Pith. sign in

REVIEW 3 major objections 6 minor 48 references

PDEformer-2: A Versatile Foundation Model for Two-Dimensional Partial Differential Equations

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single pretrained model takes a PDE's symbolic form as input and returns mesh-free solutions across dozens of equation families, domains, and boundary conditions.

desk verdict A substantial and mostly honest 2D PDE foundation model whose 'versatility' is narrower than advertised; deserves peer review with major revisions on evaluation rigor and scope. read the letter →

arxiv 2507.15409 v2 pith:IFCV5YCN submitted 2025-07-21 math.NA cs.NA

Zhanhong Ye , Zining Liu , Bingyang Wu , Hongjie Jiang , Leheng Chen , Minyan Zhang , Xiang Huang , Qinghe Meng. Jingyuan Zou
show 2 more authors
Hongsheng Liu Bin Dong
This is my paper
classification math.NAcs.NA MSC 65M9968T07
keywords PDEfoundationmodeltwo-dimensionalpartialdifferentialequationscomputationalgraphrepresentationimplicitneuralmesh-freepredictionszero-shotpredictionfew-shotadaptationinverseproblems
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

PDEformer-2 is a single neural network pretrained to solve two-dimensional partial differential equations of many kinds. The paper's central claim is that one model, given the PDE's symbolic form as a computational graph along with initial conditions and coefficient fields, predicts solutions at arbitrary spatio-temporal coordinates with zero-shot accuracy for equations that resemble its pretraining distribution, and that on new equations it adapts faster and from fewer samples than specialized neural operators. The claim is supported by a 40 TB pretraining dataset of 3.29 million examples spanning eight equation families, by a graph-transformer encoder fused with an implicit neural representation decoder, and by forward, few-shot, and inverse-problem experiments. If the paper is right, a single differentiable, mesh-free surrogate could replace per-equation operator training and case-by-case mesh construction for 2D PDEs, at least within the families and regimes the pretraining data covers.

What carries the argument

The load-bearing object is the computational graph representation of the PDE, a directed acyclic graph whose nodes are typed mathematical objects (unknown field variables, scalar coefficients, coefficient fields, initial conditions, signed distance functions) and operations (partial derivatives, sums, products, nonlinear terms). Scalar numbers enter through an MLP and numerical fields enter through a CNN function encoder; a graph Transformer with shortest-path attention biases fuses the symbolic and numeric information into per-variable latent codes; an adapted Poly-INR decoder then maps any coordinate $(t,x,y)$ to a predicted solution value using layer-wise modulations. The signed distance function representation of the domain and its boundary curves is what lets the model handle non-rectangular domains and mixed boundary conditions without a fixed mesh.

What would settle it

Regenerate reference solutions for PDEs in the same eight families but in regimes the spectral solver tends to reject—for example, wave equations with high-frequency sinusoidal initial conditions or sharp pulses near absorbing boundaries—using a robust adaptive solver that succeeds on them; then measure PDEformer-2's zero-shot nRMSE. If the error on those solver-rejected instances is several times the 8.7% training average and stays above specialized operators even after finetuning on hundreds of samples, the central versatility claim is falsified exactly where the teacher solvers fail.

Watch

Extended reading notes

Core claim

The paper's central claim is that the symbolic form of a 2D PDE can be the input to a foundation model. PDEformer-2 encodes the equation, boundary conditions, and domain shape as a directed computational graph whose nodes are mathematical objects and operations, embeds scalar coefficients and numerical fields through separate encoders, processes the graph with a graph Transformer, and decodes per-variable solutions with an implicit neural representation (adapted Poly-INR) that can be queried at any point in $[0,1]\times\Omega$. Pretrained on 3.29 million solver-generated samples (about 40 TB) covering diffusion-convection-reaction, wave, multi-variable, divergence-constrained, shallow-water, and elasticity equations, the model reports a mean test nRMSE of 8.7% (base) and 10.3% (fast) across all families, extrapolates to five coupled variables, and, after finetuning, outperforms from-scratch specialized operators when fewer than 100 samples are available. It also serves as a differentiable surrogate for inverse problems, recovering coefficient scalars, sparse system structure, source fields, and wave velocity fields from noisy observations.

