Domain Agnostic Fourier Neural Operators

Fourier neural operators (FNOs) can learn highly nonlinear mappings between function spaces, and have recently become a popular tool for learning responses of complex physical systems. However, to achieve good accuracy and efficiency, FNOs rely on the Fast Fourier transform (FFT), which is restricted to modeling problems on rectangular domains. To lift such a restriction and permit FFT on irregular geometries as well as topology changes, we introduce domain agnostic Fourier neural operator (DAFNO), a novel neural operator architecture for learning surrogates with irregular geometries and evolving domains. The key idea is to incorporate a smoothed characteristic function in the integral layer architecture of FNOs, and leverage FFT to achieve rapid computations, in such a way that the geometric information is explicitly encoded in the architecture. In our empirical evaluation, DAFNO has achieved state-of-the-art accuracy as compared to baseline neural operator models on two benchmark datasets of material modeling and airfoil simulation. To further demonstrate the capability and generalizability of DAFNO in handling complex domains with topology changes, we consider a brittle material fracture evolution problem. With only one training crack simulation sample, DAFNO has achieved generalizability to unseen loading scenarios and substantially different crack patterns from the trained scenario. Our code and data accompanying this paper are available at https://github.com/ningliu-iga/DAFNO.

Paper

References (51)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer BzBv5/10 · confidence 5/52023-06-15

Summary

This paper presents a novel method of extending Fourier Neural Operators (FNO) to irregular geometries using an indicator function to represent the shape of the geometric area. This area is then extended to a larger, regular area, thereby allowing FNO to handle irregular geometries. This is indeed an innovative approach, as the use of indicator functions in neural operators to handle irregular areas has not been done before. The methodology bears some similarities to the use of neural networks in PINNs to learn hard constraints, ensuring regular equations comply with constraints on the geometric boundary. Experimental results indicate that the proposed method achieves commendable performance with fewer parameters.

Strengths

The paper's idea is innovative, using indicator functions in neural operators to handle irregular geometries is a novel approach. The authors show through experimental results that the proposed method performs well with fewer parameters, a significant advantage.

Weaknesses

1. The review of related work seems insufficient. The recent development of methods represented by transformers [1,2,3] that handle irregular geometric areas well is not mentioned or cited at all. Moreover, last year's ICLR Factorized FNO [4] also dealt with several non-uniform geometric area problems. 2. Regarding the technical aspect of the paper, I have a concern. The original irregular grid may be adaptive, implying that we may need a very high resolution when embedding it into a uniform grid. This might mean that the uniform grid's spacing needs to be smaller than the smallest spacing of the original grid, leading to considerable computation waste. The proposed method does not seem to avoid this waste. 3. I suggest that for datasets like airfoil and hyperelasticity, which have been detailed in other works, there is no need for further extensive description in the main text. From a machine learning researcher's perspective, the dataset's attributes, sizes, and challenges are more important than detailed principles of its generation, especially when these have already been thoroughly discussed in other studies. Considering the above points, the authors need first add a few key references, carefully consider the technical questions raised, and then heavily revise the paper. If the core technical challenges (question 2 ) can be addressed, I may consider increasing the score after the rebuttal. Otherwise, I would suggest that the authors continue to revise this work for submission to another conference or journal. 1. Choose a Transformer: Fourier or Galerkin (https://arxiv.org/abs/2105.14995) 2. Transformer for Partial Differential Equations' Operator Learning (https://arxiv.org/abs/2205.13671) 3. GNOT: A General Neural Operator Transformer for Operator Learning (https://arxiv.org/abs/2302.14376) 4. Factorized Fourier Neural Operators (https://arxiv.org/abs/2111.13802)

Questions

None.

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

3 good

Limitations

None.

Reviewer dnax6/10 · confidence 3/52023-07-04

Summary

The paper explores the extension of Fourier Neural Operators (FNOs) to irregular geometries and topology changes. To leverage the computational speed benefits of the fast Fourier transform (FFT) employed by FNOs, they enclose the physical domain with a period box. They then adapt the FNO kernel by multiplying the integrand with a smoothed characteristic function that encodes domain information. This encoding enables their method to generalize to new geometries while preserving the FFT computational efficiency. They outperform several FNO baselines on two benchmark datasets, and showcase the properties of their method on a real-world crack propagation dataset, which contains evolving topology.

Strengths

The paper is well written and clear. The different choices of architecture are well justified. - It tackles an interesting and useful problem for real-world application of FNO. - The idea of incorporating the domain information in the kernel is simple, but effective and grounded, and the smoothing of the domain characteristic function helps avoiding dicontinituty problems. - Experimental results on three datasets are strong. - The crack propagation dataset is well-considered and helps highlighting the advantages of DAFNO, especially its ability to adapt to evolving topologies.

Weaknesses

- The comparison with several baselines is interesting, but given the recent improvements seen over FNO and Geo-FNO, having F-FNO [1] as a baseline seems important. In addition to neural operators, other grid-independent methods that can be used on irregular domains have been developed, for instance [2] implemented a continuous model using Implicit Neural Representations (INRs). It would have been interesting to see a comparison or at least a discussion about the advantage of DAFNO in the light of these recent improvements in the community (for e.g. computational cost?). - As explained in the limitations and clear from the methodology itself, the architecture struggles to handle non-uniformly meshed grids due to the inability to use FFT on such grids. This limitation significantly restricts the range of problems that can be efficiently tackled using this architecture. For example, if applied to real-world airfoil surrogate modeling problems like [3], it seems that DAFNO would not offer significant improvements over the original FNO. - While the idea behind DAFNO is interesting and novel, the architectural improvement upon the regular FNO architecture seems limited. The authors address this issue in a ‘remark’ paragraph, but I am not totally convinced that it is a sufficient improvement, especially since the architecture does not appear to efficiently handle non-uniformly meshed grids. - This is more a remark than a weakness. In Figure 3, the zero-shot super-resolution prediction from eDAFNO does not effectively highlights the resolution invariant properties, since there is no ground truth to compare it to. While some degree of "deblurring" is noticeable, it remains unclear if the network successfully captures higher frequency phenomena that may occur at higher resolutions. The appendix, specifically Figure 10, sheds light on this property for the crack propagation dataset. However, Figure 3 does not truly demonstrates this resolution invariant property. [1]: Tran, Alasdair, et al. "Factorized Fourier Neural Operators." The Eleventh International Conference on Learning Representations. 2022. [2]: Yin, Yuan, et al. "Continuous PDE Dynamics Forecasting with Implicit Neural Representations." The Eleventh International Conference on Learning Representations. 2023. [3]: Bonnet, Florent, et al. "AirfRANS: High Fidelity Computational Fluid Dynamics Dataset for Approximating Reynolds-Averaged Navier–Stokes Solutions." Advances in Neural Information Processing Systems 35 (2022): 23463-23478.

Questions

1. I wonder why Geo-FNO is not in the baselines for the Crack propagation dataset. It would have been an interesting baseline, especially to see if it could adapt to evolving geometries better than a masked FNO. It is claimed that Geo-FNO cannot handle topology change, but once a continuous mapping is learned, it can be applied. I agree that it should not adapt well, but is there a specific reason for its absence in these experimental results? 2. Just to confirm my understanding, for the FNO baseline, the grid is interpolated as in Geo-FNO, and then a mask is given as an additional input? 3. I don’t quite understand the link from equation 5 to the equation before 6. Why is $\chi(x)$ in factor of $W^l h(x) + c^l$ ? Is it to ensure that DAFNO will not produce results for points outside of the domain? 4. Given the link of the smoothing function to a signed distance function (SDF) with the $\\tanh(\\beta \\text{dist}(x, \partial \Omega))$ formulation, is there a reason for not directly using a SDF instead of the smoothing function ? It may have given more (or a different) information to the network, as in different previous works (for e.g. [4]). 5. Out of curiosity, have you experimented with using other smoothing functions besides tanh (for e.g. a logistic function)? [4]: Guo, Xiaoxiao, Wei Li, and Francesco Iorio. "Convolutional neural networks for steady flow approximation." Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 2016.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

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

3 good

Contribution

3 good

Limitations

The main limitations are already presented in the article.

Reviewer UPhM6/10 · confidence 3/52023-07-05

Summary

The Fourier Neural Operator (FNO) is a model in the field of neural operators that has successfully interpreted various physical phenomena. However, one of the issues with FNO is its limitation to learn only on rectangular domains. In Geo-FNO, this problem is addressed by lifting irregular domains to the latent space of rectangular domains. The authors propose a novel convolution integral kernel through the domain characteristic function X(x) and introduce DAFNO (Domazin-Adapted Fourier Neural Operator). DAFNO outperforms the existing baseline in problems involving airfoil and hyperelasticity.

Strengths

Conducting research on a domain-agnostic physics simulator is a crucial endeavor. In particular, the field of neural operators, which the authors have explored, offers a highly efficient approach zto interpreting recent physical information, thus holding significant potential.

Weaknesses

Could you provide more detailed explanation about the mathematical motivation behind the domain characteristic function? It would be necessary to provide specific explanations for Eq(4) and Eq(5). Due to the varying model sizes between the baseline models and the proposed model in Table 1, it is difficult to claim that the comparison is fair. A more equitable comparison can be achieved when all model sizes are standardized to be the same. The authors proposed a new neural operator model called DAFNO. However, besides modifying the internal structure of the iterative layers in Geo-FNO, specifically by splitting the structure of the iterative layers in Figure 2, it is challenging to identify any other distinct novelties. Adding more baselines in the field of neural operators could lead to better experiments and improved evaluations. Such as, MWT (Multiwavelet-based neural operator), WNO (Wavelet neural operator)

Questions

Is there no experimental result for the benchmark dataset (Pipe) based on the Navier-Stokes equation in the Geo-FNO paper? I regard it as one of the significant experiments in this domain. Intuitively, it might be expected that DAFNO, having designed the Iterative layers in parallel, could have larger training and inference times compared to the baseline models like FNO and Geo-FNO. Can you provide any measurement results to demonstrate this?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

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

There is no limitations.

Reviewer fNa37/10 · confidence 4/52023-07-06

Summary

In this paper, the authors propose the Domain Agnostic Fourier Neural Operator (DAFNO), an FNO that can deal with irregular boundaries. While the classical FNO is limited by construction to irregular domains, DAFNO simply includes a smoothed function $I(\cdot)$ of the characteristic function of the domain on which the data is defined. DAFNO is shown to outperform relevant baselines on 2-dimensional problems with irregular boundaries and one problem with topology changing over time.

Strengths

The paper is about a very simple yet effective idea: including a smoothed mask to model boundaries inside the integral operator. This approach is original to my knowledge and effective in terms of strong experimental evaluation relevant to practical applications, including being able to handle topology changes. The paper is clear in its description and well-placed in the current literature on neural operators and deep surrogate models for scientific ML.

Weaknesses

In my opinion, there is no major weakness in this paper. The main limitation could be that the approach is (particularly) simple, but I believe this does not need to be seen as a negative point. Sometimes, simpler is better. The experimental results are solid but do not encompass more complex settings such as 3D fluid dynamics. Another weakness is the lack of source code, which I believe should be provided, given the simplicity of the setting.

Questions

- Can the proposed method be extended to modeling systems ≥ than 3 dimensions? - At line 262: “In general, the topology evolution rule that determines Ω(t) can be obtained from a separate neural network or from physics as is the case in the current example”, did you test out a separate neural network in the experiments as well?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

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

3 good

Limitations

The main limitation (about uniformly meshed grids) is mentioned. Other limitations are covered in the “weaknesses” section.

Reviewer FUid6/10 · confidence 4/52023-07-10

Summary

Fourier Neural Operators (FNOs) are a popular method for modeling physical systems such as different types of PDEs. However, in order to use the FFT to make FNO efficient, the input needs to be a regular grid. The authors study the question of irregular grid inputs for FNO, as well as problems with changing topologies. They propose to add a smoothed characteristic function in the integral layer architecture. This directly encodes the topology into the architecture, and it allows the architecture to continue to use FFT while allowing irregular grid inputs. The authors apply this domain agnostic method to the regular FNO as well as implicit FNO. The authors run experiments on material modeling, airfoil simulation, and material fractures.

Strengths

- This is the first neural operator that can handle learning with topology changes. This is a useful property to handle fracture problems, such as material fracture or earthquakes. - The technique is relatively simple: adding a characteristic function to the integral kernel and adding a smoothing term, but it is novel and elegant that the characteristic function encodes the topology directly in the architecture. This is how the authors’ technique can learn changing topologies, when, say Geo-FNO learns a fixed mapping and cannot learn changing topologies. - The authors experiment on three problems: hyperelasticity, airfoils, and crack propagation, and they also perform ablation studies to separate the characteristic function and the smoothness.

Weaknesses

- My biggest concern is that the technique may not be able to handle highly irregular topologies or topologies with fine-grained features, for two separate reasons. As shown in Figure 1, the resulting grid from the authors’ technique is still uniform, but with a characteristic added, to tell what is inside or outside the topology. Since the grid is still uniform, this may not work well for a topology that is very irregular or has fine features. This is in contrast to Geo-FNO, which learns a mapping, allowing it to handle highly irregular topologies and also to focus more grid points on the “important” parts of the topology and fewer grid points on the coarse / less important parts. Furthermore, the smoothing function could magnify the above problem, making it even harder to model delicate parts of the topology. I think of this as a weakness because the authors say that handling irregular topologies is one of the main contributions of their work. - The technique is fairly simple: it boils down to adding a characteristic function with smoothing. But on the other hand, this technique is the first to handle dynamically changing topologies. - The authors did not release code or mention releasing code. If they authors released code (anonymously during submission) it would have a bigger impact. - Why not compare airfoil with the other baselines? Also, why do the 3rd and 4th panels of Figures 4 look like there is a vertical line discontinuity in the center of the image? - Why not compare the cracking to more baselines than just FNO, such as Geo-FNO?

Questions

Can the authors comment on all of the weaknesses listed above? Also, I wonder if Geo-FNO can be combined with DAFNO, to handle more complex topologies and also be able to handle changing topologies? Do you plan to release your code? It could even be released during the rebuttal with e.g. https://anonymous.4open.science/. I would raise my score if some or all of the points from the "weaknesses" section are addressed.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

The authors discuss limitations in Section 5: they only consider changing domains where the grid is uniform, and they only consider problems with the same boundary conditions as the domain changes. I agree that it would be good to add a mapping for grid deformation in future work.

Reviewer dnax2023-08-17

Thank you for your rebuttal

Thank you for all your differents answers, which I found insightful. My two main reservations concerned the limited related work missing the F-FNO reference, as well as the potential inapplicability of the architecture to non-uniformly meshed grids. The authors have effectively addressed both of these concerns, consequently I will improve my evaluation.

Authorsrebuttal2023-08-19

Thank you!

We thank the reviewer for their kind response and for raising the score. We sincerely appreciate the reviewer's valuable time and suggestions, which helped us to improve the quality of this work.

Reviewer FUid2023-08-18

Thank you for preparing the rebuttal to my review and the other reviews. I appreciate that you released code, added the F-FNO baseline, and ran airfoil with more baselines. I agree with the clarification of the shock wave phenomenon and about running baselines for airfoil and cracking. I also particularly like the example of airfoil with an irregular grid. Overall, I am impressed by the rebuttal and I raised my score. I have a few more questions. What is F-FNO? I think it is the same thing as tensorized FNO (TFNO)? In that case, I am wondering why the parameter count is higher for TFNO than for FNO? Did other hyperparameters change? I also wonder if DAFNO can be applied with the tensorized technique? And, although the new experiment with DAFNO on an irregular grid is great, it would be interesting to have DAFNO with a trainable mapping. It seems like a learnable irregular mapping would complement the domain adaptive part well? Note that I know it is only 4 days before the end of the discussion period, so this is not a request for more experiments, but a question, e.g., for future work.

Authorsrebuttal2023-08-19

Thank you!

We thank the reviewer for their kind response. We sincerely appreciate the reviewer's valuable time and suggestions, which helped us to improve the quality of this work. **Connections between F-FNO and T-FNO**: The reviewer is correct that F-FNO [1] has fundamentally similar architectures to TFNO [2] since they both adopt Fourier factorization, with some small differences such as the improved residual connection and training technique in F-FNO. The reason why F-FNO has more parameters than FNO (as also pointed out in the F-FNO paper) is that F-FNO used a bigger hidden size in the elasticity problem. To guarantee a fair comparison, we kept the original settings of F-FNO as in [1] since the total number of parameters are on a similar level as other methods. [1] Tran, Alasdair, et al. "Factorized Fourier Neural Operators." The Eleventh International Conference on Learning Representations. 2022. [2] Kossaifi, Jean, et al. "Multi-Grid Tensorized Fourier Neural Operator for High Resolution PDEs." (2022). **Future extensions on DAFNO**: We appreciate the reviewer's valuable suggestion. Both the tensorized architecture and the data-driven grid mapping should be readily applicable on DAFNOs. We agree with the reviewer that the combination of these two techniques with DAFNO would indeed be very interesting directions, which we will consider in the future work.

Authorsrebuttal2023-08-19

Thank you!

We thank the reviewer for their kind response and for raising the score. We sincerely appreciate the reviewer's valuable time and suggestions, which helped us to improve the quality of this work.

Reviewer fNa32023-08-19

Thanks

Thanks for your reply. (The code has not been shared with us reviewers, but I trust the AC will check it.) Given my concerns have mostly been solved and the response to the other reviewers, I am happy to raise my score!

Authorsrebuttal2023-08-19

Thank you!

We thank the reviewer for their kind response and for raising the score. We sincerely appreciate the reviewer's valuable time and suggestions, which helped us to improve the quality of this work.

Reviewer BzBv2023-08-22

Feedback

After reading your rebuttal, most of my concerns were resolved. I have raised my score to 5, but not higher due to the concern that I still think computing grid transformations to a uniform grid (like Geo-FNO) is inefficient and has many limitations. By reading your revision I found this is still a challenge. For example, for 3D airfoil problems, you will need very high-resolution grids which might be computationally expensive.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC