acceptodds
Under review as a conference paper at ICLR 2027

GRIFDIR: Graph Resolution-Invariant Diffusion Models over Irregular Domains

Abstract

Score-based diffusion models in infinite-dimensional function spaces provide a mathematically principled framework for modelling function-valued data, offering key advantages such as resolution invariance and the ability to handle irregular discretisations. However, practical implementations have struggled to fully realise these benefits. Existing backbones like Fourier neural operators are often biased towards regular grids and fail to generalise to complex domain topologies. We introduce an architecture, GRIFDIR, for function-space diffusion models that represents generalised graph convolutional kernels as finite element functions, allowing the score network to operate directly on unstructured meshes over domains of arbitrary shape. We demonstrate the efficacy of our network architecture through a series of unconditional and conditional sampling experiments across diverse geometries, including non-convex and multiply-connected domains. Our results show that the proposed method maintains resolution invariance and achieves high fidelity in capturing functional distributions on non-trivial geometries.

Primary area:
generative models
Keywords:
Graph neural operators, infinite-dimensional diffusion model, resolution invariance

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_Pvk6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduced a multiscale GNO for score-based diffusion models on an infinite-dimensional function space. It uses a resolution-invariant convolution kernel to build a bilinear finite-element function on a fixed reference patch. For a given patch, it projects its neighbors on the mesh onto its coordinate frame. It uses a hierarchical mesh architecture with a latent diffusion transformer. Both unconditional and conditional generation are evaluated in this paper on four benchmark datasets across different metrics.

Critical strengths
  1. The core methodology (FEM-Conv) of this paper has a rigorous theoretical guarantee. The propositions in the paper demonstrate that its construction is a generalization of the standard convolution and can converge under mesh refinement, with a guaranteed computational complexity of O(4|N(i)|) operations per node at any filter size.

  2. The method proposed by the paper achieves the best SWD and MMD across all tested resolutions, including those that are OOD, as shown in Table 1. This is strong evidence of the effectiveness of this resolution-invariant representation,

Critical weaknesses
  1. The method only clearly wins on the synthetic uniform mesh (the GaussianBlobs data). On non-uniform or externally sourced datasets, the method is only second-best on unconditional MMD/SWD. Also, for the DPS conditional-sampling metrics, the method fails 3/4. On the big picture, this is concerning, because the clearest win only comes from the regime where discretization invariance matters the least.

  2. On DeepJEB, there is a huge performance gap between the proposed method and the supervised baselines. While I acknowledge the distinction between the generative paradigm and the supervised paradigm, I am still not convinced why the generative paradigm is more ideal, since the task already has a single ground-truth field for each bracket.

Questions for authors
  1. Even though MINO is using mean aggregation, why does it still outperform GRIFDIR on non-uniform meshes? It is hypothesized in the paper that this mean aggregation strategy would hurt MINO's performance on non-uniform Cylinder Flow meshes, but the results on the non-uniform GaussianBlobs test say otherwise. Does the lumped quadrature weight in GRIFDIR introduce another type of bias on a graded mesh? I am trying to understand under what circumstances the quadrature-based design in this work would actually help.

  2. Please provide a brief sensitivity analysis on the model with regard to the covariance operator C and the noise schedule. Even though this is marked as future work in the manuscript, having those would help understand the robustness of the current hyperparameter choice.

4 comments

Thanks a lot for your review! We're glad you find our core methodology to have rigorous theoretical guarantees, and find the evidence of of the resolution invariant representation to be strong. We address your critical weaknesses and questions below.

Weakness 1 and Questions 1 (Quadrature and Performance): Thanks for the question. You're right, our quadrature computation was slightly incorrect leading to a bias of sorts. Please refer to our General Response for details. After updating this, we see that the quadrature based design indeed significantly improves results on the graded meshes over MINO. The fix also substantially improves transfer to unseen discretisations on Pinball and the unconditional metrics on CylinderFlow.

Weakness 2 (DeepJEB). We would like to comment generative models are an established tool in scientific machine learning for tasks such as full-state reconstruction from sparse measurements, and this is the setting GRIFDIR targets. We agree with the reviewer that for predicting a single stress field, a supervised operator with a specialised architecture such as GINOT works best. DeepJEB is also a difficult target: the stress field changes sign sharply and locally, which a pointwise error punishes heavily. We have nonetheless improved our results, with the fixed quadrature and Fourier positional encodings (e.g. 0.645 →\to 0.587 at 6k nodes). We also now report the average of four samples for each bracket and sparse-sensor reconstruction with DPS (e.g. 0.615 →\to 0.470 with 10% of the nodes observed at 24K nodes). Both improve the prediction, and the second uses measurements without any retraining, which the supervised baselines here are not built to do. However, the main aim of the DeepJeb experiment was to show that the method scales to 3D, and we keep it in the paper for that reason.

DeepJEBnodesrel-L2L^2 ↓w-rel-L2L^2 ↓IoU@1% ↑IoU@5% ↑IoU@10% ↑r ↑rel-L$^2$, mean of 4 ↓no sensors ↓1% sensors ↓5% sensors ↓10% sensors ↓
GINOT (supervised)full0.308—————0.308————
GRIFDIR (old)6k0.6450.5750.5230.4790.5450.7640.508————
GRIFDIR (old)24k0.6380.5640.5100.4830.5410.7710.510————
GRIFDIR (fixed)6k0.5870.4730.6080.5340.5810.8070.4630.5890.5760.5420.524
GRIFDIR (fixed)24k0.6150.5050.5880.5080.5600.7920.4900.6180.5820.5230.470
GRIFDIR (fixed)full0.6330.5120.5840.4930.5430.7860.5090.6380.6090.5150.465

Question 2 (Ablations for covariance and noise scheduling): We trained the Pinball model with the covariance length scale ℓ=0.01,0.05,0.1,0.2,0.5\ell=0.01, 0.05, 0.1, 0.2, 0.5. The default of the paper was ℓ=0.1\ell = 0.1, following the setup in [1]. We see that for the Pinball dataset actually ℓ=0.2\ell = 0.2 gives slightly better results. We have included these results in the revised paper in Section F.2 Ablations.

MMDℓ\ell = 0.010.050.1 (paper)0.20.5
GRIFDIR0.1230.1130.0940.0600.062

For the noise schedule, we use the standard variance-exploding schedule of Karras et al. [2]. We agree that tuning the schedule is interesting, and optimal schedules for diffusion models are still an open research question, so further gains are likely. However, the schedule is a property of the forward SDE and sampler, not of the score network. Our contribution is the backbone: a score model architecture that makes the theoretical advantages of infinite-dimensional diffusion usable in practice. We therefore keep the schedule fixed, and identical across all baselines, so that differences in performance reflect the architecture alone.

[1] Yao et al. "Guided Diffusion Sampling on Function Spaces with Applications to PDEs" (2025)

AuthorsOpenReview ↗

Final quick summary

Hi Reviewer Pvk6, thanks again for the review. Here's a quick summary of our rebuttal, we'd really appreciate a quick reply before the discussion closes on the 18th.

You asked why MINO beat us on non-uniform meshes: your question led us to a bug in our quadrature normalisation. With it fixed, GRIFDIR gets 0.079 mean MMD on the graded meshes against 0.116 for MINO, over five training runs each (Tables 1 and 10). We feel this supports our argument that proper quadrature is a necessary aspect for mesh transfer and has led to improved performance in our model.

Thank you for your careful rebuttal, which has addressed most of my concerns. I am happy to raise my score.

Thank you for the discussion and for raising your score. Your questions led to changes that improved our paper, and we appreciate the review.

Reviewer_vL1XOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FEM-Conv, a convolution operator for score networks in infinite-dimensional diffusion models that acts directly on unstructured meshes. The filter is a finite-element function on a fixed physical patch, and neighbours are aggregated with mass-lumped quadrature weights, so the operator does not depend on the discretisation. FEM-Conv is embedded in a multiscale mesh hierarchy with a latent transformer at the coarsest level. Validation covers resolution and mesh-graded transfer on Gaussian Blobs, eight domain shapes, Pinball, Cylinder Flow and 3D DeepJEB, together with conditional sampling for inverse problems.

Critical strengths

S1. A diffusion backbone that operates on unstructured meshes and transfers across resolutions without retraining, demonstrated on four 2D/3D datasets. S2. A clean formalisation of the operator (Props. C.1–C.2) with component, kernel and radius ablations (Tables 10–14).

Critical weaknesses

W1. The authors acknowledge (F.3) that with mean aggregation FEM-Conv matches first-degree SplineConv, leaving only the fixed physical patch and the lumped weights. Both have precedents that are neither cited nor discussed: KPConv (radii scaled with cell size) and quadrature/density-weighted aggregation (MC-Conv, PointConv, Liu-Schiaffini et al. 2024). The gain over known components is therefore hard to identify. W2. On the graded meshes the method is designed for, GRIFDIR trails MINO (mean MMD/SWD 0.160/0.057 vs 0.128/0.046, Table 1) and wins on only one of four meshes (Table 8), though lumped aggregation targets exactly this setting (p.5).

Questions for authors

Q1. At 1152 points the default configuration reports MMD 0.063 (Table 1), 0.064 (Table 10), 0.0585 (Table 11) and 0.163 (Table 13). Which is the main result, and do these differ by run or seed? Q2. Does FEM-Conv differ from the kernel representations in KPConv, MC-Conv and Liu-Schiaffini et al. (2024) beyond the fixed patch and the quadrature weights? If so, what would demonstrate this empirically?

4 comments

Thank you for the constructive review and helpful feedback. We address the weaknesses and questions below.

Weakness 1 (Gain over known components): Please see the General Response for details. In short, our contribution is not a new convolutional layer per se, but the combination of these components into a diffusion modelling framework that operates directly on irregular domains. The revision now explicitly states the connection of FEM-Conv to prior work (Remark 3.2). In particular, FEM-Conv can be seen as a DISCO convolution.

Weakness 2 (Performance on graded meshes): You are right that the lumped aggregation targets exactly this setting. However, our implementation computed the normalisation incorrectly, which introduced a bias on non-uniform meshes (see General Response). After the fix, GRIFDIR outperforms MINO on three of the four graded meshes and on average (MMD/SWD 0.079/0.031 vs 0.128/0.045).

Question 1 (reported numbers): Thanks for pointing this out. The model used in Table 1 and Table 10 is the same, we just rounded differently, so this was an error on our side. Both the models in Table 11 (radius ablation) and in Table 13 (Kernel ablation) were different runs. The performance of the model in Table 12 was significantly worse than the model in Table 1/Table 10 and Table 11. To analyse the variability of the training we have now trained GRIFDIR on the GaussianBlobs dataset for 5 different training runs. Below, we show the MMD and SWD with the mean (min, max) across the runs. This is now also included in the paper in section F.2 Ablations.

GaussianBlobs1,152 pts. (coarse)2,048 pts. (train res.)8,192 pts.
GRIFDIR (fixed), MMD ↓0.071 (0.043, 0.097)0.078 (0.074, 0.083)0.093 (0.035, 0.122)
GRIFDIR (fixed), SWD ↓0.027 (0.020, 0.035)0.030 (0.029, 0.032)0.036 (0.019, 0.045)

Question 2 (difference from KPConv, MC-Conv and Liu-Schiaffini et al.): The kernel parametrisations are closely related. The linear correlation around kernel points in KPConv and the hat-function basis in Liu-Schiaffini et al., result in similar kernels as in our FEMConv. The main difference lies in the aggregation. KPConv sums over neighbours without accounting for sampling density, so its output changes with mesh resolution. In MC-Conv the authors use a random subset of nodes in the neighbourhood to implement the convolutional layer, whereas we take all neighbours into account.

Thanks for the detailed response and the additional experiments.

The corrections partly address my concerns. In particular, after fixing the normalisation of the lumped aggregation, GRIFDIR now outperforms MINO on the graded meshes, which resolves W2.

However, the fact that the core component of the paper was implemented incorrectly in the original submission makes the reported conclusions appear less rigorous. It also remains unclear whether all results that depend on the lumped aggregation have been re-run with the corrected implementation.

In addition, the rebuttal acknowledges that the kernel parametrisation of FEM-Conv is closely related to KPConv and to the DISCO-type kernels of Liu-Schiaffini et al., with the main difference being the aggregation. This tones down the methodological contribution claimed in the paper.

As this submission is outside my main area of expertise, I keep a reserved position and will defer to the more expert reviewers in the final discussion.

Thank you for the follow-up and for acknowledging that W2 is resolved.

Re-run experiments: All experiments that depend on the lumped aggregation have been re-run with the corrected implementation, and the revised manuscript reports only these results. To quantify the training variability, we additionally trained GRIFDIR on the Gaussian Blob dataset with 5 independent runs (approx. 24 h each) and report mean, min and max.

Implementation error: the error we uncovered in the rebuttal was a refinement of the original implementation. The original implementation was designed to correctly weight quadrature inside the domain, the refinement was a more careful management of the weighting contribution of cells on the boundary of the domain. After the fix GRIFDIR not only outperforms MINO on the regular mesh but also the graded meshes and on average (MMD/SWD 0.079/0.031 vs 0.128/0.045). The conclusions of the paper are unchanged.

Methodological contribution: We agree that the kernel parametrisation is closely related to KPConv and to DISCO-type kernels, and Remark 3.2 now also states this in the main paper, note DISCO convolutions do not work on non-parametric meshes as they rely on the analytic form for the quadrature on regular shapes. However, to your comment "With the main difference being the aggregation", we feel this is essential, as our previous and now improved results show. As we said in the General Response, our main contribution is a score backbone for function-space diffusion models that operates directly on unstructured meshes:

  1. the score network is a neural operator between function spaces
  2. its layers are discretisations of an integral operator, so the model can be trained on one mesh and sampled on another
  3. it supports conditional sampling (e.g. DPS, DAPS) for PDE inverse problems on these meshes.

The specific kernel parametrisation is a necessary component for this goal.

Please notice the new results and edited text are already loaded in the new 10 page submitted pdf.

We appreciate your response and hope these clarifications address the remaining concerns.

Thank you for your prompt response and clarifications, which partially address my concerns. I will maintain my original rating of Weak Accept, as my limited expertise in this area does not allow me to confidently reassess the remaining issues.

Reviewer_Dw8pOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes GRIFDIR, a multi scale graph neural operator that uses infinite dimensional diffusion on irregular geometries. The main layer is a resolution-invariant FEM convolution, where the convolutional filter is a weighted sum of a bilinear basis on a fixed reference patch. The multiscale is done using a V-cycle U-Net with a Perciever style latent transformer at the bottleneck. The authors prove that FEM-Conv is a standard CNN on uniform grids and converges to the true integral as the mesh is refined. The authors compare their method against some baselines over several resolutions and discretizations on GaussianBlobs, Pinball, and CylinderFlow.

Critical strengths

The main strength of this paper is that its motivation and the reason to use infinite-dimensional diffusion seems justified. The listed desiderata in particular makes it clear that a multi-scale operator approach is well suited. The writing and presentation is also clear.

Critical weaknesses

The main weakness of this paper is its limited novelty. I think the FEM-Conv with a Q1 kernel on a regular grid reduces to basically the same approach as SplineConv. The quadrature weighting was also done previously in QuadConv (the current version only cites QuadConv for its MLP filter). MINO also has a similar bottlneck design, where it uses a radius kernel encoder onto latent tokens and a cross attention decoder. Finally, there are very similar prior works. ∞\infty-Diff (which was not cited) also combines a diffusion model in an infinite-dimensional Hilbert space with integral operators on irregular coordinates and multiscale processing. DISCO also uses a locally supported integral kernel expanded in hat functions and discretized with quadrature weights, and Liu-Schiaffinit et al. used this as a discretization agnostic neural operator on general meshes.

References:

  • Bond-Taylor and Willcocks. "∞-diff: Infinite resolution diffusion with subsampled mollified states", ICLR 2024
  • Liu-Schiaffini et al. "Neural operators with localized integral and differential kernels", ICML 2024
Questions for authors
  1. For the 1152 and 8192 element and graded-mesh evaluations, was rlr^{l} is fixed from the training mesh and could you explain how the coarser hierarchy levels were obtained? Were they rebuilt with the same coarsening ratio, reused from training, or regenerated at some fixed value?
4 comments

Thank you for the review. We first address novelty and then answer the question about the mesh hierarchy. Please also see the General Response for a broader discussion of our contribution and its relation to prior work.

Novelty We agree that the individual building blocks have precedents, and the revision now cites and discusses each of them. We now introduce FEM-Conv as a special case of a DISCO convolution and of the localised integral kernel layers of Liu-Schiaffini et al. (Remark 3.2). Regarding the methods mentioned in the review:

  • SplineConv with degree-one splines and PP basis functions per dimension uses the same kernel space as FEM-Conv. We previously noted this in the appendix and have now moved it to the main text. However, SplineConv was designed for GNNs and uses mean aggregation, which is not discretisation-invariant.
  • QuadConv uses an MLP kernel with quadrature weights at the nodes, and we now cite it for the quadrature as well as the filter.
  • MINO uses a kernel integral layer to map the input to latent grid of fixed size (mostly 32×3232\times32). In a similar way ∞\infty-Diff also uses a first integral layer to map the irregular input to a regular grid (128×128128 \times 128) and then use a U-Net model. Instead in GRIFDIR the lowest resolution is not a latent grid, but really a coarse representation of the input mesh, which still includes the structure (e.g., holes).

In the revised manuscript we now cite these methods and discuss the relationship in detail. However, we want to stress that none of these methods has the specific combination we do.

Question (Mesh creation) For the uniform meshes, including the 1152- and 8192-element meshes, the hierarchy is regenerated for each mesh by halving the resolution in each dimension, i.e. [(nx, ny), (nx//2, ny//2), (nx//4, ny//4), (nx//8, ny//8)], so the coarsening ratio between consecutive levels is always the same. The graded meshes have the same node budget as the training mesh and use the same hierarchy (2,048 →\to 512 →\to 128 →\to 32). We first generate the coarsest mesh and then obtain the finer levels by refinement with gmsh (mesh.refine). Details on mesh generation are now given in Appendix E.1.1. The convolution radius is defined in physcial coordinates and fixed across discretisations.

AuthorsOpenReview ↗

Final quick summary

Hi Reviewer Dw8p, thanks again for the review. Here's a quick summary of our rebuttal, we'd really appreciate a quick reply before the discussion closes on the 18th.

On novelty, you were right that the building blocks have precedents, and the revision now makes the relation explicit (Remark 3.2): FEM-Conv, DISCO, QuadConv and SplineConv are all discretised local integral operators, differing in kernel and aggregation weights. Where we differ is that we work on unstructured meshes, where the quadrature has to be built from the mesh geometry, and we combine this into a function-space diffusion backbone that transfers across discretisations and supports conditional sampling. Your mesh-hierarchy question is answered in Appendix E.1.1.

I thank the authors for their response. The additional information for the convolution radius cleared up my question on how the coarser hierarchy levels were created. However, I think the framing that FEM-Conv is now presented as a special case of DISCO does not help with the limited novelty. DISCO layers already work on general meshes and so I'm not sure that GRIFDIR differentiates itself enough by working on unstructured meshes. On remark 3 comparing MINO and ∞\infty-Diff, GRIFDIR also seems to maps the coarsest mesh to a set of latent tokens shared across domains due to cross-attention (Sec 4.3 and App. D), so the bottleneck is not mesh-structured. Due to this, I will maintain my score.

Thank you for the follow-up. We address your remaining points in case this is useful for the discussion period.

You are right that Liu-Schiaffini et al. (2024) extend the DISCO convolutions of Ocampo et al. (2023) to unstructured meshes, and FEM-Conv belongs to the same family. In our updated paper we cite them clearly in Remark 3.2.

Our main contribution is on the generative side. To our knowledge, no prior work trains a function-space diffusion model on unstructured meshes, samples it on a different mesh, and uses it for conditional sampling (DPS, DAPS) on PDE inverse problems. We introduced the mesh-based convolution as a required component of this; it is not our central claim.

Two clarifications on the architecture:

  • Where the mesh is used. Liu-Schiaffini et al. use one level mesh only on the encoder and decoder, with FNO blocks on a regular grid in between (their App. C.5). GRIFDIR applies the integral operator on the mesh at every level.

  • The bottleneck. The transformer acts directly on the coarsest mesh nodes, with the mesh coordinates as positional encoding. The shared latent tokens are an add-on for the multi-domain task, where the shape of the domain changes and not just the mesh.

We hope this places our contribution more accurately, and thank you for the discussion.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank the reviewers for their careful reading and constructive feedback. Two concerns came up in several reviews. The first is novely relative to exisisting convolution and neural operator layers (Dw8p, vL1X). The second is the performance on graded meshes (vL1X, Pvk6). We address both here and give further details in the individual responses.

Contribution and Novelty We want to clarify what we see as the core contribution of our work. Our main contribution is to develop a score backbone for diffusion models in function spaces that operate directly on unstructured meshes, with the ingredients:

  1. the score network is a neural operator between function spaces
  2. its layers are discretisations of an integral operator, so the model can be trained on one mesh and sampled on another
  3. it supports conditional sampling (e.g. DPS, DAPS) for PDE inverse problems on these meshes

For 2) in particular, an accurate quadrature of the discretised integral operator is essential. For example, MINO and SplineConv use mean aggregation, which does not converge to the continuous operator under irregular discretisations. While FEM-Conv can be interpreted as a special case of a DISCO convolution, extending it to our mesh-based setting is not straightforward. DISCO focuses on group convolutions on the sphere, where exact quadrature rules are available, whereas on unstructured meshes the quadrature must be constructed from the mesh geometry. The normalisation issue described below shows how much this matters in practice. We added Remark 3.2 to the revised paper, where we discuss the relationship to SplineConv and DISCO.

To our knowledge, no prior method combines all of these. Most existing function-space diffusion models for PDEs operate on regular grids. The closest work, MINO, operates on unstructured meshes but is unconditional.

MethodFunction SpaceHandles Unstructured GridConditional Sampling
DiffusionPDE [1]✗✗✓ (DPS-like)
FunDPS [2]✓✗✓ (DPS-like)
Fun-DDPS [3]✓✗✓ (DPS-like)
DDIS [4]✓✗✓ (DAPS-like)
FunDiff [5]✓✗✓ (cond. model)
∞\infty-Diff [6]✓✓✓ (cond. model)
MINO [7]✓ (Functional Flow Matching)✓✗
GRIFDIR✓✓✓ (DPS, DAPS, cond. model)

Note that ∞\infty-Diff does only perform experiments on regular grids.

[1] Huang et al. "DiffusionPDE: Generative pde-solving under partial observation" (2024)

[2] Yao et al. "Guided diffusion sampling on function spaces with applications to PDEs" (2026)

[3] Ju et al. "Function-Space Decoupled Diffusion for Forward and Inverse Modeling in Carbon Capture and Storage" (2026)

[4] Lin et al. "Decoupled diffusion sampling for inverse problems on function spaces" (2026)

[5] Wang et al. "FunDiff: Diffusion Models over Function Spaces for Physics-Informed Generative Modeling" (2025)

[6] Bond-Taylor et al. "∞\infty-Diff: Infinite Resolution Diffusion with Subsampled Mollified States" (2023)

[7] Shi et al. "Mesh-Informed Neural Operator : A Transformer Generative Approach" (2025)

Correction to Quadratrure

Prompted by the reviewers questions about the performance on the graded mesh transfer experiment, we looked again at and managed to uncovered an error in our implementation. Our claim was that correctly calculating quadrature was nessasary for mesh transfer. However in our previous implementation of the FEM convolution, each neighbouring cell was counted either fully inside or fully outside the convolution support, depending on where its centre fell. Instead, the cells should be weighted by the fraction of its area that actually lies inside the ball.

