Geometry-Informed Neural Operator for Large-Scale 3D PDEs

We propose the geometry-informed neural operator (GINO), a highly efficient approach to learning the solution operator of large-scale partial differential equations with varying geometries. GINO uses a signed distance function and point-cloud representations of the input shape and neural operators based on graph and Fourier architectures to learn the solution operator. The graph neural operator handles irregular grids and transforms them into and from regular latent grids on which Fourier neural operator can be efficiently applied. GINO is discretization-convergent, meaning the trained model can be applied to arbitrary discretization of the continuous domain and it converges to the continuum operator as the discretization is refined. To empirically validate the performance of our method on large-scale simulation, we generate the industry-standard aerodynamics dataset of 3D vehicle geometries with Reynolds numbers as high as five million. For this large-scale 3D fluid simulation, numerical methods are expensive to compute surface pressure. We successfully trained GINO to predict the pressure on car surfaces using only five hundred data points. The cost-accuracy experiments show a $26,000 \times$ speed-up compared to optimized GPU-based computational fluid dynamics (CFD) simulators on computing the drag coefficient. When tested on new combinations of geometries and boundary conditions (inlet velocities), GINO obtains a one-fourth reduction in error rate compared to deep neural network approaches.

Paper

References (54)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer zgWV5/10 · confidence 4/52023-07-04

Summary

This paper presents geometry-informed neural operator (GINO), solving computational fluid dynamics (CFD) problems. It combines graph neural operations (GNO) and Fourier neural operators (FNO) to adapt to irregular discretized grids. The authors have tested the model on two large-scale datasets.

Strengths

1. The authors propose combining GNO and FNO to leverage the advantages of both methods, such as analyzing local and global information and efficiently processing irregular grids. They also conduct experiments demonstrating that GINO exhibits discretization invariance over the latent grid and the input-output mesh. 2. The authors have generated two CFD datasets using various vehicle datasets. It would greatly benefit the learning physical simulation community if these datasets were made publicly available upon the paper's acceptance.

Weaknesses

1. The definition of the $\kappa$ operator in the graph operator block is unclear. It is not specified whether it measures the distance between two points or the similarity of their features. Additionally, it would be helpful to know if the $\kappa$ operator has any learnable parameters. Providing more details in the paper would make it self-contained and enable readers to understand the methodology. 2. The paper mentions the input of SDF features and surface points. It would be beneficial to clarify if these two types of data are fed in a uniform manner. For instance, are the point locations associated with their corresponding SDF values, while the surface points are assigned a value of 0? 3. The authors claim efficient graph construction in the paper. However, it appears that once the graph operator block finishes processing the input, it results in a regular grid. It would be important to address whether the implementation takes this into consideration and ensures the efficiency of the graph construction process.

Questions

See weakness.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

See weakness.

Reviewer 5VBN5/10 · confidence 5/52023-07-05

Summary

This paper introduces a novel approach for applying Fourier or other Neural operators to complex geometries by prepending a “learnable projection step” via Graph Neural Operator (GNO). Unlike previous methods that morphe complex geometries into regular domains, this approach projects (learnable) sampled nodes onto nearby regular grids, providing advantages such as discretization invariance and reduced computational overhead by sub-sampling. However, the method is limited to simple geometries and fails to account for variations in geometry and subtle geometry features. **The paper also lacks sufficient datasets and comparisons with relevant methods, such as the GNN family.** While the ideas are somewhat novel, though not groundbreaking, **the paper requires further evaluation to demonstrate its strengths and limitations.** **A borderline rejection is assigned, with reconsideration if the issues are addressed during the revision period.** ## After rebuttal - The authors improved their references, presentation, and empirical evaluation; Hence I increased my score to 5

Strengths

- Discretization invariant - Improved efficiency and scalability by sub-sampling - Improved empirical performance

