pith. sign in

arxiv: 2605.03548 · v2 · pith:SMJWUTUYnew · submitted 2026-05-05 · 💻 cs.LG · cs.AI

PerFlow: Physics-Embedded Rectified Flow for Efficient Reconstruction and Uncertainty Quantification of Spatiotemporal Dynamics

Hao Zhou , Rui Zhang , Han Wan , Hao Sun This is my paper

Pith reviewed 2026-05-19 16:47 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords rectified flowphysics-informed generative modelsspatiotemporal reconstructionuncertainty quantificationPDE constraintssparse measurementsfast samplingprojection methods
0
0 comments X

The pith

PerFlow embeds hard physics constraints into rectified flows to reconstruct sparse spatiotemporal fields quickly and with uncertainty estimates.

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

The paper targets the reconstruction of fields governed by partial differential equations from sparse and irregular measurements, where standard deterministic methods fail to capture uncertainty and generative models suffer from slow sampling due to combined guidance for data and physics. PerFlow decouples these by feeding observations straight into the rectified flow dynamics for conditioning while applying a separate projection step to enforce physical rules such as conservation or incompressibility. Theoretical invariance guarantees ensure that the generated trajectories remain on the physics-consistent manifold at every step. This design yields competitive accuracy on PDE systems alongside much faster inference, using roughly 50 steps instead of thousands.

Core claim

PerFlow performs guidance-free conditioning by feeding observations into rectified-flow dynamics and embeds hard physics via a constraint-preserving projection, with invariance guarantees ensuring trajectories remain on the physics-consistent manifold throughout sampling, which enables competitive reconstruction accuracy, sound physics consistency, efficient 50-step conditional sampling, and up to 320x faster inference than 2000-step guided diffusion baselines.

What carries the argument

The constraint-preserving projection applied after each rectified-flow step to enforce invariants like incompressibility or conservation while preserving the learned distribution.

If this is right

  • Competitive reconstruction accuracy on various PDE systems with maintained physics consistency.
  • Reliable uncertainty quantification arising from sampling the learned generative distribution.
  • Conditional sampling performed efficiently in around 50 steps rather than thousands.
  • Inference speedups reaching 320 times compared with guided diffusion baselines that require 2000 steps.

Where Pith is reading between the lines

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

  • The separation of conditioning and projection steps could transfer to other flow-based or diffusion models for enforcing hard constraints in different domains.
  • Real-world sensor networks with irregular spacing might benefit directly, provided the projection operator remains computationally cheap.
  • The invariance property opens a route to hybrid models that combine learned flows with traditional numerical integrators for long-horizon forecasting.

Load-bearing premise

The constraint-preserving projection can be applied after each rectified-flow step without distorting the learned distribution or violating the invariance guarantees that keep trajectories on the physics-consistent manifold.

What would settle it

An experiment showing that repeated application of the projection after flow steps produces samples that leave the target distribution or violate the claimed physics manifold invariance would disprove the central decoupling mechanism.

Figures

Figures reproduced from arXiv: 2605.03548 by Han Wan, Hao Sun, Hao Zhou, Rui Zhang.

Figure 1
Figure 1. Figure 1: (a) Relative ℓ2 error vs. sampling steps on Darcy (S3GM, DiffPDE, PerFlow). (b) DiffPDE sensitivity on Poisson: Relative ℓ2 error vs. PDE-loss weight λ. et al., 2020; Lu et al., 2021; Li et al., 2024b]. Most exist￾ing surrogate models are deterministic and trained on grid￾ded inputs (e.g., complete initial conditions), yet real-world measurement data is often sparse and irregular, rendering field reconstru… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of PerFlow. (a) PerFlow learns a rectified-flow transport between initial and target distributions on the physics-consistent manifold. (b) Training and inference pipelines. During training, sparse observations are formed by masking ground-truth fields, while prior￾projected noise is combined with the target field using weights (1 − t) and t to construct xt. PerFlow learns a conditional velocity fi… view at source ↗
Figure 3
Figure 3. Figure 3: Sparse reconstruction physical fields (left) and the distribution of the point-wise mean view at source ↗
Figure 4
Figure 4. Figure 4: Uncertainty quantification under sparse observations, including Darcy coefficient view at source ↗
Figure 5
Figure 5. Figure 5: Reconstruction error versus sampling steps on ( view at source ↗
Figure 6
Figure 6. Figure 6: Poisson data-scaling study for PerFlow: test Rel- view at source ↗
Figure 7
Figure 7. Figure 7: Sparse reconstruction on four PDE benchmarks: view at source ↗
Figure 8
Figure 8. Figure 8: Darcy flow reconstruction under varying numbers of observations ( view at source ↗
read the original abstract

Reconstructing PDE-governed fields from sparse and irregular measurements is challenging due to their ill-posed nature. Deterministic surrogates are trained on dense fields that struggle with limited measurements and uncertainty quantification. Generative models, by learning distributions over spatiotemporal fields, can better handle sparsity and uncertainty. However, existing generative approaches enforce data consistency and PDE constraints simultaneously via sampling-time gradient guidance, resulting in slow and unstable inference. To this end, we propose PerFlow, a Physics-embedded rectified Flow for efficient sparse reconstruction and uncertainty quantification of spatiotemporal dynamics. PerFlow decouples observation conditioning from physics enforcement, performing guidance-free conditioning by feeding observations into rectified-flow dynamics while embedding hard physics via a constraint-preserving projection (e.g., incompressibility or conservation). Theoretically, we establish invariance guarantees to ensure that trajectories remain on the physics-consistent manifold throughout sampling. Experiments on various PDE systems demonstrate competitive reconstruction accuracy with sound physics consistency, while enabling efficient conditional sampling (e.g., 50 steps) and up to 320x faster inference than 2000-step guided diffusion baselines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes PerFlow, a physics-embedded rectified flow for reconstructing PDE-governed spatiotemporal fields from sparse measurements. It decouples observation conditioning (via guidance-free feeding of observations into rectified-flow dynamics) from physics enforcement (via post-step constraint-preserving projections such as for incompressibility), while claiming invariance guarantees that keep sampling trajectories on the physics-consistent manifold. This is asserted to yield competitive accuracy, sound physics consistency, 50-step conditional sampling, and up to 320x faster inference than 2000-step guided diffusion baselines.

Significance. If the invariance guarantees survive the projections without distorting the learned velocity field or breaking trajectory straightness, PerFlow would represent a meaningful efficiency gain for physics-informed generative modeling of PDEs, enabling practical uncertainty quantification and reconstruction where guided diffusion is too slow.

major comments (2)
  1. [§4] §4 (Invariance Guarantees): The central claim that trajectories remain on the physics-consistent manifold after each constraint-preserving projection rests on unshown conditions (linearity, commutativity with the flow, or measure preservation). Without these, the projection risks curving the straight-line paths that rectified flows rely on for few-step sampling, directly undermining the advertised 50-step efficiency and 320x speedup.
  2. [§5] §5 (Experiments): The reported competitive accuracy and physics consistency lack error bars, dataset sizes, or ablation on projection frequency; the 320x inference comparison to 2000-step baselines is therefore difficult to evaluate and is load-bearing for the efficiency claim.
minor comments (1)
  1. [§3] Notation for the projection operator P and its integration after each rectified-flow step is introduced without a clarifying diagram or pseudocode, making the decoupling of conditioning and physics enforcement harder to follow.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful comments on our manuscript. We have carefully considered each point and provide detailed responses below. Where appropriate, we will revise the manuscript to address the concerns raised.

read point-by-point responses

  1. Referee: [§4] §4 (Invariance Guarantees): The central claim that trajectories remain on the physics-consistent manifold after each constraint-preserving projection rests on unshown conditions (linearity, commutativity with the flow, or measure preservation). Without these, the projection risks curving the straight-line paths that rectified flows rely on for few-step sampling, directly undermining the advertised 50-step efficiency and 320x speedup.

    Authors: We appreciate this observation regarding the theoretical foundations. The invariance guarantees in Section 4 are established under the assumption that the constraint-preserving projections are linear operators that commute with the rectified flow vector field and preserve the relevant measure on the manifold. In the revised version, we will explicitly list these conditions and provide a more detailed proof that demonstrates how these properties ensure the sampling trajectories remain straight lines on the physics-consistent manifold. This clarification will reinforce that the few-step sampling efficiency is indeed preserved. revision: yes

  2. Referee: [§5] §5 (Experiments): The reported competitive accuracy and physics consistency lack error bars, dataset sizes, or ablation on projection frequency; the 320x inference comparison to 2000-step baselines is therefore difficult to evaluate and is load-bearing for the efficiency claim.

    Authors: We agree that the experimental section would benefit from additional details for reproducibility and evaluation. In the revised manuscript, we will add error bars computed over multiple independent runs, explicitly report the sizes of the training and test datasets for each PDE system, and include an ablation study varying the projection frequency during the sampling process. Furthermore, we will provide more precise details on the baseline implementations and hardware used for the inference time comparisons to better substantiate the reported speedups. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation builds on external rectified-flow and projection primitives without self-referential reduction

full rationale

The paper's core construction decouples observation conditioning (via direct injection into rectified-flow dynamics) from physics enforcement (via post-step constraint-preserving projection) and claims invariance guarantees that keep trajectories on the manifold. No quoted equation or step reduces a claimed performance gain or invariance property to a fitted parameter or self-citation by construction; the method explicitly references prior rectified-flow literature as an independent base and presents the projection as an added operator whose compatibility is asserted via new theoretical guarantees rather than tautology. Experiments on PDE systems supply the empirical support for efficiency and accuracy claims, keeping the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review based on abstract only; no explicit free parameters, axioms, or invented entities are detailed enough to enumerate. The projection operator and invariance guarantees are presented as core but unexpanded components.

pith-pipeline@v0.9.0 · 5722 in / 1130 out tokens · 49220 ms · 2026-05-19T16:47:09.171287+00:00 · methodology

Review history (3 revisions) →

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages · 8 internal anchors

  1. [1]

    Neural operators for accelerating scientific simulations and design.Nature Reviews Physics, 6(5):320–328,