This error depends on the size and shape of the cells. On the uniform training mesh, where cells have a similar shape, the normalisation error is nearly a uniform "bias". However, on the graded mesh, where the cells are shaped differently, it lead to an 11−16%11-16\% error in the kernel integral.

Fix: We replace the cell count with an exact quadrature, where each cell is weighted by the area of its part inside the support. These normalisation weights can be computed once per mesh. This adds no new trainable parameters or hyperparameters, and the architecture and training configuration are unchanged. It did, however, require re-training the GRIFDIR models. This has resulted in the following improvement to our results:

GaussianBlob dataset: we now train all the models for five different seeds and now show the mean performance and standard deviation over training runs. With the wrong normalisation, performance on the graded meshes varied considerably, from an MMD of 0.231 on fine left to 0.062 on fine edge. With the corrected normalisation, performance is consistent across all meshes.

Gaussian Blobs, MMD ↓1,1522,0488,192fine leftfine rightfine centrefine edgegraded avg
MINO0.111 ± 0.0130.100 ± 0.0070.101 ± 0.0080.122 ± 0.0100.122 ± 0.0110.056 ± 0.0040.163 ± 0.0150.116 ± 0.009
GRIFDIR (old)0.0630.0890.0530.2310.2390.1100.0620.160
GRIFDIR (fixed)0.071 ± 0.0210.078 ± 0.0040.093 ± 0.0380.070 ± 0.0160.082 ± 0.0130.083 ± 0.0210.083 ± 0.0180.079 ± 0.016
Gaussian Blobs, SWD ↓1,1522,0488,192fine leftfine rightfine centrefine edgegraded avg
MINO0.042 ± 0.0040.039 ± 0.0020.037 ± 0.0030.044 ± 0.0030.045 ± 0.0030.024 ± 0.0020.058 ± 0.0050.043 ± 0.002
GRIFDIR (old)0.0280.0320.0250.0790.0800.0440.0260.057
GRIFDIR (fixed)0.027 ± 0.0060.030 ± 0.0010.036 ± 0.0110.028 ± 0.0050.031 ± 0.0040.034 ± 0.0070.030 ± 0.0050.031 ± 0.005

Multi-domain blob samples - we have replaced Figure 5 in the paper, we found with the improved quadrature the samples are move faithful to the elipsoid shapes of the distribution.

Pinball: We also added resolution-transfer experiments on the Pinball dataset, using meshes with 7525 and 1934 nodes. Here too, the fixed model transfers much better to new resolutions than the old model and performs close to MINO.

PinballMMD ↓\downarrowSWD ↓\downarrowDPS-MSE ↓\downarrowDPS-ES ↓\downarrowMMD ↓\downarrow (7525 / 1934 / 226)MMD ↓\downarrow (1934 / 557 / 140)
MINO0.0840.0920.06414.260.0840.080
GRIFDIR (old)0.0910.0950.05511.530.2090.376
GRIFDIR (fixed)0.0940.0960.05311.470.0940.089

CylinderFlow: we additionally experimented with the distance to the nearest domain boundary (wall signal) as an extra input feature, which further improves the results. However, the normalisation fix alone already yields a clear improvement (MMD 0.0197 →\to 0.0096).

CylinderFlowMMD ↓SWD ↓DPS-RMSE ↓DPS-ES ↓
MP-GNN0.01870.01150.01160.4246
MINO0.03070.01650.01040.4429
GRIFDIR (old)0.01970.01170.01110.3788
GRIFDIR (fixed), without wall signal0.00960.00770.01130.3809
GRIFDIR (fixed), with wall signal0.00610.00650.01120.3899

All of these results are updated in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.