Weaknesses

- Insufficient number of datasets and comparisons: - More datasets should have been included, such as cylinders and airfoils from [1]. - Comparisons with representative methods from the GNN family, such as MeshGraphNet[1], MSGNN-Grid[2], and BSMSGNN[3], should have been discussed or ideally conducted. - Among the mentioned papers [1] to [3], a crucial comparison would be with [2], which also utilizes a background grid as a helper but differs in the backbone as it did not use FNO. This would have provided a valuable benchmark for evaluation. - The overly ambitious illustrations: such as the claim in the abstract that "...(the method) can be applied to any geometry", which is not the case. A more objective approach in illustrating both the strengths and limitations would be appreciated by readers. - Lack of clarity in dataset presentation and results: - Fig.2 suggests that the shapes are overly simplistic and the pressure distributions appear uniformly smooth, indicating ease of learning. It is essential for the author to provide a more comprehensive and clear explanation, including typical examples of datasets and variations in geometry among examples. [1] Pfaff, Tobias, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter W. Battaglia. "Learning Mesh-Based Simulation with Graph Networks." Link: https://openreview.net/forum?id=roNqYL0_XP [2] Lino, Mario, Chris Cantwell, Anil A Bharath, and Stathi Fotiadis. "Simulating continuum mechanics with multi-scale graph neural networks.". Link: https://arxiv.org/abs/2106.04900 [3] Cao, Yadi, Menglei Chai, Minchen Li, and Chenfanfu Jiang. "Efficient learning of mesh-based physical simulation with bi-stride multi-scale graph neural network.". Link: https://openreview.net/forum?id=2Mbo7IEtZW

Questions

See the last point in **Weaknesses**

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The method is limited to very simple geometries due to 2 facts: 1. Only relies on the point cloud. A counter-example is porous material where the point cloud can be the same but the tunnel is very different. - In other words, although this method is discretization invariant, it is also geometry-ignorant, which is not desired. 2. Sub-sampling, if the key flow feature is determined by some subtle geometries, which is very common, this method fails again. The limitation also is reflected in the dataset, as the pressure seems to look really smooth for all cases, and all geometries are very simple. I recommend the authors objectively illustrate these facts, even at the beginning (you do not have to emphasize them; mentioning them is enough). More experiments objectively showing the limitation (maybe in the appendix) are also appreciated.

Reviewer RPer4/10 · confidence 3/52023-07-06

Summary

The authors address the task of learning to solve large-scale PDEs based on a geometry-informed neural operator. The combination of graph neural operators (GNO) and Fourier neural operators (FNO) allows the exploration of the benefits of being able to handle irregular grids and locality of operations to allow efficiency (due to GNO) as well as capturing global interactions (due to FNO), thereby overcoming the limitations of the individual approaches. In more detail, the surface (e.g. point cloud) is input to a geometry encoder that encodes the irregular grid information on a regular grid structure based on local kernel integration layers through GNO with graph operations. Then the result is concatenated with signed distance features. Then, a sequence of FNO layers is used (i.e. on latent space) for global kernel integration. The respective intermediate result is projected back to the domain of the input geometry. To show the potential of their approach, the authors carry out experiments on a novel own dataset as well as a ShapeNet car dataset, where they report quantitative results regarding training and test errors.

Strengths

Technical soundness and novelty: The method seems novel and performs beneficial over some potential alternatives in terms of speed and accuracy, since it allows a significant speed-up in comparison to the OpenFOAM solver and offers more accuracy than GNO, GeoFNO and U-Net. Evaluation: The authors provide quantitative and qualitative results including comparisons to baselines regarding training/test errors. Exposition: The paper is well-structured and mostly readable. Figure/tables and their captions are informative.

Weaknesses

Technical soundness and novelty - The discussion of limitations provided by the authors is rather short. Blending out the benefits of physics-informed approaches, that led to extreme speed-ups over OpenFOAM particularly for fluid simulation as well as offer generalization to novel scenes without being limited to object categories, limits the conclusions drawn from the presented work. The relation to these should be better emphasized to clarify what the presented method adds and whether it can be combined with these. - The datasets are described in a very short manner. Especially for the new dataset, it would be relevant to see more details such as a systematic overview on the key aspects. In addition, clarifications on whether the trainings of different approaches converged within the used 100 iterations would be relevant. In Figure 2, the rightmost part is also difficult to interpret given the used color scale. Evaluation: - Reporting training and test errors only gives limited insights on where the errors are better/worse in comparison to previous approaches. E.g., it is not demonstrated that physical phenomena (such as the Magnus effect or Karman vortex streets) are accurately represented. - What are limitations/failure cases that cannot be handled that well with the presented operator? References: Discussing other developments such as Brandstetter et al., CLIFFORD NEURAL LAYERS FOR PDE MODELING -> usage of multivector representations together with Clifford convolutions. The authors show the benefit of Clifford neural layers by replacing convolution and Fourier operations in common neural PDE surrogates by their Clifford counterparts on 2D Navier-Stokes tasks or physics-informed upgrades of U-Net such as e.g. Wandel et al., Teaching the incompressible Navier–Stokes equations to fast neural surrogate models in three dimensions -> an example of physics-informed fluid simulation based on physics-informed U-Net or the splitting of the solver into region-wise optimization such as Balu et al., Distributed Multigrid Neural Solvers on Megavoxel Domain would improve the discussion of the presented approach and its potential in the context of related work, especially since the presented approach involves quite strong assumptions (dependence on training data/shape category) and not leveraging physics-informed models therefore should be discussed.

Questions

Please discuss the comments mentioned under 'Weaknesses'.

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Limitations are shortly discussed, but seem to be severely limiting (category-specific approach, lacking generalization capabilities, unclear relation to speed-ups achieved based on physics-informed approaches).

Reviewer rW9d5/10 · confidence 4/52023-07-06

Summary

This paper proposes a framework to learn a neural operator for large-scale 3D PDEs. The framework uses a well-implemented Graph Neural Operator (GNO) to transform the irregular grid into a regular grid, so that it enables the powerful Fourier Neural Operator (FNO) to work on irregular input data, such as point clouds, in a discretization invariant way. The main contribution of the paper is to use GNO to transform large-scale irregular grids (point clouds and SDF) into regular grids, to make the FNO suitable. Based on this, GINO realized 100,000x speed-up compared to GPU-based simulators on large-scale stable CFD problems. The model is proven to be working on datasets with a high level of complexity and realism. The paper also generates two large-scale CDF datasets, which require a large amount of time to simulate and generate, which is valuable.

Strengths

1. Clear written about the problem setting and equations and symbols. 2. The main idea is useful but not too complicated to understand. An efficient combination of existing methods. 3. The method has a general potential for various kinds of PDEs. 4. The method has shown great performance in engineering-level experiments, fulfilling its great potential for applications.

Weaknesses

Although the experiments have demonstrated the main ability of the model, the experiments are not complete enough to support all claims and novelties. 1. Since the model is a new combination of two existing methods GNO and FNO, the main contribution is the new usage of GNO. Then the main thing required to be demonstrated should be “GNO is more proper and has great encoding and decoding ability between irregular and regular grids”. Regarding this, some different encoding methods should be compared, such as GNN, kNN. 2. Discretization invariance is said to be important, but no direct experiments could support this. Although Table 5 is relevant to this, but the variables are not kept so it’s not a direct support. In other words, a GNN+FNO baseline should be added.

Questions

1. What is the difference between your proposed GNO and Continuous Convolution (CtsConv) from “Lagrangian Fluid Simulation with Continuous Convolutions”? It seems there is no obvious difference based on equations. And one of the claimed contributions is a well-implemented GNO, which makes it much more efficient, but CtsConv is also well-implemented based on Open3D and hashing table. 2. What is the performance for much finer latent resolutions such as 128^3 and 256^3? Can we keep increasing the test error by increasing the latent resolution?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

2 fair

Contribution

3 good

Limitations

1. As the paper stated, the trained model is limited to a specific observed category of shapes. The training for the operator requires a training dataset of high quality. 2. The proposed framework should be general for various kinds of PDEs, but only stable NS equations are tested. 3. The framework is for 3D PDEs, which could be modified for time-dependent ones.

Reviewer zqvK5/10 · confidence 3/52023-07-08

Summary

This paper proposes a geometry-informed neural operator for arbitrary geometry, to facilitate the learning on the solution operator for large-scale 3D CFD simulation.

Strengths

The proposed GINO model applies graph-kernel blocks for the encoder and decoder, for processing features in the latent uniform space, with the Fourier blocks running on the latent space to capture the global interaction. The proposed method provides significant runtime speedup.

Weaknesses

For the experiment, the authors only perform the evaluation on the car category of ShapeNet dataset. It’ll be more persuasive to have the proposed model evaluate the other categories rather than only a single category.

Questions

As the authors claim the significant speedup provided by the proposed method. Could the authors report the scale of the parameters of the proposed method, and the comparison to the other related methods?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

As pointed out by the authors, this work is constrained to a specific category and limited to CFD with more complex shapes. It'll be with practical significance if the proposed work could tackle these limitations.

Reviewer 5VBN2023-08-10

Increase my score to 5 after reading rebuttal

Thanks the authors resolved partial of my concerns. I would hence increase my score to 5. I apologize for not being able to edit the original comments (maybe because of passing due).

Reviewer zgWV2023-08-11

Read the rebuttal, maintain the same rating.

Authorsrebuttal2023-08-14

Looking forward to your feedbacks

We are happy to note that our responses have addressed the concerns of several reviewers. As the discussion deadline nears, we kindly urge the remaining reviewers to share their feedback. Your insights and comments are invaluable in enhancing the quality of our paper. Please inform us if there are any further concerns. We believe that there's an agreement on the complexity and potential of 3D aerodynamics simulation. Our approach combined the flexibility of graphs with the efficacy of the spectral method, paving the way for scalable, complex 3D simulations. And we hope our dataset will lead to exciting future research in the community. If our rebuttal has partially resolved your concerns, we kindly request you to adjust the score accordingly. Your understanding and contributions are greatly appreciated.

Reviewer RPer2023-08-15

Follow-up on Rebuttal

I thank the authors for discussing my concerns. The following aspects remain vague: 1) The authors mention to add a more systematic description of the dataset, however, no further details are provided. 2) The authors mention to add a drag coefficient study in the revision, however, no further details are provided. 3) The benefits over the mentioned references remain unclarified. More details on this would clarify the contribution of the submission.

Authorsrebuttal2023-08-16

Thank you to the reviewer for the response. It appears there has been a misunderstanding. The rebuttal does not permit the submission of revisions. Therefore, we have included our new experiments and updates in the general response and the 1-page supplemental pdf. These can be found at the **General response** prompt near the top of the webpage. We will also elaborate on them here: ## 1. The description of the dataset is in Section 3 of the pdf file. We further add the illustrations of the Ahmed-body dataset. The figure on the left shows the velocity field and the pressure field. The velocity field, represented with 7 million nodes, has complex vortexes at the rear of the body. The pressure field, represented with 100 thousand nodes, is steep at the front and also the legs. Such aerodynamic simulations are extremely expensive. Each simulation takes 7-19 hours on 2 Nvidia v100 GPUs with 16 CPU cores. It is extremely costly to generate a 3D dataset with multiple shapes. We continue to generate simulations on new shapes and increase the instances from 500 to 800. Industry-standard Ahmed-body geometries are characterized by six design parameters: length, width, height, ground clearance, slant angle, and fillet radius. Refer to the wiki (https://www.cfd-online.com/Wiki/Ahmed_body) for details on Ahmed body geometry. In addition to these design parameters, we include the inlet velocity to address a wide variation in Reynolds number. We identify the design points using the Latin hypercube sampling scheme for space filling design of experiments and generate around 800 design points. The aerodynamic simulations were performed using the GPU-accelerated OpenFOAM solver for steady-state analysis, applying the SST K-omega turbulence model. These simulations consist of 7.2 million mesh points on average, but we use the surface mesh as the input to training which is roughly around 70-100k mesh nodes. ## 2. The drag coefficient study is in Section 2 of the pdf file and the general response. To compare the performance of our model against the industry-standard OpenFOAM solver, we perform a full cost-accuracy trade-off analysis. The result shows GINO is 26,000x faster at computing the drag coefficients. Figure [1] below shows the cost-accuracy curve, measured in terms of inference time needed for a relative error in the drag coefficient for GINO and OpenFOAM. The cost of GINO is computed as the time, averaged over the test set, needed to predict the drag coefficient by running the model. This time includes both data pre-processing (computing the SDF) as well as the model run-time and the drag calculation given the predicted fields. All models are ran on a single NVIDIA V100 GPU. The cost for OpenFOAM is computed as described in the next paragraph and is averaged over the test set. The solver is ran on two NVIDIA V100 GPUs in parallel. We observe a four to five order of magnitude speed-up when using GINO. At a $3\%$ relative error, we find the speed-up from our model which includes drag in the loss to be $26,000 \times$. As we increase the size of the latent space, the cost of GINO grows, however, we observe a plateau in the drag error. This is common in machine learning models as the error from using finite data starts to dominate the approximation error. Furthermore, we use only the size of the latent space as a hyper-parameter, keeping the number of learnable parameters fixed. It is interesting to explore further how parametrically scaling the model impacts predictive power. During data generation, we keep track of the drag coefficient predicted by OpenFOAM after every iteration. While the coefficient converges with more iterations, this convergence is not monotone and can often appear quite noisy. This makes computing the error from the raw data not possible. We therefore apply a box filter to the raw signal to compute a filtered version of the drag which acts as a smoother. We take as the reference drag, the drag at the last iteration of the filtered signal. To compute the number of iterations it takes for the solver to predict a drag coefficient at a given relative error, we trace back the predictions from the filtered signal and return the first time at which this prediction incurs the given error with respect to the reference drag. An example of this methodology is shown in Figure [2]. The errors for our GINO model are computed with respect to the true drag coefficient from the last iteration of the solver. This is because we take as ground truth the pressure and wall shear stress from this last iteration and train our model to predict them.

Authorsrebuttal2023-08-16

## 3. Here we add the discussion of the references > References: Discussing other developments such as > - Brandstetter et al., CLIFFORD NEURAL LAYERS FOR PDE MODELING -> usage of multivector representations together with Clifford convolutions. The authors show the benefit of Clifford neural layers by replacing convolution and Fourier operations in common neural PDE surrogates by their Clifford counterparts on 2D Navier-Stokes tasks The Clifford neural layers [1] use the Clifford algebra to compute multivectors, which improves the effiency for simulations involving multiple fields, and they provide Clifford-FNO implementations as an extension of FNO. While in this work, we only considered the pressure field so Clifford may not be helpful, it will be interesting to explore using Clifford-FNO within our GINO model to address complex 3d geometry when modeling more fields (e.g. velocity field and stress tensor field) in the future. > or physics-informed upgrades of U-Net such as e.g. > > - Wandel et al., Teaching the incompressible Navier–Stokes equations to fast neural surrogate models in three dimensions -> an example of physics-informed fluid simulation based on physics-informed U-Net The work [2] uses a physics-informed loss with U-Net for 3D fluid simulation. The researchers consider channel flow in three geometries: box, ball, and cylinder, in a resolution of 128x64x64 (0.5M) voxels. The Reynolds numbers studied range from 0.64 to 800. They also test the model on the shapes of a fish and three boxes. The model employs finite-difference schemes (eq 6-9) to compute the physics-informed loss. The model has a great performance on these 3D simulations. In contrast, in our work, we consider a higher Reynolds number regime ranging from 4.35 × 10^5 to 6.82 × 10^6 generated on 7.2M mesh points. We think this scenario is more realistic and intriguing. The physics-informed setting is a promising direction we aim to explore further. However, the finite-difference scheme in [2] introduces truncation errors, which can be problematic in turbulent regimes. Thus, we plan to explore physics-informed neural operators that compute the exact derivatives from the Fourier space. Further, the neural operators are discretization-invariant, meaning they can be training with lower data resolution and fine-tune with the physics-informed loss at higher resolutions, which is an efficient way to combine data and physics. > or the splitting of the solver into region-wise optimization such as > > - Balu et al., Distributed Multigrid Neural Solvers on Megavoxel Domain The work [3] innovatively proposes the use of multigrid training for neural networks. Similar to the multigrid iteration, the methods explore V, W, and F cycles. It is shown that the deep learning model enjoys a significant speedup. In numerical analysis, multigrid methods are introduced to aid the slow convergence of low-frequency errors, given that the standard Jacobi iteration converges faster for high-frequency errors. On the other hand, data-driven neural network models appear to converge faster on low-frequency modes, which possess higher energy. It would be intriguing to conduct a spectral analysis and study the convergence of frequency. [1] Brandstetter, Johannes, et al. "Clifford neural layers for PDE modeling." arXiv preprint arXiv:2209.04934 (2022). [2] Wandel, Nils, Michael Weinmann, and Reinhard Klein. "Teaching the incompressible Navier–Stokes equations to fast neural surrogate models in three dimensions." Physics of Fluids 33.4 (2021). [3] Balu, Aditya, et al. "Distributed multigrid neural solvers on megavoxel domains." Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. 2021.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC