy to provide any additional clarification during the discussion period. To address the reviewer’s concerns, we conducted additional experiments covering component-wise ablations, comparisons with GINO-style transfer and an alternative backbone, the significance of Propositions 1–2, hyperparameter sensitivity, computational cost, and generalization across different resolutions. All results are reported in terms of mean relative ℓ2\ell_2 error (%) on the test set and are summarized in the table below.

Point to grid transferDarcy flowAirfoil-flap
L03.405.53
L11.255.21
L21.372.85
L30.691.81
L40.631.74
L4S0.691.79
GL1.23#
L3F1.523.83
(# denotes experiments that are currently running, results will be reported once completed.)

where the different experiments are defined as follows:

  • L0: Nearest-cell transfer, where each node is assigned to its nearest lattice cell.
  • L1: L0 replaced with linear B-spline interpolation, distributing each node across its 2rd2^{r_d} neighboring cells.
  • L2: L1 + harmonic extension to fill unoccupied cells.
  • L3: L2 + occupancy mm and density log⁡(1+w)\log(1+w) as additional lattice channels, with Wdet⁡=0W_{\det}=0.
  • L4: L3 + the proposed nodal detail branch Wdet⁡dW_{\det}d, representing the complete LiGNO architecture.
  • L4S: L4 with the nodal detail branch replaced by a conventional pointwise skip connection WvWv.
  • GL: The proposed geometric lifting interface replaced with GINO's encoder-decoder, while retaining the LiNO backbone.
  • L3F: L3 interface with the LiNO backbone replaced by a parameter-matched FNO.

W1 & W5: Contribution over LiNO, and comparison of geometry transfer and 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. The methodological insight is therefore twofold. First, how irregular data are transferred to the lattice matters as much as the operator acting on it and secondly, given a well-designed transfer, the multiresolution lifting backbone is substantially more accurate than a spectral backbone, on both a smooth elliptic problem and an airfoil-flap problem with shocks.

W2: Rationale and evidence for each component, and how they combine

We thank the reviewer for raising this important question regarding the motivation and rationale behind the individual components of the LiGNO architecture. In response, we discuss our reasoning behind using such architecture and what purpose each component serves with empirical evidence through targeted ablation studies to demonstrate its contribution to the model's performance. Each component addresses a distinct challenge in operator learning on irregular domains.

The linear B-spline synthesis reproduces affine fields and interpolates smooth fields with O(h2)O(h^2) error, compared with O(h)O(h) for nearest-cell transfer. This reduces the Darcy flow error by 2.7× (L0 → L1). Harmonic extension smoothly fills unoccupied cells and avoids artificial jumps from zero filling. It improves airfoil-flap accuracy by 1.8× (L1 → L2), while having a negligible effect on Darcy flow, which has no interior holes. The occupancy mask identifies observed cells, while the density channel log⁡(1+w)\log(1+w) captures sub-cell boundary locations. Together, they roughly halve the error on both benchmarks (L2 → L3). The nodal residual pathway preserves information that may be lost during grid projection, although its impact on accuracy is modest as seen in the table above (see L3, L4, and L4S). Finally, multiresolution coefficient processing expands the receptive field across scales, enabling efficient modelling of non-local interactions. Increasing depth from L=1 to L=4 reduces Darcy error from 2.69% to 0.68% (see table in W4). Thus, the components work together to address interpolation error, unoccupied regions, geometric information loss, preservation of nodal details, and non-local dependencies.

W3: Significance of Propositions 1 and 2 We thank the reviewer for these constructive questions regarding the theoretical and practical roles of Propositions 1 and 2. While Proposition 1 establishes Euclidean adjointness (Sh=Γh⊤S_h = \Gamma_h^\top) for the unnormalized operators, combining it with weight normalization Πh=Dw−1Sh\Pi_h = D_w^{-1} S_h yields exact adjointness under the sampling-density-weighted inner product, ⟨Πhv,z⟩Dw=⟨Shv,z⟩=⟨v,Γhz⟩\langle \Pi_h v, z \rangle_{D_w} = \langle S_h v, z \rangle = \langle v, \Gamma_h z \rangle. Because the harmonic extension Hm\mathcal{H}_m solves a discrete Dirichlet problem that fixes values on occupied cells Λocc\Lambda occ, it acts as the identity on the sampled support while smoothly extending into unoccupied cells so boundary discontinuities do not inject spurious high-frequency oscillations into the subsequent second-generation wavelet lifting transform. Furthermore, because analysis (Πh\Pi_h) and synthesis (Γh\Gamma_h) share the same B-spline basis βk\beta_k, forward feature aggregation and backward gradient propagation remain spatially aligned without interpolation mismatch. Regarding Proposition 2, we agree with the reviewer that exact reconstruction is a structural property rather than a source of predictive accuracy. It follows directly from defining the nodal detail as the residual and guarantees that no input information is lost at the first level, thereby ensuring the invertibility of the transform. However, it does not inherently improve prediction. Our new ablation results support this observation: removing the nodal branch (L3) or replacing it with a standard skip connection (L4S) yields comparable accuracy to the full model (0.69% and 0.69% vs. 0.63% on Darcy flow, and 1.81% and 1.79% vs. 1.74% on airfoil-flap, respectively). Accordingly, we will revise the manuscript to clarify that Proposition 2 establishes the invertibility of the proposed transform, rather than serving as an explanation for the observed accuracy improvements.

W6: Accuracy–resource trade-offs

Parameter count reflects a model's architectural capacity, but memory consumption is also strongly influenced by how the model represents the input data. In LiGNO, memory and training time are dominated by the latent lattice resolution and by the number of Jacobi sweeps retained for back-propagation, while the number of input nodes enters only linearly through the pointwise encoder, splat and gather. On Darcy flow, memory increases from 1.5 to 4.6 GB as the lattice is refined from 32232^2 to 80280^2, and from 2.6 to 6.6 GB as TT increases from 0 to 40. On airfoil-flap (2402240^2 lattice), the harmonic extension alone increases memory from 12.1 to 25.4 GB and training time from 3.4 to 5.5 h (L1 vs. L2). This is consistent with the higher memory requirements observed on benchmarks with larger latent lattices. Our sensitivity study also identifies more efficient configurations with accuracy comparable to the default (C=32C=32: 0.68% with 0.96M parameters and 2.9 GB; T=5T=5: 0.66% with 3.1 GB). Moreover, since the encoder is affine and the harmonic extension is linear and channel-wise, the extension can be applied to the few input channels rather than to all CC latent channels without changing the output; we will adopt this to further reduce memory.

Regarding measurement conditions: in Tables 4–5, memory was recorded at each model's training batch size (LiGNO batch size: 32 on Darcy flow, 24 on airfoil-flap, and 12 on DrivAer), whereas GINO and Transolver can process one geometry at a time (batch size 1), which largely accounts for the differences in reported memory, particularly on DrivAer and airfoil-flap.

W7: Resolution transfer and super-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 [0.35,0.7][0.35, 0.7]. 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 ℓ2\ell_2 error in % computed over the test set.

Trained on ↓\downarrow / Tested on →\rightarrow35%50%70%100% (Full Res)
35-70%1.181.050.981.22

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.

W4: Component ablations and hyperparameter sensitivity

We thank the reviewer for this constructive suggestion. We agree that a hyperparameter sensitivity analysis is important for understanding the influence of key tunable parameters on model performance. The component ablations suggested by the reviewer are presented in a summary table and are discussed in W2. For sensitivity, we varied one hyperparameter at a time on Darcy flow around the default (s=80s=80, L=4L=4, C=64C=64, τ=10−6\tau=10^{-6}, T=20T=20), reporting accuracy alongside cost:

SettingValueError (%)Params (M)Train (h)Memory (GB)
default–0.632.400.554.62
lattice ss32 / 48 / 641.04 / 0.80 / 0.762.400.27 / 0.31 / 0.391.50 / 2.25 / 3.29
depth LL0 / 1 / 2 / 326.31 / 2.69 / 0.66 / 0.740.14 / 0.70 / 1.27 / 1.840.31 / 0.42 / 0.45 / 0.473.48 / 4.31 / 4.53 / 4.60
width CC32 / 48 / 960.68 / 0.69 / 0.830.96 / 1.58 / 4.650.35 / 0.42 / 0.692.90 / 3.75 / 6.45
threshold τ\tau10−210^{-2} / 10−110^{-1} / 0.50.68 / 0.74 / 0.692.400.494.62
Jacobi sweeps TT0 / 5 / 10 / 400.69 / 0.66 / 0.66 / 0.662.400.37 / 0.39 / 0.43 / 0.612.62 / 3.12 / 3.62 / 6.62

Lifting depth LL Without the multiresolution hierarchy (L=0L=0) the model fails (26.31%), with a single level it remains far from the default (2.69%), and from L=2L=2 onward it recovers the default accuracy (0.63–0.74%). Despite comparable parameter counts (0.70M vs. 0.96M), the single-level model performs substantially worse than the four-level model (2.69% vs. 0.68%), indicating that the performance came from multiresolution coefficient processing discussed in W2.

Lattice resolution ss controls the accuracy–cost trade-off. The error decreases monotonically from 1.04% to 0.63% as the lattice is refined from 32232^2 to 80280^2, while memory and training time grow with the lattice size (from 1.50 to 4.62 GB and from 0.27 to 0.49 h).

The occupancy threshold τ\tau, the number of Jacobi sweeps TT and the width CC have a minor effect on Darcy flow: for τ\tau from 10−610^{-6} to 0.50.5, from 0 to 40 sweeps, and for C=32C=32 and 4848, the error stays within about 0.1 percentage points of the default. The number of sweeps mainly affects memory (from 2.62 GB at T=0T=0 to 6.62 GB at T=40T=40), so a small number of sweeps (T=5T=5) gives comparable accuracy at lower cost. Similarly, C=32C=32 and C=48C=48 reach 0.68–0.69% with 40–66% of the default parameter count, whereas increasing the width to C=96C=96 does not improve accuracy at a fixed training budget. We expect TT and τ\tau to matter more on non-smooth, multi-body problems such as airfoil-flap, where the harmonic extension has a much larger effect (1.8× improvement from L1 to L2, against a neutral effect on Darcy flow), since both control how unoccupied cells around the bodies are filled.

We performed this study on Darcy flow because each airfoil-flap run takes about 4–5 hours, making it difficult to accommodate a full sweep within the rebuttal period. Later, we will extend the sensitivity study to the non-smooth airfoil-flap benchmark.

Reviewer_tsa7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LiGNO, a neural operator for PDEs on unstructured meshes with varying geometries. The method combines a geometric lifting mechanism, consisting of B-spline splatting, harmonic extension, and nodal residual preservation, with a learnable multiresolution lifting backbone. The authors evaluate LiGNO on four steady-state PDE benchmarks and a long-horizon autoregressive airfoil-flow benchmark, comparing it against GINO, Transolver, and PCNO.

Critical strengths
  1. The paper addresses an important problem in scientific machine learning: learning PDE solution operators on irregular discretizations and geometry-dependent domains without requiring explicit mesh connectivity.

  2. The proposed architecture provides a practical combination of structured multiscale processing and node-level residual preservation, allowing the model to accommodate varying numbers of nodes.

  3. The experimental results are promising. LiGNO achieves the best reported performance on three of the four steady-state benchmarks. In particular, the substantial improvements in the 400-step autoregressive AIRFOIL experiments are noteworthy.

Critical weaknesses

1. The methodological novelty and the contribution of the geometric lifting mechanism could be better established.

The multilevel learnable predict-update transformations and the evolution of coarse and directional detail coefficients are largely based on LiNO (Pandey et al., 2026) [1]. In addition, the general framework of mapping irregular discretizations to a structured latent grid for operator learning has been explored in GINO (Li et al., NeurIPS 2023) [2].

The main methodological contribution of LiGNO therefore appears to lie in its geometric lifting interface, which combines B-spline projection, harmonic extension, and nodal residual preservation.

However, the exact reconstruction property in Proposition 2 follows directly from the definition of the residual coefficients. Furthermore, Eq. (19) can be algebraically reformulated as a pointwise skip connection combined with a latent-grid correction, assuming the stated shared linear detail mapping. It remains unclear whether the proposed residual representation provides additional benefits beyond conventional residual connections.

I recommend adding several targeted ablation experiments to better isolate the contribution of the proposed mechanism:

  • Compare LiGNO with a standard point-to-grid adapter using the same LiNO backbone.
  • Replace the proposed nodal detail branch with a conventional pointwise skip connection.
  • Evaluate the proposed geometric interface with an alternative backbone, such as FNO or CNO.

These experiments would help clarify whether the performance improvements arise primarily from the geometric lifting mechanism, the existing multiresolution backbone, or their combination.

2. The experimental comparison could be strengthened with more recent geometry-aware neural operators.

The current evaluation considers GINO (2023), Transolver (2024), and PCNO (2025). While these baselines represent different approaches to neural operator learning, several recent and closely related methods are not included.

In particular, RIGNO (Mousavi et al., NeurIPS 2025) [3] addresses operator learning on arbitrary domains and investigates generalization across spatial resolutions, making it highly relevant to the proposed setting.

For the large-scale 3D DrivAer benchmark, Transolver++ (Luo et al., ICML 2025) [4] and GAOT (Wen et al., NeurIPS 2025) [5] provide relevant comparisons for scalable geometry-aware PDE learning.

GINOT (Liu et al., CMAME 2026) [6] is also closely related, especially considering that the Poisson benchmark used in this paper is attributed to the corresponding work.

I suggest including representative comparisons with these recent methods, particularly:

  • RIGNO on Darcy flow or another variable-domain benchmark.
  • GINOT on the Poisson benchmark.
  • Transolver++ or GAOT on the DrivAer benchmark, where computational resources permit.

I do not expect all recent architectures to be reproduced across every dataset. A small number of carefully selected comparisons under matched data splits and evaluation protocols would provide a more convincing assessment of LiGNO's accuracy and computational efficiency relative to recent state-of-the-art methods.

3. Additional experiments would help substantiate the claims regarding discretization generalization.

The current experiments demonstrate that LiGNO can process samples with varying geometries and node counts. However, varying node counts across different samples do not necessarily establish generalization across discretizations.

I recommend adding a controlled cross-discretization experiment. For example, the authors could train LiGNO on a particular range of mesh resolutions and evaluate the trained model on independently remeshed versions of the same physical problems, using different node densities or sampling distributions.

Such an experiment would help distinguish robustness to varying input sizes from genuine discretization generalization.

Additionally, the AIRFOIL autoregressive benchmark uses a fixed mesh. Although its long-horizon results are impressive, this experiment primarily demonstrates temporal prediction performance.

I suggest clarifying this distinction in the experimental discussion and ensuring that the generalization claims are consistent with the specific properties evaluated by each benchmark.

Questions for authors
  1. Could the authors provide ablations that separate the contributions of the geometric lifting interface, the LiNO backbone, and the nodal residual branch? In particular, how does the proposed mechanism compare with a standard point-to-grid adapter and a conventional pointwise skip connection?

  2. Could the authors consider additional comparisons with recent geometry-aware neural operators, such as RIGNO, GINOT, and Transolver++ or GAOT, on the most relevant benchmarks?

  3. Could the authors provide evidence of generalization to unseen mesh resolutions or sampling distributions? For the autoregressive AIRFOIL experiment, were all baselines evaluated under identical training targets and rollout protocols?

References

[1] Pandey et al. LiNO: Lifting Based Multiresolution Neural Operator. arXiv:2607.02715, 2026.

[2] Li et al. Geometry-Informed Neural Operator for Large-Scale 3D PDEs. NeurIPS, 2023.

[3] Mousavi et al. RIGNO: A Graph-based Framework for Robust and Accurate Operator Learning for PDEs on Arbitrary Domains. NeurIPS, 2025.

[4] Luo et al. Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries. ICML, 2025.

[5] Wen et al. Geometry Aware Operator Transformer as an Efficient and Accurate Neural Surrogate for PDEs on Arbitrary Domains. NeurIPS, 2025.

[6] Liu et al. Geometry-Informed Neural Operator Transformer for Partial Differential Equations on Arbitrary Geometries. Computer Methods in Applied Mechanics and Engineering, 2026.

7 comments

We sincerely thank the reviewer for their positive and encouraging feedback on our work. We greatly appreciate their recognition of the significance of the problem, the architectural design of LiGNO, and its promising experimental performance. We appreciate the feedback and address each of the reviewer's points below. We would also be happy to provide any additional clarification during the discussion period.

Q1: Ablation of architectural components? We thank the reviewer for this valuable suggestion. To address this point, we have conducted additional ablation studies to isolate the contributions of the geometric lifting interface, the LiNO backbone, and the nodal residual branch. We also compare the proposed geometric lifting with a standard point-to-grid adapter (as GINO) and the nodal residual branch with a conventional pointwise skip connection. These experiments are performed on two benchmark problems, particularly Darcy flow (smooth solutions), and airfoil-flap (multiresolution solutions). The corresponding experiments and their results are presented below.

Ablation results: Mean relative ℓ2\ell_2-error in % across test samples

Point to grid transferDarcy flowAirfoil-flap
L03.405.53
L11.255.21
L21.372.85
L30.691.81
L40.631.74
L4S0.691.79
GL1.23#
L3F1.523.83
(# denotes experiments that are currently running, results will be reported once completed.)
where the different experiments are defined as follows:
  • L0: Nearest-cell transfer, where each node is assigned to its nearest lattice cell.
  • L1: L0 replaced with linear B-spline interpolation, distributing each node across its 2rd2^{r_d} neighboring cells.
  • L2: L1 + harmonic extension to fill unoccupied cells.
  • L3: L2 + occupancy mm and density log⁡(1+w)\log(1+w) as additional lattice channels, with Wdet⁡=0W_{\det}=0.
  • L4: L3 + the proposed nodal detail branch Wdet⁡dW_{\det}d, representing the complete LiGNO architecture.
  • L4S: L4 with the nodal detail branch replaced by a conventional pointwise skip connection WvWv.
  • GL: The proposed geometric lifting interface replaced with GINO's encoder-decoder, while retaining the LiNO backbone.
  • L3F: L3 interface with the LiNO backbone replaced by a parameter-matched FNO.