Load-bearing premise

The load-bearing premise is that the randomly sampled PDEs the traditional solvers managed to solve (success rates 10%–90% per family, as reported in Supplementary C.1) are representative enough of 'most common' 2D PDEs to support the zero-shot and few-shot claims; the paper itself notes in the Discussion that the model is not yet validated on real-world data and that accuracy and speed are not universally better than baselines.

Editorial extensions

If this is right

  • Zero-shot solving moves from per-equation training to one pretrained inference: a 2D PDE expressible in the supported node vocabulary and close to the pretraining distribution can be queried at arbitrary spatio-temporal points without retraining.
  • With fewer than 100 training samples, finetuned PDEformer-2 reports lower test nRMSE than from-scratch DeepONet, FNO, U-Net, and Geo-FNO variants on the tested datasets, and it converges in about ten seconds on the INS-Pipe problem.
  • The model extrapolates to five coupled unknown variables even though pretraining contained at most four, suggesting that the computational graph representation transfers beyond the variable-count distribution seen during pretraining.
  • The same differentiable surrogate solves inverse problems in the tested settings: recovering scalar coefficients, identifying near-zero coefficients (system identification), reconstructing source fields (average nRMSE 0.159), and recovering wave velocity fields from noisy, sparsely observed data.
  • On at least some test cases the model's predictions are closer to a second solver's reference than to the non-physical output of the teacher solver that produced its training labels.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the central claim generalizes, expressing a PDE as a computational graph becomes the natural interface for symbolic regression: the same graph input that drives forward prediction could be optimized to discover governing equations from data, making system identification a direct by-product of the architecture rather than a separate task.
  • The mesh-free INR output, combined with auto-differentiation, opens an adjoint-free route to shape and boundary-condition optimization; the paper's own speed tables suggest that practical use would rely on the encoder-only mode (produce the INR) and query only selected points, as the inverse experiments already do.
  • The reported weak spots—high-frequency initial conditions and absorbing-boundary wave problems—are exactly the regimes where spectral solvers struggle, which suggests a testable scaling hypothesis: zero-shot accuracy across PDE families should track the teacher solver's success rate, and pretraining with multiple solvers of complementary strengths should widen the model's reliable range.
  • A direct extension would be to measure how zero-shot nRMSE changes when the data generator's rejection rate is deliberately varied within one family; if error tracks rejection rate, dataset coverage, not model capacity, is the current bottleneck for versatility.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces PDEformer-2, a foundation model for two-dimensional PDEs that takes a computational graph representation of the PDE as input and produces mesh-free spatio-temporal solutions through an implicit neural representation decoder. The model is pretrained on approximately 3.29 million samples from eight generic PDE families (totaling 40 TB) generated by the classical solvers Dedalus and FEniCSx. The authors report an average zero-shot nRMSE of 8.7% on the held-out pretraining test set, faster few-shot adaptation than specialized neural operators on several datasets, and demonstrations of inverse problems including scalar coefficient recovery, system identification, source-field recovery, and wave-velocity recovery. The paper also includes ablations on imperfect PDE specifications and a study of extrapolation to PDEs with more variables.

Significance. If the reported results hold, PDEformer-2 represents a substantial step toward a practical foundation model for PDEs: its computational-graph input is flexible across equation forms, domains, boundary conditions, and numbers of variables, and its INR output allows queries at arbitrary spatio-temporal coordinates. The paper ships code, pretrained weights, and a large dataset, which are valuable community assets. The few-shot experiments against several baseline neural operators and the inverse-problem case studies are extensive and are reported in a generally transparent manner. The authors are also explicit about several limitations, including degraded performance on out-of-distribution regimes and on PDEs that the teacher solver could not handle. The central claims are supported by the experiments as presented, though the scope of the zero-shot claim is narrower than the title and abstract suggest, and the evaluation protocol has methodological weaknesses that need addressing.

major comments (3)
  1. [Section 2.1; Supplementary C.1] The pretraining dataset is generated by randomly sampling PDE forms and retaining only the cases that Dedalus or FEniCSx could solve, with success rates between 10% and 90%. Consequently, the phrase in the abstract, 'PDEs that resemble the pretraining ones,' effectively means 'PDEs that the teacher solver could solve.' The paper's own Supplementary Figure 24 shows that zero-shot nRMSE against FEniCSx is unsatisfactory on most of the 100 Dedalus-failure cases, and Figure 9d-e show degradation when the target PDE deviates from the pretraining distribution. The claim of a 'versatile foundation model for 2D PDEs' therefore overstates the demonstrated scope. Please temper the title/abstract or provide evidence that the solver-solvable subset is representative of the 'most common' 2D PDEs, for example by characterizing the failure modes of Dedalus and showing that the retained samples cover the intended application regimes.
  2. [Supplementary D.2; Figure 9] The evaluation protocol records the best test-set accuracy over training epochs for all models, rather than selecting checkpoints on an independent validation set. The authors acknowledge this is less rigorous and assert that validation-set and test-set errors are close, but they do not present the supporting comparison. Since the few-shot superiority claim (e.g., Figures 9a-9e) is a central result, the current protocol could optimistically bias the reported numbers, especially for the baselines. Please re-evaluate with validation-based early stopping or provide quantitative evidence that the two protocols give similar results.
  3. [Figures 9-11; Tables 2-3] All reported accuracy numbers appear to come from a single training or finetuning run, with no error bars or seed variation. At small sample sizes (e.g., 4 or 20 training samples in Figure 9), the observed differences between PDEformer-2 and the baselines could be within run-to-run variability. The central claims about efficient adaptation and superior few-shot accuracy would be considerably strengthened by reporting means and standard deviations over at least three independent seeds, or at minimum clearly stating that the results are single runs and discussing the potential variability.
minor comments (6)
  1. [Section 2.3] The text contains a typo: 'assmue' should be 'assume'.
  2. [Section 1] Two typos appear in the introduction: 'Cartisian' should be 'Cartesian' and 'priliminarily' should be 'preliminarily'.
  3. [Figure 9 caption] The word 'verticle' should be 'vertical'.
  4. [Section 4] The word 'intepretation' should be 'interpretation'.
  5. [References] References [33] and [34] refer to the same Poly-INR paper; one duplicate should be removed.
  6. [Section 2.1; Figure 6] The claim that PDEformer-2 'outperforms its teacher solver' is based on a selected example and a scatter plot. Please state more carefully that this is an illustrative observation on a small set of cases, not a general superiority claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PDEformer-2 is evaluated on held-out samples and external references, and its self-citations are architectural lineage rather than load-bearing derivations.

full rationale

PDEformer-2 is a supervised foundation model: it is pretrained on solver-generated labels and tested on held-out samples drawn from the same distribution, which is standard empirical evaluation rather than circular reasoning. Zero-shot predictions are genuine forward inferences at test time; no parameter is fitted to the test targets, and few-shot results are produced by finetuning on separate training samples and measuring error on held-out labels. Inverse problems use PDEformer-2 as a differentiable surrogate in an optimization loop against noisy observations, which is a standard downstream application and not a renamed fit of the reported metrics. The self-citations to PDEformer-1 concern architecture lineage and hyperparameter selection only; the versatility claims are independently demonstrated through eight PDE families and external benchmarks such as Wave-Gauss, Wave-C-Sines, and FEniCSx reference solutions. The paper explicitly acknowledges the scope limitation imposed by solver-filtered pretraining data, including unsatisfactory performance on teacher-solver failure cases and on out-of-distribution PDEs. These are honest boundary conditions, not circular steps. No equation is shown to reduce to another by construction, no fitted parameter is relabeled as a prediction, and no uniqueness theorem is imported from the authors' prior work.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No physics-based free parameters or invented entities. The model's weights are learned from data, and architecture hyperparameters (N=4, L=11, d_e=768) are copied from PDEformer-1. The assumptions listed are the load-bearing data and representation premises.

