Pith. sign in

REVIEW 3 major objections 8 minor 29 references

Projecting physical fields onto adaptive local bases yields Transformer tokens that solve PDEs on general meshes at state-of-the-art accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 12:45 UTC pith:3ZWSH3JI

load-bearing objection Solid engineering architecture for PDE Transformers with a real efficiency win on full-mesh decode; industrial SOTA claims need same-split verification before you trust the leaderboard. the 3 major comments →

arxiv 2607.24513 v1 pith:3ZWSH3JI submitted 2026-07-27 cs.LG

Physics Transformer: Tailoring Transformer for General PDE Prediction

Guoze Sun , Rui Zhang , Jiankai Tang , Mengtao Yan , Runze Mao , Zhi X. Chen , Hao Sun This is my paper
classification cs.LG
keywords Physics Transformerneural operatorsPDE predictionfunction projectionphysics tokensfactorized attentionirregular meshesCFD surrogates
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Physical fields are continuous functions sampled on meshes, not words or image patches, so a Transformer needs tokens that respect that functional structure. Physics Transformer partitions any discretization into locality-preserving patches, learns adaptive basis functions inside each patch, and projects the local field onto those bases to form compact physics tokens. Factorized attention then couples the same physical-state index across patches and mixes states inside each patch, after which the tokens decode back to fields at arbitrary query points. Across 2D dynamics, irregular multiphysics, and industrial 3D CFD, the method reports the best predictive accuracy, stable multi-step rollouts, and full-mesh inference from a modest support set with no accuracy loss. The practical claim is that function projection is a workable foundation for Transformer PDE solvers on general geometries.

Core claim

Function-projection tokenization inside locality-preserving patches, followed by factorized attention over aligned physics-token indices and within-patch states, produces a general Transformer operator that captures fine-grained physical structure and achieves state-of-the-art field prediction from 2D PDE dynamics through industrial-scale 3D CFD, including stable autoregressive rollouts and accurate full-mesh decoding from cached support tokens.

What carries the argument

Patch-wise physics tokenization: inside each Hilbert-ordered spatial patch a shared generator builds M adaptive basis functions; the local field is projected onto them to yield M physics tokens whose indices stay aligned across patches, enabling factorized space-then-state attention and arbitrary-query de-tokenization.

Load-bearing premise

That a shared basis generator keeps the same token index meaning the same latent physical state in every patch, so factorized attention across those indices is enough and denser coupling is unnecessary.

What would settle it

On a multiscale irregular-mesh task, force token indices to be unaligned or replace factorized attention with full attention over all patch-token pairs; if accuracy and rollout stability collapse relative to the reported tables, the alignment premise fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Function projection can replace global slice aggregation as the default way to tokenize continuous fields for Transformers.
  • Industrial CFD surrogates can train on ~100K support points and decode full multi-million-point meshes from cached tokens without accuracy loss.
  • Stable multi-step rollout on stiff multiphysics becomes achievable without geometry-specific architectures.
  • Complexity drops from quadratic in mesh size to costs linear in points plus quadratic only in patches and tokens per patch.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same local-projection idea could transfer to other continuous-field operators outside fluids, such as elasticity or electromagnetics on unstructured meshes.
  • If index alignment is imperfect, a light cross-state long-range path or learned index routing might close remaining gaps without restoring full quadratic cost.
  • Cached-token decoding suggests a natural path to interactive design loops where geometry changes only re-encode a small support set.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 8 minor

Summary. The manuscript proposes Physics Transformer, a Transformer-based neural operator for PDE field prediction on regular grids, irregular meshes, and point clouds. Points are serialized by a Hilbert space-filling curve and split into P locality-preserving patches; within each patch, a shared, input-adaptive generator produces M softmax-normalized local basis functions (tokenization and de-tokenization bases), onto which the local field is projected to yield M "physics tokens" per patch. Global interaction is factorized: attention across patches within each token index m, then attention across the M tokens within each patch. A Galerkin/Petrov–Galerkin interpretation is given (§4), including a rank argument that patch-local bases provide up to PM effective basis functions versus M for global slices. Experiments cover six benchmarks (TRL2D, Gray–Scott, REALM rocket combustor, F1-Airfoil, NASA-CRM, DrivAerML) with one-step and 20-step rollout errors, integrated Cd/Cl R² on DrivAerML, ablations on ordering/patch count/basis count/mesh scale, efficiency comparisons against Transolver-3, and a cached-token arbitrary-query mechanism that decodes full-resolution meshes from a 100K-point support set with no accuracy loss (Table 9).

Significance. If the results hold, this is a solid contribution to Transformer-based operator learning: it identifies a real limitation of global-slice tokenization (over-smoothing of local fine-scale structure) and offers a clean fix with a plausible Galerkin-style rationale. The strongest evidence is concrete and falsifiable: six benchmarks with reproduced baselines under matched protocols, ablations over patch count P, basis count M, ordering scheme, and mesh scale; cached-token full-mesh decoding with no measurable accuracy loss (Table 9); and reported 15.3×/7.5× full-resolution inference speedups over Transolver-3 on ~8M/~15M-cell meshes (Table 12), which addresses a genuine industrial pain point (accurate Cd/Cl quadrature requires full-surface evaluation). Training configurations are tabulated (Table 10) and official implementations are used for baselines, which supports reproducibility. The arbitrary-query mechanism (§3.5, Algorithm 1) is a practically useful addition beyond the accuracy story.

major comments (3)
  1. [§B.2 / Tables 3–4] §B.2 and Tables 3–4: the Transolver-3 and AB-UPT rows on NASA-CRM and DrivAerML are adopted from those papers, while the authors use their own ad hoc 400/50/50 split of DrivAerML (§B) because the dataset has no official partition. If the source papers used a different partition, metric variant, or training budget (per §B.2, Transolver-3 is trained 800/600 epochs vs. 500 here), those rows are not commensurable with the rest of the table. The authors already re-ran GeoTransolver precisely because its paper reported relative L1, so the same scrutiny should be applied to the adopted numbers. One mitigation: the reproduced, same-protocol GeoTransolver is stronger than the adopted Transolver-3 row on DrivAerML and NASA-CRM, so the headline ordering likely survives even if the adopted rows are dropped — but this should be made explicit. Requested: mark adopted rows in Tables 3–4, state the sour
  2. [Tables 3–4, §5.2] No multi-seed statistics are reported anywhere in the paper. This is load-bearing specifically for Table 4, where the headline integrated-coefficient result is C_d R² = 0.996 (a tie with GeoTransolver) and C_l = 0.993 vs. 0.991 — differences well below typical seed-to-seed variance for models of this size. The pointwise field-error margins over the strongest reproduced baseline (e.g., DrivAerML p_s 3.35 vs. 3.64) are comfortable enough that the main SOTA claim is probably robust, but the Table 4 wording ('achieves R² scores above 0.99') and the boldface ordering should be supported by at least 3-seed mean/std on the DrivAerML runs, or the claim should be softened to a statistical tie.
  3. [§3.2–3.3, §4.3] The central architectural premise — that a shared score projection W_s makes token index m semantically aligned across heterogeneous patches, so factorized attention (spatial attention per fixed m, then token attention per patch) is an adequate substitute for coupling all PM tokens — is asserted in §3.2–3.3 and underpins both the complexity story (§3.6) and the PM-basis-capacity interpretation (§4.3), but is never directly ablated. The provided ablations vary P, M, ordering, and mesh scale, but none compares against (i) full PM×PM attention at small scale, (ii) a variant without the cross-patch same-m stage, or (iii) a shuffled-index control that would test whether the same-m alignment matters at all. A controlled comparison on a small benchmark (e.g., TRL2D at P=512, M=32, where full PM attention is affordable) would either substantiate the factorization or bound what it gives up; eithe
minor comments (8)
  1. [§3.2] §3.2: the tokenization equation reads T_p = K^⊤_p V_p, but K_p is never defined; §4.2 writes the same step correctly as Φ^⊤_p V_p. Please fix the symbol in §3.2 (and check the similar 'sprojection W_s' typo in the following paragraph).
  2. [§5, captions, §B] §5 (first paragraph) and §5.1 ('Large-scale industrial CFD'): the model is referred to as 'SGNO' twice — an apparent leftover from another manuscript. Also 'DriverML' (§1, §B.2) and 'DrivAerML' are used inconsistently; 'papar' and 'Expect for Transolver3' in §B.3/§B.2; 'Trl2D' in Fig. 5 caption; 'traning'/'predication' in Fig. 3 caption; 'Deeper red color express' in Fig. 4 caption.
  3. [Tables 2–3] Table 2 caption does not state which metric is reported per benchmark (VRMSE for TRL2D/GSRD, normalized MSE for REALM per §B.1), nor whether values are percentages; Table 3 values are stated to be percentages only in §B.1. Please annotate the tables.
  4. [Tables 2, 5; abstract] Table 2: on Gray–Scott rollout, FFNO (0.47) beats Physics Transformer (0.49), and Table 5 shows BallTree's τ error (5.25) is slightly worse than Hilbert's (5.24) while the text claims BallTree gives a 'marginal improvement over Hilbert.' The abstract and §5.2 'state-of-the-art' phrasing should be qualified to 'best overall' rather than uniformly best.
  5. [§3.6] §3.6: the dominant attention term O(MP²·d_t) is quadratic in P, and the default P=1024 makes this non-trivial; Transolver-style global slices scale linearly in N·M. The efficiency tables (11–12) address wall-clock cost, but a brief FLOPs or scaling discussion clarifying when the P² term bites would make the complexity claim more precise.
  6. [§3.5, Appendix A] §3.5/Algorithm 1: for arbitrary queries, the de-tokenization weights q_q depend on optional query features a_q — please state what a_q is at previously unseen locations (geometry only? boundary conditions?) and what happens when it is unavailable, since the full-mesh DrivAerML decoding in Table 9 relies on this path.
  7. [§5.6, Figure 3] §5.6: 'increasing the inference scale from 100K to 1M improves the relative L2 error by less than 1%' is ambiguous (absolute percentage points or relative improvement?). Also Figure 3b appears to show saturation, not monotonic improvement — please align text and figure.
  8. [§5.7, Table 8] Missing citation context: Point Transformer V3's serialization, Erwin's ball-tree partitioning, and SpiderSolver are compared empirically in Table 8, which is welcome; however the comparison adapts PTv3/SpiderSolver with replaced heads — please report their parameter counts alongside the results (PTv3 'largest configuration' is much larger than 30M) so the comparison is interpretable.

