Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

GITO: Graph-Informed Transformer Operator for Learning Complex Partial Differential Equations

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read GITO, a hybrid graph–transformer operator, reports lower relative L2 errors than six neural operator baselines on the Navier–Stokes, heat-conduction, and airfoil benchmarks, with improvements between 3.6% and 46.7%.

desk verdict GITO is a plausible new architecture, but the benchmark supremacy claim is not yet established because the baselines are imported from other papers and there are no error bars. read the letter →

arxiv 2506.13906 v1 pith:37AAUDIY submitted 2025-06-16 cs.LG

Milad Ramezankhani , Janak M. Patel , Anirudh Deodhar , Dagnachew Birru This is my paper
classification cs.LG
keywords neuraloperatorstransformersgraphnetworkspartialdifferentialequationsirregulargeometriesoperatorlearningzero-shotsuper-resolution
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

This paper tries to establish that combining a graph neural network with a transformer—through a dedicated self-attention fusion layer—produces a neural operator that predicts PDE solutions on irregular, non-uniform meshes more accurately than existing operators. The proposed architecture, GITO, reads input functions and query points as graphs, encodes local neighborhood relations with a GNN and long-range dependencies with global attention, then fuses the two views before a cross-attention decoder evaluates the solution at arbitrary spatial locations. Reported results on three benchmark datasets—steady 2D Navier–Stokes flow, multi-layer heat conduction, and transonic airfoil flow—show consistently lower relative L2 errors than FNO, Geo-FNO, MIONet, GNOT, GKT, and OFormer, with relative gains of roughly 3.6% to 46.7% over the best prior numbers. If these comparisons hold, GITO would be a competitive mesh-agnostic surrogate for engineering PDE problems, with zero-shot super-resolution as a built-in property.

What carries the argument

The central object is the hybrid graph transformer (HGT) module followed by a transformer neural operator (TNO) module. HGT encodes the input function and the query points as k-nearest-neighbor or radius graphs, runs a graph attention network (GATv2) for local message passing in parallel with a linear-complexity global self-attention layer, then concatenates the two representations and passes them through a self-attention fusion layer that produces a combined relational embedding. TNO then uses linear-complexity cross-attention with the encoded query embeddings as queries and the HGT output as keys and values, followed by self-attention among enriched queries and an MLP decoder, which makes the model discretization-invariant. The linear attention, adopted from the GNOT work, keeps the cost linear in the number of points and supports evaluating the output at arbitrary query locations, which is what the paper means by zero-shot super-resolution.

What would settle it

Retrain FNO, Geo-FNO, MIONet, GNOT, GKT, and OFormer on the same train/test splits of the three datasets with matched parameter counts, identical normalization, the same relative L2 metric, and the same or larger training budget, then compare. If any baseline achieves lower relative L2 error than GITO on any variable, the paper's central outperformance claim would have to be revised. A cheaper first check is to compute GITO's error on the exact test indices and normalization used in the original GNOT and Geo-FNO repositories and see whether the published baseline numbers reproduce under those conditions.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the hybrid graph transformer (HGT) module—in particular its fusion layer, which applies self-attention to the concatenated outputs of a GNN and a global-attention block—is what makes the operator accurate, and that the transformer neural operator (TNO) with linear-complexity cross-attention enables the model to answer queries at any point in the domain. GITO consistently achieves the lowest relative L2 error across all three datasets and all reported variables, outperforming both convolutional-style operators and prior transformer-based operators such as GNOT, GKT, and OFormer. The paper attributes the gains to richer relational embeddings on graph-structured geometry and to the fusion mechanism, which an ablation shows yields lower error than simply summing the GNN and attention outputs even when the sum variant is given twice the hidden size.

Load-bearing premise

The central ranking claim assumes that the baseline error numbers imported from two earlier papers (the GNOT paper for Navier-Stokes and heat conduction, and the Transolver paper for airfoil) were produced under the same data splits, normalization, evaluation metric, and training budgets as GITO's, since those baselines were not retrained in this work. If the imported numbers are not comparable, the reported improvement percentages and the 'lowest error' claim could change.

Editorial extensions

If this is right

  • The same trained GITO model can evaluate the solution field at arbitrary query points, so a model trained on one mesh performs zero-shot super-resolution on finer or differently discretized meshes without retraining.
  • If the reported numbers hold, GITO's largest relative gains appear on the Airfoil geometry task (about 46.7% relative L2 improvement over GKT, the best baseline), suggesting the graph-informed encoding is most beneficial for sparse, irregular geometric domains.
  • The ablation shows that removing the fusion layer hurts accuracy even when the simplified model has more parameters, implying the fusion mechanism itself—not raw capacity—is the source of the gain.
  • Because the attention mechanism scales linearly, the architecture can in principle be applied to larger numbers of input and query points than quadratic-attention operators, though the paper only demonstrates this on the three reported benchmarks.
  • The graph construction strategy is dataset-dependent: a moderate-radius circular graph works best for the dense NS dataset, a small KNN graph works best for the sparse Heat dataset, and KNN with 16 neighbors is chosen for Airfoil, meaning practitioners need to tune the graph to the geometry.

Reading between the lines

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

  • If the outperformance survives a unified comparison protocol, the 'fuse local message passing with global attention through an attention layer' design could be dropped into other operator architectures (e.g., in place of the branch or trunk encoders in DeepONet-style operators, or the token mixing in Fourier/Galerkin operators) and plausibly yield similar gains; the paper does not test this transf
  • The dataset-dependent best graph construction hints that making the connectivity itself learnable—rather than tuned per dataset—could remove a hyperparameter and improve generalization to unseen mesh geometries; a natural next experiment would be to learn edge weights or neighbor selection jointly with the operator.
  • The reported improvement is smallest for the pressure variable on the Navier-Stokes dataset (3.6% relative), which may indicate that the hybrid graph encoding helps most with geometrically driven fields (velocity, Mach number) and less with pressure, a pattern the paper does not discuss but a reader could test.
  • A concrete testable extension would be to apply GITO to a 3D steady-state problem or a time-dependent rollout to see whether the linear-complexity attention and graph encoding keep their advantage outside the 2D steady-state setting.
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 / 4 minor

Summary. The paper proposes GITO, a graph-informed transformer operator for learning PDE solution operators on irregular geometries and non-uniform meshes. The architecture combines a hybrid graph transformer (HGT), which interleaves a GNN and global self-attention with a fusion layer, with a transformer neural operator (TNO) that uses linear-complexity cross- and self-attention to enable discretization-invariant queries. Experiments on three benchmark datasets (Navier-Stokes, Heat, Airfoil) report lower relative L2 errors than several baselines, and ablations study the fusion layer and graph construction strategies. The paper also claims zero-shot super-resolution and mesh-agnostic generalization as contributions.

Significance. If the reported results hold under a controlled comparison, GITO would be a competitive mesh-agnostic neural operator for 2D steady-state PDE families, with the HGT/GNN-local and transformer-global combination being a sensible design that addresses a real limitation of pure transformer and pure GNN operators. The paper gives credit for using public datasets, reporting ablations of the fusion layer and graph construction choices, and providing architecture details. However, the central 'outperforms' claim depends on cross-paper baseline numbers that are not re-run, no error bars or repeated-seed statistics are provided, and the zero-shot super-resolution claim is not experimentally demonstrated. These issues currently make the headline accuracy advantages difficult to verify.

major comments (3)
  1. [Section 4.1, Table 1] This is the main load-bearing point.
  2. [Abstract / Section 5] This is part of the paper's stated contribution, not merely a future-work remark.
  3. [Table 3, Section 4.3.2] The typo itself is local, but it undermines the reliability of the ablation evidence.
minor comments (4)
  1. [Introduction] The sentence 'Neural operators extend the concept of surrogate modeling by directly mappings infinite-dimensional input-output spaces' contains a typo ('mappings' should be 'mapping').
  2. [Figure 4] The colorbar label reads 'Abosulte Error'; this should be 'Absolute Error'.
  3. [Appendix A.2] The hyperparameter table reports GITO hidden sizes before fusion and notes that exact parameter matching with GNOT was not possible. The narrative is understandable, but it would be clearer to state the final effective hidden dimension after concatenation and to give the parameter counts for all compared baselines, not only GNOT and GITO.
  4. [Section 4.2] The caption of Table 1 says 'The top first and second best results are highlighted', but no highlighting is visible in the text version; please either render the highlights or describe them in the caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GITO's benchmark claims are evaluated against external datasets and external baselines, with no parameter fitted to the target metric and then reported as a prediction.

full rationale

This is an empirical architecture paper whose stated goal is improved accuracy on three benchmark PDE datasets. The target result — relative L2 error on held-out test sets — is measured against data produced by external repositories (the GNOT and Geo-FNO datasets) and against numbers published by other groups (Hao et al. and Wu et al.). GITO's architecture is a new combination of a graph transformer and a transformer neural operator; the paper does not derive a physical constant or fit a parameter to the benchmark metric and then rename that fit as a prediction. The ablations compare design variants (fusion layer, graph construction) under controlled hyperparameters; choosing the best-performing variant is model selection, not circularity, because the chosen variant is still evaluated on the same external test procedure. The only load-bearing comparison issue is that most baselines are imported from prior papers rather than re-run under one identical protocol, and that is a validity/fairness concern about the empirical ranking, not a self-definitional or self-citation circularity. No passage equates an input with an output by construction, and no claim is justified solely by a citation to the authors' own prior work; the only self-reference, Ramezankhani and Milani [6], is a related-work citation and is not load-bearing. The unsupported zero-shot super-resolution statement in the abstract is a missing-experiment concern, not circular reasoning. Accordingly, no circular step can be quoted, and the honest finding is no significant circularity with score 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The reported comparisons rest on external benchmark datasets and imported baselines; no physical constants or new entities are introduced. The central claim depends on per-dataset tuning of graph construction and model capacity, plus the comparability of imported baseline numbers.

free parameters (2)
  • Graph construction strategy (KNN k or radius) = NS: circular radius 0.0525; Heat: KNN k=4; Airfoil: KNN k=16
    Selected per dataset via ablations in Section 4.3.2 to minimize relative L2 error; the reported final results use these tuned choices.
  • Model capacity hyperparameters (hidden size, attention layers, MLP layers, experts) = NS: hidden 96, 2 attention layers, 2 experts; Heat: hidden 128, 3 layers, 3 experts; Airfoil: hidden 96, 2 layers, 2…
    Chosen by hand per dataset (Table 5) to keep GNOT and GITO parameter counts similar; the central comparison depends on these settings.
assumptions (3)
  • domain assumption Published baseline errors from Hao et al. [13] and Wu et al. [15] are directly comparable to GITO's own runs.
    Section 4.1 and Table 1 import these numbers; different training protocols, normalization, or data splits could change the relative ranking.
  • domain assumption Linear-complexity attention from GNOT [13] is a faithful replacement for softmax attention in operator learning.
    Adopted in Section 3.4 for all attention layers; the paper provides no proof or systematic comparison of the two attention forms.
  • domain assumption KNN and radius graphs capture the local spatial structure needed for these PDEs.
    Section 3.1 and Section 4.3.2 choose graph construction per dataset to fit validation error, not from a physical first-principles rule.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GITO: Graph-Informed Transformer Operator for Learning Complex Partial Differential Equations." pith.science (2026). https://pith.science/paper/37AAUDIY

@misc{pith2026250613906,
  author       = {Pith},
  title        = {Pith review of: GITO: Graph-Informed Transformer Operator for Learning Complex Partial Differential Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/37AAUDIY}},
  note         = {Machine review of arXiv:2506.13906}
}
read the original abstract

We present a novel graph-informed transformer operator (GITO) architecture for learning complex partial differential equation systems defined on irregular geometries and non-uniform meshes. GITO consists of two main modules: a hybrid graph transformer (HGT) and a transformer neural operator (TNO). HGT leverages a graph neural network (GNN) to encode local spatial relationships and a transformer to capture long-range dependencies. A self-attention fusion layer integrates the outputs of the GNN and transformer to enable more expressive feature learning on graph-structured data. TNO module employs linear-complexity cross-attention and self-attention layers to map encoded input functions to predictions at arbitrary query locations, ensuring discretization invariance and enabling zero-shot super-resolution across any mesh. Empirical results on benchmark PDE tasks demonstrate that GITO outperforms existing transformer-based neural operators, paving the way for efficient, mesh-agnostic surrogate solvers in engineering applications.

Figures

Figures reproduced from arXiv: 2506.13906 by the authors.

Figure 1
Figure 1. Overall architecture of GITO. The input function and query points are first converted into graph representations [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (Top) The hybrid graph transformer (HGT) module consists of a GNN layer, a self-attention global layer, and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of GITO’s predictions against ground truth, and the corresponding absolute error plots for a test [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison between GITO’s predicted Mach number and the ground truth for a representative test sample [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of mesh points depicting the domain geometry of a single sample from a) the NS dataset and b) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Visualization of mesh points depicting the domain geometry of a single sample from the Airfoil dataset [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. DETNO: A Diffusion-Enhanced Transformer Neural Operator for Long-Term Traffic Forecasting

    cs.LG 2025-08 conditional novelty 5.0 of 10

    DETNO couples a transformer neural operator with a diffusion refiner, achieving lower rollout error and better high-frequency fidelity on synthetic LWR traffic forecasts than ONTraffic and GNOT.

Reference graph

Works this paper leans on

37 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    Springer, 2014

    Peter J Olver et al.Introduction to partial differential equations, volume 1. Springer, 2014

    work page 2014

  2. [2]

    Bayesian deep convolutional encoder–decoder networks for surrogate modeling and uncertainty quantification.Journal of Computational Physics, 366:415–447, 2018

    Yinhao Zhu and Nicholas Zabaras. Bayesian deep convolutional encoder–decoder networks for surrogate modeling and uncertainty quantification.Journal of Computational Physics, 366:415–447, 2018

    work page 2018

  3. [3]

    Prediction of aerodynamic flow fields using convolutional neural networks.Computational Mechanics, 64:525–545, 2019

    Saakaar Bhatnagar, Yaser Afshar, Shaowu Pan, Karthik Duraisamy, and Shailendra Kaushik. Prediction of aerodynamic flow fields using convolutional neural networks.Computational Mechanics, 64:525–545, 2019

    work page 2019

  4. [4]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics, 378:686–707, 2019

    2019

  5. [5]

    Gpt-pinn: Generative pre-trained physics-informed neural networks toward non-intrusive meta-learning of parametric pdes.Finite Elements in Analysis and Design, 228:104047, 2024

    Yanlai Chen and Shawn Koohy. Gpt-pinn: Generative pre-trained physics-informed neural networks toward non-intrusive meta-learning of parametric pdes.Finite Elements in Analysis and Design, 228:104047, 2024

    work page 2024

  6. [6]

    Milad Ramezankhani and Abbas S Milani. A sequential meta-transfer (smt) learning to combat complexities of physics-informed neural networks: Application to composites autoclave processing.Composites Part B: Engineering, 283:111597, 2024

    work page 2024

  7. [7]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature machine intelligence, 3(3): 218–229, 2021

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature machine intelligence, 3(3): 218–229, 2021

    2021

  8. [8]

    Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

    arXiv 2010

Show all 37 references
  1. [9]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  2. [10]

    Choose a transformer: Fourier or galerkin.Advances in neural information processing systems, 34: 24924–24940, 2021

    Shuhao Cao. Choose a transformer: Fourier or galerkin.Advances in neural information processing systems, 34: 24924–24940, 2021

    2021

  3. [11]

    Predicting physics in mesh-reduced space with temporal attention.arXiv preprint arXiv:2201.09113, 2022

    Xu Han, Han Gao, Tobias Pfaff, Jian-Xun Wang, and Li-Ping Liu. Predicting physics in mesh-reduced space with temporal attention.arXiv preprint arXiv:2201.09113, 2022

    2022 arXiv

  4. [12]

    Transformer for partial differential equations’ operator learning.arXiv preprint arXiv:2205.13671, 2022

    Zijie Li, Kazem Meidani, and Amir Barati Farimani. Transformer for partial differential equations’ operator learning.arXiv preprint arXiv:2205.13671, 2022

    2022 arXiv

  5. [13]

    Gnot: A general neural operator transformer for operator learning

    Zhongkai Hao, Zhengyi Wang, Hang Su, Chengyang Ying, Yinpeng Dong, Songming Liu, Ze Cheng, Jian Song, and Jun Zhu. Gnot: A general neural operator transformer for operator learning. InInternational Conference on Machine Learning, pages 12556–12569. PMLR, 2023

    2023

  6. [14]

    Universal physics transformers: A framework for efficiently scaling neural operators.Advances in Neural Information Processing Systems, 37:25152–25194, 2024

    Benedikt Alkin, Andreas Fürst, Simon Schmid, Lukas Gruber, Markus Holzleitner, and Johannes Brandstetter. Universal physics transformers: A framework for efficiently scaling neural operators.Advances in Neural Information Processing Systems, 37:25152–25194, 2024

    2024

  7. [15]

    Transolver: A fast transformer solver for pdes on general geometries.arXiv preprint arXiv:2402.02366, 2024

    Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for pdes on general geometries.arXiv preprint arXiv:2402.02366, 2024. 8 Graph-informed transformer operator for learning complex PDEsA PREPRINT

    2024 arXiv

  8. [16]

    Hamlet: Graph transformer neural operator for partial differential equations.arXiv preprint arXiv:2402.03541, 2024

    Andrey Bryutkin, Jiahao Huang, Zhongying Deng, Guang Yang, Carola-Bibiane Schönlieb, and Angelica Aviles-Rivero. Hamlet: Graph transformer neural operator for partial differential equations.arXiv preprint arXiv:2402.03541, 2024

    2024 arXiv

  9. [17]

    Transformers for modeling physical systems.Neural Networks, 146: 272–289, 2022

    Nicholas Geneva and Nicholas Zabaras. Transformers for modeling physical systems.Neural Networks, 146: 272–289, 2022

    2022

  10. [18]

    Continuous spatiotemporal transformers.arXiv preprint arXiv:2301.13338, 2023

    Antonio H de O Fonseca, Emanuele Zappala, Josue Ortega Caro, and David Van Dijk. Continuous spatiotemporal transformers.arXiv preprint arXiv:2301.13338, 2023

    2023 arXiv

  11. [19]

    Scalable transformer for pde surrogate modeling.Advances in Neural Information Processing Systems, 36:28010–28039, 2023

    Zijie Li, Dule Shu, and Amir Barati Farimani. Scalable transformer for pde surrogate modeling.Advances in Neural Information Processing Systems, 36:28010–28039, 2023

    2023

  12. [20]

    Positional knowledge is all you need: Position-induced transformer (pit) for operator learning.arXiv preprint arXiv:2405.09285, 2024

    Junfeng Chen and Kailiang Wu. Positional knowledge is all you need: Position-induced transformer (pit) for operator learning.arXiv preprint arXiv:2405.09285, 2024

    2024 arXiv

  13. [21]

    Message passing neural pde solvers.arXiv preprint arXiv:2202.03376, 2022

    Johannes Brandstetter, Daniel Worrall, and Max Welling. Message passing neural pde solvers.arXiv preprint arXiv:2202.03376, 2022

    2022 arXiv

  14. [22]

    Neural operator: Graph kernel network for partial differential equations.arXiv preprint arXiv:2003.03485, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Graph kernel network for partial differential equations.arXiv preprint arXiv:2003.03485, 2020

    2003 arXiv

  15. [23]

    Graph networks as learnable physics engines for inference and control

    Alvaro Sanchez-Gonzalez, Nicolas Heess, Jost Tobias Springenberg, Josh Merel, Martin Riedmiller, Raia Hadsell, and Peter Battaglia. Graph networks as learnable physics engines for inference and control. InInternational conference on machine learning, pages 4470–4479. PMLR, 2018

    2018

  16. [24]

    Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023

    2023

  17. [25]

    Graph neural networks exponentially lose expressive power for node classification

    Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. arXiv preprint arXiv:1905.10947, 2019

    1905 arXiv

  18. [26]

    On the bottleneck of graph neural networks and its practical implications.arXiv preprint arXiv:2006.05205, 2020

    Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications.arXiv preprint arXiv:2006.05205, 2020

    2006 arXiv

  19. [27]

    A generalization of transformer networks to graphs.arXiv preprint arXiv:2012.09699, 2020

    Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs.arXiv preprint arXiv:2012.09699, 2020

    2012 arXiv

  20. [28]

    Do transformers really perform badly for graph representation?Advances in neural information processing systems, 34:28877–28888, 2021

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. Do transformers really perform badly for graph representation?Advances in neural information processing systems, 34:28877–28888, 2021

    2021

  21. [29]

    Graphit: Encoding graph structure in transformers.arXiv preprint arXiv:2106.05667, 2021

    Grégoire Mialon, Dexiong Chen, Margot Selosse, and Julien Mairal. Graphit: Encoding graph structure in transformers.arXiv preprint arXiv:2106.05667, 2021

    2021 arXiv

  22. [30]

    Recipe for a general, powerful, scalable graph transformer.Advances in Neural Information Processing Systems, 35:14501–14515, 2022

    Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer.Advances in Neural Information Processing Systems, 35:14501–14515, 2022

    2022

  23. [31]

    Exphormer: Sparse transformers for graphs

    Hamed Shirzad, Ameya Velingker, Balaji Venkatachalam, Danica J Sutherland, and Ali Kemal Sinop. Exphormer: Sparse transformers for graphs. InInternational Conference on Machine Learning, pages 31613–31632. PMLR, 2023

    2023

  24. [32]

    Scaling physics-informed hard constraints with mixture- of-experts.arXiv preprint arXiv:2402.13412, 2024

    Nithin Chalapathi, Yiheng Du, and Aditi Krishnapriyan. Scaling physics-informed hard constraints with mixture- of-experts.arXiv preprint arXiv:2402.13412, 2024

    2024 arXiv

  25. [33]

    How attentive are graph attention networks?arXiv preprint arXiv:2105.14491, 2021

    Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks?arXiv preprint arXiv:2105.14491, 2021

    2021 arXiv

  26. [34]

    Fourier neural operator with learned deformations for pdes on general geometries.arXiv preprint arXiv:2207.05209, 2022

    Zongyi Li, Daniel Zhengyu Huang, Burigede Liu, and Anima Anandkumar. Fourier neural operator with learned deformations for pdes on general geometries.arXiv preprint arXiv:2207.05209, 2022

    2022 arXiv

  27. [35]

    Mionet: Learning multiple-input operators via tensor product.arXiv preprint arXiv:2202.06137, 2022

    Pengzhan Jin, Shuai Meng, and Lu Lu. Mionet: Learning multiple-input operators via tensor product.arXiv preprint arXiv:2202.06137, 2022

    2022 arXiv

  28. [36]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017

    2017

  29. [37]

    Elsevier, 2014

    Gerald Farin.Curves and surfaces for computer-aided geometric design: a practical guide. Elsevier, 2014. 9 Graph-informed transformer operator for learning complex PDEsA PREPRINT A Datasets and Model Hyperparameters. A.1 Datasets NS.We use a two-dimensional steady-state fluid ...

    2014

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.