The ablation table adds one interface component at a time while keeping the LiNO backbone, lattice, and training fixed. The main finding is that the geometric interface is the primary contributor to accuracy, and each component provides clear benefits in its intended role. The B-spline transfer reduces Darcy flow error 2.7× (3.40 → 1.25%), because linear interpolation is far more accurate than nearest-cell on smooth fields. The harmonic fill improves airfoil-flap error 1.8× (5.21 → 2.85%), because the airfoil and flap leave holes in the lattice, unlike on Darcy flow, which has no holes. The geometry channels halve the error on both problems (Darcy flow 1.37 → 0.69%, airfoil-flap 2.85 → 1.81%). The backbone matters too, with our interface fixed, FNO is 2.2× worse on both datasets, smooth Darcy flow and shocked airfoil-flap.

Further, using GINO’s learned adapter with the same backbone, our interface is 1.8× more accurate (0.69% vs 1.23%). Against plain transfers with the same backbone (L0 and L1), the LiGNO interface is 1.9–5.2× better on Darcy flow and about 3× better on airfoil-flap.

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 therefore 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.

Q2: Comparison with recent geometry-aware neural operators? We have included additional comparisons with RIGNO, GINOT, Transolver++ on the respective benchmarks suggested by the reviewer, namely Darcy flow, Poisson equation, and DrivAer respectively. The corresponding results are presented below.

Relative ℓ2\ell_2-error (mean ±\pm standard deviation) in % computed over the test set.

Darcy flow

GINOTransolverPCNORIGNOLiGNO
0.97 ±\pm 0.712.15 ±\pm 1.361.14 ±\pm 0.854.79 ±\pm 0.390.63 ±\pm 0.08

Poisson equation

GINOTransolverPCNOGINOTLiGNO
1.53 ±\pm 0.720.61 ±\pm 1.170.33 ±\pm 0.241.16 ±\pm 0.490.48 ±\pm 0.18

DrivAer

GINOTransolverPCNOTransolver++LiGNO
20.83 ±\pm 3.8419.57 ±\pm 4.1225.32 ±\pm 2.8318.54 ±\pm 3.7217.85 ±\pm 3.43

Due to time constraints, Transolver++ was trained using 50,000 randomly sampled points per geometry at each training step, while evaluation was performed on the full mesh. We use this setting to make the comparison feasible.

Q3: Generalization to unseen mesh resolutions or sampling distributions? We thank the reviewer for the insightful comment. To evaluate LiGNO's generalization across mesh resolutions, we conduct additional experiments on Darcy flow. We train LiGNO on meshes whose node sets were thinned by farthest-point sampling to a fraction uniformly drawn from [0.35,0.7][0.35, 0.7]. 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 ℓ2\ell_2 error in % computed over the test set.

Trained on ↓\downarrow / Tested on →\rightarrow35%50%70%100% (full res)
35-70%1.181.050.981.22

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 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.

Regarding airfoil flow protocol: Yes, for the autoregressive AIRFOIL experiment, all baselines were evaluated under identical training targets and rollout protocols.

Q1: Thank you for the thoughtful and constructive response. I appreciate the substantial additional ablation studies and the authors' willingness to acknowledge the limited contribution of the nodal residual branch. These results address much of my original concern.  My remaining question concerns the attribution of performance improvements. The L3F results suggest that the existing LiNO backbone also plays a substantial role. Could the authors clarify whether the GL comparison uses matched geometric input channels, model capacity, and training settings? The pending GL result on Airfoil-flap would also be helpful.  A clearer distinction between the independent contribution of the geometric lifting interface and that of the existing LiNO backbone would further strengthen the paper. 

Q2&Q3: Thank you for the additional comparisons with RIGNO, GINOT, and Transolver++, as well as the cross-resolution experiments. These additions meaningfully strengthen the evaluation. I also appreciate the clarification regarding the AIRFOIL rollout protocol.  I have a few remaining questions: 

  1. On DrivAer, the standard deviation of LiGNO changes from 1.92 in the original manuscript to 3.43 in the rebuttal. Could the authors clarify this discrepancy and briefly discuss the effect of using 50,000 training points for Transolver++?
  2. For cross-resolution generalization, only the 100% node-density setting lies outside the training range of 35%–70%, and the error increases from 0.98% to 1.22%. I suggest clarifying the scope of this result, particularly the distinction between density generalization and independent remeshing.  I appreciate the authors' efforts and remain open to reconsidering my assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.