Circularity Check

0 steps flagged

No significant circularity: standard supervised operator learning with empirical ablations, not a self-forcing derivation.

full rationale

Physics Transformer is an architecture paper: patch-wise function projection, factorized attention, and arbitrary-query decoding are design choices trained end-to-end against held-out simulation targets (VRMSE, normalized MSE, relative L2). Reported errors are not algebraically forced by fitted constants renamed as predictions. Section 4’s Galerkin/projected-operator write-up is an interpretive analogy (inspired by Cao 2021) that describes what one forward pass does conditionally on input-dependent bases; it does not claim a first-principles derivation that entails the SOTA tables. Hyperparameters P and M are selected by validation ablations (Tables 5–7), which is ordinary ML practice. Self-citations (e.g., REALM as a benchmark co-authored by overlapping authors) supply datasets, not load-bearing uniqueness theorems that forbid alternatives. Comparability concerns about adopted baseline numbers or ad hoc DrivAerML splits affect experimental fairness, not circularity of a derivation chain. No step reduces Eq. X to Eq. Y by construction or fits a quantity and relabels it a prediction.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 2 invented entities

Load-bearing content is architectural and empirical, not a short axiom list deriving a constant. Claims rest on standard operator-learning setup, locality of Hilbert patches, shared-index semantic alignment, sufficiency of factorized attention, and faithfulness of public/industrial CFD labels. Free parameters are the usual ML knobs (P, M, depth/width, sampling counts) selected by ablation rather than physical constants.

free parameters (5)
  • Number of patches P (default 1024 in main runs; 512 on TRL2D config table) = 1024 (main); task-dependent
    Controls spatial granularity and quadratic patch-attention cost; chosen via ablation Table 6, not derived.
  • Physics tokens per patch M (basis count) = 32 or 64
    Local projection rank; ablated 16–128 in Table 7; set to 32 or 64 by task complexity.
  • Support-set size for industrial training/inference (~100K points) = ~100K
    Subsampling level where accuracy saturates on DrivAerML (§5.6); practical fit to memory, not mesh-independent theorem.
  • Model width/depth/heads and training hyperparameters = e.g. L=6–16, C=128–192, AdamW, cosine LR
    L, H, C, LR, epochs, optimizer per Table 10; tuned for validation performance and parameter parity with baselines.
  • Hilbert lattice resolution 2^L and Trans-Hilbert axis permutations σ_ℓ
    Implementation choices for patch boundaries across layers (§3.1); not uniquely determined by physics.
axioms (7)
  • domain assumption Observed fields are samples of functions on a domain; operator learning seeks G mapping input samples to field samples.
    Standard neural-operator framing in §1 and §3 opening; underpins tokenization-as-projection.
  • domain assumption Hilbert (or grid-window) contiguous segments are sufficiently locality-preserving patches for local bases to be meaningful.
    §3.1 and §5.3; Morton vs Hilbert vs BallTree comparison assumes coherent patches improve tokens.
  • ad hoc to paper Shared projection weights make token index m comparable across patches (aligned latent states).
    Stated in §3.2; architectural hypothesis enabling factorized spatial attention per m.
  • ad hoc to paper Factorized attention (across patches per m, then across m per patch) captures needed long-range and local state coupling without full PM self-attention.
    §3.3 design choice; complexity claims and accuracy rest on this sufficiency.
  • ad hoc to paper Softmax-normalized score maps Φ_p, Ψ_p act as useful adaptive test/trial bases (Petrov–Galerkin analogy).
    §3.2 definitions and §4 interpretation; analogy to Cao (2021), not a proved convergent discretization.
  • domain assumption Nearest-support patch assignment plus query MLP weights suffice to decode unseen coordinates from cached tokens.
    §3.5 / Appendix A; empirical Table 9 supports on DrivAerML smoothness regime.
  • domain assumption Benchmark CFD/PDE datasets and official metrics are adequate proxies for predictive quality (including integrated Cd/Cl).
    §5 and Appendix B evaluation protocols.
invented entities (2)
  • Patch-wise physics tokens from input-adaptive local bases (Φ_p, Ψ_p) no independent evidence
    purpose: Compress each spatial patch into M latent physical-state coefficients for Transformer mixing and query decode.
    Core proposed representation; learned end-to-end, not an external physical particle/field.
  • Factorized Physics Transformer block (spatial MHSA per token index + token MHSA per patch) no independent evidence
    purpose: Replace dense point attention with structured global/local mixing over P×M tokens.
    Architectural module defined in §3.3; validated only via downstream errors.

pith-pipeline@v1.2.0-grok45-kimik3 · 27505 in / 4202 out tokens · 85029 ms · 2026-07-31T12:45:46.028465+00:00 · methodology

0 comments
read the original abstract

Transformer architectures have attracted increasing attention for solving partial differential equations (PDEs), owing to their flexibility in handling irregular discretizations and their ability to capture long-range physical dependencies. However, unlike discrete language tokens or fixed-resolution image patches, observed physical fields are finite samples of underlying infinite-dimensional functions. Consequently, effectively applying Transformers to PDEs requires a tokenizer that respects the functional nature of physical fields and constructs physically expressive tokens from arbitrary discretizations.To this end, we propose \methodname{Physics Transformer}, a function-projection-based Transformer architecture for physical field prediction. Physics Transformer treats a physical field as a continuous function and partitions its discretization into locality-preserving spatial patches. Within each patch, it dynamically learns a set of adaptive local basis functions and projects the sampled field onto these bases to obtain compact physics tokens. The resulting tokens capture diverse latent physical states while preserving fine-scale spatial structures, enabling efficient global interaction through factorized attention across space and physical states. The projected representation further supports efficient decoding at arbitrary query locations. Extensive experiments on diverse benchmarks, ranging from two-dimensional PDE dynamics to industrial-scale three-dimensional CFD simulations, demonstrate that Physics Transformer accurately captures fine-grained physical structures and achieves state-of-the-art predictive performance. These results establish function projection as a practical and effective foundation for designing Transformer architectures for PDE solving.

Figures

Figures reproduced from arXiv: 2607.24513 by Guoze Sun, Hao Sun, Jiankai Tang, Mengtao Yan, Rui Zhang, Runze Mao, Zhi X. Chen.

Figure 1
Figure 1. Figure 1: The main structure of Physics Transformer. A general overview is on top left. And a [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: We test Physics Transformer on six challenging tasks covering different discretizations [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: In part a, we show the satisfying predication accuracy of integrated quantities of Physics [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: We illustrate two representative attention scores from the last layer of Physics Transformer [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results on Trl2D dataset. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Results on GSDR dataset. 33 [PITH_FULL_IMAGE:figures/full_fig_p033_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Results on REALM dataset [PITH_FULL_IMAGE:figures/full_fig_p034_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Results on F1-Airfoil Dynamics dataset. 34 [PITH_FULL_IMAGE:figures/full_fig_p034_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Full results on NASA-CRM dataset. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Full results on DrivAerML dataset. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 4 linked inside Pith

  1. [1]

    Learning Nonlinear Operators via

    Lu, Lu and Jin, Pengzhan and Pang, Guofei and Zhang, Zhongqiang and Karniadakis, George Em , journal =. Learning Nonlinear Operators via. 2021 , doi =

    2021

  2. [2]

    International Conference on Learning Representations , year =

    Fourier Neural Operator for Parametric Partial Differential Equations , author =. International Conference on Learning Representations , year =

  3. [3]

    Neural Operator: Learning Maps Between Function Spaces With Applications to

    Kovachki, Nikola and Li, Zongyi and Liu, Burigede and Azizzadenesheli, Kamyar and Bhattacharya, Kaushik and Stuart, Andrew and Anandkumar, Anima , journal =. Neural Operator: Learning Maps Between Function Spaces With Applications to. 2023 , url =

    2023

  4. [4]

    arXiv preprint arXiv:2003.03485 , year =

    Neural Operator: Graph Kernel Network for Partial Differential Equations , author =. arXiv preprint arXiv:2003.03485 , year =

    Pith/arXiv arXiv 2003

  5. [5]

    Proceedings of the 38th International Conference on Machine Learning , year =

    Learning Mesh-Based Simulation with Graph Networks , author =. Proceedings of the 38th International Conference on Machine Learning , year =

  6. [6]

    Fourier Neural Operator with Learned Deformations for

    Li, Zongyi and Huang, Daniel Zhengyu and Liu, Burigede and Anandkumar, Anima , journal =. Fourier Neural Operator with Learned Deformations for. 2022 , url =

    2022

  7. [7]

    Geometry-Informed Neural Operator for Large-Scale 3D

    Li, Zongyi and Kovachki, Nikola Borislavov and Choy, Chris and Li, Boyi and Kossaifi, Jean and Otta, Shourya Prakash and Nabian, Mohammad Amin and Stadler, Maximilian and Hundt, Christian and Azizzadenesheli, Kamyar and Anandkumar, Anima , booktitle =. Geometry-Informed Neural Operator for Large-Scale 3D. 2023 , url =

    2023

  8. [8]

    arXiv preprint arXiv:2205.13671 , year =

    Transformer for Partial Differential Equations' Operator Learning , author =. arXiv preprint arXiv:2205.13671 , year =

    Pith/arXiv arXiv

  9. [9]

    2023 , url =

    Hao, Zhongkai and Wang, Zhengyi and Su, Hang and Ying, Chengyang and Dong, Yinpeng and Liu, Songming and Cheng, Ze and Song, Jian and Zhu, Jun , booktitle =. 2023 , url =

    2023

  10. [10]

    Advances in Neural Information Processing Systems , volume =

    Choose a Transformer: Fourier or Galerkin , author =. Advances in Neural Information Processing Systems , volume =. 2021 , url =

    2021

  11. [11]

    International Conference on Learning Representations , year =

    An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale , author =. International Conference on Learning Representations , year =

  12. [12]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =. 2021 , doi =

    2021

  13. [13]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Point Transformer V3: Simpler, Faster, Stronger , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2024 , doi =

    2024

  14. [14]

    Transolver: A Fast Transformer Solver for

    Wu, Haixu and Luo, Huakun and Wang, Haowen and Wang, Jianmin and Long, Mingsheng , booktitle =. Transolver: A Fast Transformer Solver for. 2024 , url =

    2024

  15. [15]

    Advances in Neural Information Processing Systems , volume =

    Universal Physics Transformers: A Framework for Efficiently Scaling Neural Operators , author =. Advances in Neural Information Processing Systems , volume =. 2024 , url =

    2024

  16. [16]

    International Conference on Learning Representations , year =

    Factorized Fourier Neural Operators , author =. International Conference on Learning Representations , year =

  17. [17]

    SpiderSolver: A Geometry-Aware Transformer for Solving

    Qi, Kai and Wang, Fan and Dong, Zhewen and Sun, Jian , booktitle =. SpiderSolver: A Geometry-Aware Transformer for Solving. 2025 , url =

    2025

  18. [18]

    Proceedings of the 42nd International Conference on Machine Learning , year =

    Erwin: A Tree-Based Hierarchical Transformer for Large-Scale Physical Systems , author =. Proceedings of the 42nd International Conference on Machine Learning , year =

  19. [19]

    Transolver++: An Accurate Neural Solver for

    Luo, Huakun and Wu, Haixu and Zhou, Hang and Xing, Lanxiang and Di, Yichen and Wang, Jianmin and Long, Mingsheng , booktitle =. Transolver++: An Accurate Neural Solver for. 2025 , url =

    2025

  20. [20]

    Geometry Aware Operator Transformer as an Efficient and Accurate Neural Surrogate for

    Wen, Shizheng and Kumbhat, Arsh and Lingsch, Levi and Mousavi, Sepehr and Zhao, Yizhou and Chandrashekar, Praveen and Mishra, Siddhartha , journal =. Geometry Aware Operator Transformer as an Efficient and Accurate Neural Surrogate for. 2025 , url =

    2025

  21. [21]

    2025 , url =

    Alkin, Benedikt and Bleeker, Maurits and Kurle, Richard and Kronlachner, Tobias and Sonnleitner, Reinhard and Dorfer, Matthias and Brandstetter, Johannes , journal =. 2025 , url =

    2025

  22. [22]

    arXiv preprint arXiv:2511.06294 , year =

    Transolver Is a Linear Transformer: Revisiting Physics-Attention through the Lens of Linear Attention , author =. arXiv preprint arXiv:2511.06294 , year =

    arXiv

  23. [23]

    2025 , url =

    Adams, Corey and Ranade, Rishikesh and Cherukuri, Ram and Choudhry, Sanjay , journal =. 2025 , url =

    2025

  24. [24]

    arXiv preprint arXiv:2602.04940 , year =

    Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries , author =. arXiv preprint arXiv:2602.04940 , year =

    Pith/arXiv arXiv

  25. [25]

    Advances in Neural Information Processing Systems , volume =

    The Well: A Large-Scale Collection of Diverse Physics Simulations for Machine Learning , author =. Advances in Neural Information Processing Systems , volume =. 2024 , url =

    2024

  26. [26]

    arXiv preprint arXiv:2512.18595 , year =

    Benchmarking Neural Surrogates on Realistic Spatiotemporal Multiphysics Flows , author =. arXiv preprint arXiv:2512.18595 , year =

    arXiv

  27. [27]

    2026 , howpublished =

    2026

  28. [28]

    arXiv preprint arXiv:2505.14704 , year =

    Towards Scalable Surrogate Models Based on Neural Fields for Large-Scale Aerodynamic Simulations , author =. arXiv preprint arXiv:2505.14704 , year =

    Pith/arXiv arXiv

  29. [29]

    2024 , url =

    Ashton, Neil and Mockett, Charles and Fuchs, Marian and Fliessbach, Louis and Hetmann, Hendrik and Knacke, Thilo and Schonwald, Norbert and Skaperdas, Vangelis and Fotiadis, Grigoris and Walle, Astrid and Hupertz, Burkhard and Maddix, Danielle , journal =. 2024 , url =

    2024