assumptions (3)
  • domain assumption Reference solutions generated by Dedalus and FEniCSx are sufficiently accurate ground truths for the PDEs in the pretraining and test sets.
    All training labels and most test references come from these solvers (Supplementary C). The paper itself documents non-physical oscillations and NaN failures (Section 2.1), so the assumption is only partially satisfied and affects the soundness of accuracy numbers.
  • domain assumption The eight generic PDE forms with random coefficient selection span the space of 'most common' 2D PDEs that the model claims to handle.
    The versatility claim rests on the pretraining distribution covering DCR, wave, multi-variable, divergence-constrained, shallow-water, and elasticity families (Table 2). If real-world PDEs fall outside this distribution, zero-shot performance degrades, as shown by the Wave-Gauss and Wave-C-Sines results (Figure 9d-e).
  • domain assumption Computational graph representation with SDF-encoded domains and boundary conditions can faithfully express all problem inputs the model is given.
    The architecture relies on converting symbolic PDEs, domains, and BCs into DAGs (Section 4, Supplementary B). Any lossy encoding would cap achievable accuracy; the paper provides examples but no guarantee of completeness for arbitrary 2D PDEs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PDEformer-2: A Versatile Foundation Model for Two-Dimensional Partial Differential Equations." pith.science (2026). https://pith.science/paper/IFCV5YCN

@misc{pith2026250715409,
  author       = {Pith},
  title        = {Pith review of: PDEformer-2: A Versatile Foundation Model for Two-Dimensional Partial Differential Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IFCV5YCN}},
  note         = {Machine review of arXiv:2507.15409}
}
read the original abstract

Partial differential equations (PDEs) play a central role in describing many physical phenomena. Various scientific and engineering applications demand a versatile and differentiable PDE solver that can quickly generate solutions with adequate accuracy, and limitations of the traditional solvers and specialized neural operators motivate the development of foundation models for solving PDEs. This paper introduces PDEformer-2, a versatile foundation model for two-dimensional PDEs. Based on our previous one-dimensional PDEformer-1 model, PDEformer-2 receives the PDE form as network input via computational graph representation, which has the flexibility to encode most common PDEs. The mesh-free predicted solutions can be directly queried at arbitrary spatio-temporal coordinates. A large (40TB) diverse dataset is employed to pretrain the current model, making it capable of simultaneously addressing PDEs with different symbolic forms, domain shapes, boundary conditions, number of variables, and time-dependency. Accurate zero-shot prediction is allowed for PDEs that resemble the pretraining ones. When adapted to new unseen PDEs, PDEformer-2 demonstrates faster learning than many specialized models, and has smaller errors given limited (less than 100) samples. Additionally, PDEformer-2 can be employed in the inverse problems thanks to its fast and differentiable nature and produces reasonable results in our experiments to recover coefficient scalars and fields of a PDE.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 33 canonical work pages

  1. [1]

    Quarteroni, A., Manzoni, A., Negri, F.: Reduced Basis Methods for Partial Differential Equations: An Introduction, pp. 1–263. Springer, ??? (2015). https: //doi.org/10.1007/978-3-319-15431-2

    doi:10.1007/978-3-319-15431-2 2015

  2. [2]

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A.: Fourier Neural Operator for Parametric Partial Differential Equations (2021)

    work page 2021

  3. [3]

    Nature Machine Intelligence3(3), 218–229 (2021) https://doi.org/10.1038/ s42256-021-00302-5

    Lu, L., Jin, P., Pang, G., Zhang, Z., Karniadakis, G.E.: Learning nonlinear operators via deeponet based on the universal approximation theorem of opera- tors. Nature Machine Intelligence3(3), 218–229 (2021) https://doi.org/10.1038/ s42256-021-00302-5

    work page 2021

  4. [4]

    Lorsung, C., Li, Z., Farimani, A.B.: Physics Informed Token Transformer (2023)

    work page 2023

  5. [5]

    Yang, L., Liu, S., Osher, S.J.: Fine-Tune Language Models as Multi-Modal Differential Equation Solvers (2024)

    work page 2024

  6. [6]

    CoRRabs/2405.17527(2024)

    Zhou, H., Ma, Y., Wu, H., Wang, H., Long, M.: Unisolver: Pde-conditional transformers are universal pde solvers. CoRRabs/2405.17527(2024)

    arXiv 2024

  7. [7]

    Transactions on Machine Learning 26 Research (2024)

    Shen, J., Marwah, T., Talwalkar, A.: UPS: Efficiently building foundation models for PDE solving via cross-modal adaptation. Transactions on Machine Learning 26 Research (2024)

    work page 2024

  8. [8]

    arXiv preprint arXiv:2502.06026 (2025)

    Negrini, E., Liu, Y., Yang, L., Osher, S.J., Schaeffer, H.: A multimodal pde foundation model for prediction and scientific text descriptions. arXiv preprint arXiv:2502.06026 (2025)

    arXiv 2025

Show all 48 references
  1. [9]

    arXiv preprint arXiv:2410.01137 (2024)

    Lorsung, C., Farimani, A.B.: Explain like i’m five: Using llms to improve pde surrogate models with text. arXiv preprint arXiv:2410.01137 (2024)

    2024 arXiv

  2. [10]

    Liu, Y., Zhang, Z., Schaeffer, H.: PROSE: Predicting Operators and Symbolic Expressions using Multimodal Transformers (2023)

    2023

  3. [11]

    Physical Review E111(3), 035304 (2025)

    Sun, J., Liu, Y., Zhang, Z., Schaeffer, H.: Towards a foundation model for partial differential equations: Multioperator learning and extrapolation. Physical Review E111(3), 035304 (2025)

    2025

  4. [12]

    arXiv preprint arXiv:2408.16168 (2024)

    Sun, J., Zhang, Z., Schaeffer, H.: Lemon: Learning to learn multi-operator networks. arXiv preprint arXiv:2408.16168 (2024)

    2024 arXiv

  5. [13]

    arXiv preprint arXiv:2409.09811 (2024)

    Liu, Y., Sun, J., He, X., Pinney, G., Zhang, Z., Schaeffer, H.: Prose-fd: A multi- modal pde foundation model for learning multiple operators for forecasting fluid dynamics. arXiv preprint arXiv:2409.09811 (2024)

    2024 arXiv

  6. [14]

    Proceedings of the National Academy of Sciences120(39), 2310142120 (2023)

    Yang, L., Liu, S., Meng, T., Osher, S.J.: In-context operator learning with data prompts for differential equation problems. Proceedings of the National Academy of Sciences120(39), 2310142120 (2023)

    2023

  7. [15]

    Journal of Computational Physics519, 113379 (2024) https://doi.org/10.1016/j.jcp.2024.113379

    Yang, L., Osher, S.J.: Pde generalization of in-context operator networks: A study on 1d scalar nonlinear conservation laws. Journal of Computational Physics519, 113379 (2024) https://doi.org/10.1016/j.jcp.2024.113379

    2024

  8. [16]

    https://arxiv.org/abs/2411.16063

    Cao, Y., Liu, Y., Yang, L., Yu, R., Schaeffer, H., Osher, S.: VICON: A Foundation Model for Multi-Physics Fluid Dynamics via Vision In-Context Operator Networks (2025). https://arxiv.org/abs/2411.16063

    2025

  9. [17]

    McCabe, M., Blancard, B.R.-S., Parker, L.H., Ohana, R., Cranmer, M., Bietti, A., Eickenberg, M., Golkar, S., Krawezik, G., Lanusse, F., Pettee, M., Tesileanu, T., Cho, K., Ho, S.: Multiple Physics Pretraining for Physical Surrogate Models (2023)

    2023

  10. [18]

    In: Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., Berkenkamp, F

    Hao, Z., Su, C., Liu, S., Berner, J., Ying, C., Su, H., Anandkumar, A., Song, J., Zhu, J.: DPOT: Auto-regressive denoising operator transformer for large-scale PDE pre-training. In: Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., Berkenkamp, F....

    2024

  11. [19]

    CoRRabs/2402.16014 (2024)

    Chen, T., Zhou, H., Li, Y., Wang, H., Gao, C., Zhang, S., Li, J.: Building flexible machine learning models for scientific computing at scale. CoRRabs/2402.16014 (2024)

    2024 arXiv

  12. [20]

    arXiv preprint arXiv:2501.18972 (2025)

    Liu, Y., Sun, J., Schaeffer, H.: Bcat: A block causal transformer for pde foundation models for fluid dynamics. arXiv preprint arXiv:2501.18972 (2025)

    2025 arXiv

  13. [21]

    Advances in Neural Information Processing Systems36, 71242–71262 (2023)

    Subramanian, S., Harrington, P., Keutzer, K., Bhimji, W., Morozov, D., Mahoney, M.W., Gholami, A.: Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. Advances in Neural Information Processing Systems36, 71242–71262 (2023)

    2023

  14. [22]

    Advances in Neural Information Processing Systems37, 104035–104064 (2024)

    Rahman, M.A., George, R.J., Elleithy, M., Leibovici, D., Li, Z., Bonev, B., White, C., Berner, J., Yeh, R.A., Kossaifi, J.,et al.: Pretraining codomain attention neural operators for solving multiphysics pdes. Advances in Neural Information Processing Systems37, 104035–104064 (2024)

    2024

  15. [23]

    In: Advances in Neural Information Processing Systems, vol

    Herde, M., Raonić, B., Rohner, T., Käppeli, R., Molinaro, R., Bézenac, E., Mishra, S.: Poseidon: Efficient foundation models for pdes. In: Advances in Neural Information Processing Systems, vol. 37 (2024). https://proceedings.neurips.cc/paper_files/paper/2024/hash/ 84e1b1ec17b...

    2024

  16. [24]

    arXiv preprint arXiv:2407.06664 (2024)

    Ye, Z., Huang, X., Chen, L., Liu, Z., Wu, B., Liu, H., Wang, Z., Dong, B.: Pdeformer-1: A foundation model for one-dimensional partial differential equations. arXiv preprint arXiv:2407.06664 (2024)

    2024 arXiv

  17. [25]

    Physical Review Research2(2), 023068 (2020) https://doi.org/10.1103/PhysRevResearch.2.023068 arXiv:1905.10388 [astro-ph.IM]

    Burns, K.J., Vasil, G.M., Oishi, J.S., Lecoanet, D., Brown, B.P.: Dedalus: A flexible framework for numerical simulations with spectral methods. Physical Review Research2(2), 023068 (2020) https://doi.org/10.1103/PhysRevResearch.2.023068 arXiv:1905.10388 [astro-ph.IM]

    2020 arXiv

  18. [26]

    preprint (2023)

    Baratta, I.A., Dean, J.P., Dokken, J.S., Habera, M., Hale, J.S., Richardson, C.N., Rognes, M.E., Scroggs, M.W., Sime, N., Wells, G.N.: DOLFINx: the next generation FEniCS problem solving environment. preprint (2023). https://doi. org/10.5281/zenodo.10447666

    2023 doi

  19. [27]

    In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pp

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, p...

    2015

  20. [28]

    Journal of Machine Learning Research24(388), 1–26 (2023)

    Li, Z., Huang, D.Z., Liu, B., Anandkumar, A.: Fourier neural operator with learned deformations for pdes on general geometries. Journal of Machine Learning Research24(388), 1–26 (2023)

    2023

  21. [29]

    arXiv preprint arXiv:2501.19205 (2025)

    Mousavi, S., Wen, S., Lingsch, L., Herde, M., Raonić, B., Mishra, S.: Rigno: A 28 graph-based framework for robust and accurate operator learning for pdes on arbitrary domains. arXiv preprint arXiv:2501.19205 (2025)

    2025

  22. [30]

    In: Proceed- ings of ICNN’95 - International Conference on Neural Networks, vol

    Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceed- ings of ICNN’95 - International Conference on Neural Networks, vol. 4, pp. 1942–19484 (1995). https://doi.org/10.1109/ICNN.1995.488968 . https://ieeexplore.ieee.org/abstract/document/488968Accessed 2024-04-03

    1995

  23. [31]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  24. [32]

    Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen, Y., Liu, T.-Y.: Do transformers really perform badly for graph representation? Advances in neural information processing systems34, 28877–28888 (2021)

    2021

  25. [33]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Singh, R., Shukla, A., Turaga, P.: Polynomial implicit neural representations for large diverse datasets. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2041–2051 (2023)

    2023

  26. [34]

    In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp

    Singh, R., Shukla, A., Turaga, P.: Polynomial implicit neural representations for large diverse datasets. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2041–2051 (2023). https://doi.org/10.1109/ CVPR52729.2023.00203

    2023

  27. [35]

    Lu, L., Meng, X., Cai, S., Mao, Z., Goswami, S., Zhang, Z., Karniadakis, G.E.: A comprehensive and fair comparison of two neural operators (with practical extensions) based on fair data. Computer Methods in Applied Mechanics and Engineering393, 114778 (2022) https://doi.org/10...

    2022

  28. [36]

    2D convolution with32output channels, stride4, kernel size4×4, with bias included

  29. [37]

    Rectified linear unit (ReLU) activation function

  30. [38]

    2D convolution with128output channels, stride4, kernel size4 × 4, with bias included

  31. [39]

    ReLU activation function

  32. [40]

    2D convolution with768output channels, stride4, kernel size4 × 4, with bias included

  33. [41]

    Flattening operation, applied to the2×2feature map with768channels

  34. [42]

    DeepONet (MLP)

    Vector spliting intoN= 4feature embeddings assigned to the branch nodes. A.2.2 Graph Transformer Initial embedding vector In the graph Transformer, the initial embedding of nodeiis given as h(0) i =x type(i) +ξ i +z − deg−(i) +z + deg+(i), where ξi ∈R de is the input feature e...

  35. [43]

    +∂ y(u0u1) +∂ xp= 0, ∂tu1 −a∆u 1 +∂ x(u0u1) +∂ y(u2

  36. [44]

    +∂ yp= 0, ∂xu0 +∂ yu1 = 0. (10) If readers are more familiar with the convection form involvingu·∇u, its mathematical equivalence8 to the conservation form can be shown using the identity∇ ·(uuT) = u· ∇u+ (∇ ·u)u=u· ∇uwhen∇ ·u= 0holds. List of datasets: • dcdcr_icA: Basic DC-D...

  37. [45]

    +∂ y(c200u0 +u 2 1) −a0∇ ·(a0∇u0) +∂ xp= 0, ∂tu1 +c 010u0 +c 011u1 +∂ x(u2

  38. [46]

    39: Visualization of the divergence-constrained DCR equation on a square domain periodic along both axes and initial conditions violating the divergence constraint

    +∂ y(u0 +u 0u1) +s 1 +∂ yp−p= 0, ∂xu0 +∂ yu1 +u 1 +c 2 = 0 Fig. 39: Visualization of the divergence-constrained DCR equation on a square domain periodic along both axes and initial conditions violating the divergence constraint. 71    ∂ttu0 +µ 0∂tu0 +u 1 +c 000u0 +b 0001u...

  39. [47]

    +∂ y(u0 +u 1) −a0∇ ·(a0∇u0) +∂ xp+ (−c 0)p= 0, ∂ttu1 +µ 1∂tu1 +u 2 1 +∂ x(b1101u0u1 +b 1100u2

  40. [48]

    41: Visualization of the divergence-constrained wave equation on a square domain periodic along both axes and initial conditions satisfying the divergence constraint

    +∂ y(u0 +c 211u1) −a1∆u1 +∂ yp+ (−c 1)p= 0, ∂xu0 +∂ yu1 +c 0u0 +c 1u1 = 0 Fig. 41: Visualization of the divergence-constrained wave equation on a square domain periodic along both axes and initial conditions satisfying the divergence constraint. 73    ∂ttu0 +µ 0...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.