    [Azizzadenesheliet al., 2024 ] Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Miguel Liu-Schiaffini, Jean Kossaifi, and Anima Anandkumar. Neural operators for accelerating scientific simulations and design.Nature Reviews Physics, 6(5):320–328,

    work page 2024

  2. [2]

    [Bhaganagar and Chambers, 2025] Kiran Bhaganagar and David Chambers. Accelerated elliptical pde solver for computational fluid dynamics based on configurable u-net architecture: Analogy to v-cycle multigrid.Machine Intel- ligence Research, 22(2):324–336,

    work page 2025

  3. [3]

    Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communications, 15(1):10416,

    [Duet al., 2024 ] Pan Du, Meet Hemant Parikh, Xiantao Fan, Xin-Yang Liu, and Jian-Xun Wang. Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communications, 15(1):10416,

    work page 2024

  4. [4]

    Pde-gcn: Novel architectures for graph neu- ral networks motivated by partial differential equa- tions.Advances in neural information processing systems, 34:3836–3849,

    [Eliasofet al., 2021 ] Moshe Eliasof, Eldad Haber, and Eran Treister. Pde-gcn: Novel architectures for graph neu- ral networks motivated by partial differential equa- tions.Advances in neural information processing systems, 34:3836–3849,

    work page 2021

  5. [5]

    Gen- erative adversarial networks.Communications of the ACM, 63(11):139–144,

    [Goodfellowet al., 2020 ] Ian Goodfellow, Jean Pouget- Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gen- erative adversarial networks.Communications of the ACM, 63(11):139–144,

    work page 2020

  6. [6]

    Geom-deeponet: A point-cloud- based deep operator network for field predictions on 3d parameterized geometries.Computer Methods in Applied Mechanics and Engineering, 429:117130,

    [Heet al., 2024 ] Junyan He, Seid Koric, Diab Abueidda, Ali Najafi, and Iwona Jasiuk. Geom-deeponet: A point-cloud- based deep operator network for field predictions on 3d parameterized geometries.Computer Methods in Applied Mechanics and Engineering, 429:117130,

    work page 2024

  7. [7]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851,

    [Hoet al., 2020 ] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851,

    work page 2020

  8. [8]

    [Holton and Hakim, 2013] James R Holton and Gregory J Hakim.An introduction to dynamic meteorology, vol- ume

    work page 2013

  9. [9]

    Diffusionpde: Generative pde-solving under partial observation.Advances in Neu- ral Information Processing Systems, 37:130291–130323,

    [Huanget al., 2024 ] Jiahe Huang, Guandao Yang, Zichen Wang, and Jeong Joon Park. Diffusionpde: Generative pde-solving under partial observation.Advances in Neu- ral Information Processing Systems, 37:130291–130323,

    work page 2024

  10. [10]

    Auto-Encoding Variational Bayes

    [Kingma and Welling, 2013] Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  11. [11]

    John Wiley & Sons,

    [Lapidus and Pinder, 1999] Leon Lapidus and George F Pin- der.Numerical solution of partial differential equations in science and engineering. John Wiley & Sons,

    work page 1999

  12. [12]

    Fourier Neural Operator for Parametric Partial Differential Equations

    [Liet al., 2020 ] Zongyi Li, Nikola Kovachki, Kamyar Az- izzadenesheli, Burigede Liu, Kaushik Bhattacharya, An- drew Stuart, and Anima Anandkumar. Fourier neural op- erator for parametric partial differential equations.arXiv preprint arXiv:2010.08895,

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  13. [13]

    Scalable transformer for pde surrogate model- ing.Advances in Neural Information Processing Systems, 36:28010–28039,

    [Liet al., 2023 ] Zijie Li, Dule Shu, and Amir Barati Fa- rimani. Scalable transformer for pde surrogate model- ing.Advances in Neural Information Processing Systems, 36:28010–28039,

    work page 2023

  14. [14]

    Generative latent neural PDE solver using flow matching.arXiv preprint arXiv:2503.22600, 2025

    [Liet al., 2025 ] Zijie Li, Anthony Zhou, and Amir Barati Fa- rimani. Generative latent neural pde solver using flow matching.arXiv preprint arXiv:2503.22600,

    work page arXiv 2025

  15. [15]

    Flow Matching for Generative Modeling

    [Lipmanet al., 2022 ] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  16. [16]

    Pde-refiner: Achieving accurate long rollouts with neural pde solvers.Advances in Neural Information Processing Systems, 36:67398–67433,

    [Lippeet al., 2023 ] Phillip Lippe, Bas Veeling, Paris Perdikaris, Richard Turner, and Johannes Brandstetter. Pde-refiner: Achieving accurate long rollouts with neural pde solvers.Advances in Neural Information Processing Systems, 36:67398–67433,

    work page 2023

  17. [17]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    [Liuet al., 2022 ] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003,

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  18. [18]

    Learn- ing nonlinear operators via deeponet based on the univer- sal approximation theorem of operators.Nature machine intelligence, 3(3):218–229,

    [Luet al., 2021 ] Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learn- ing nonlinear operators via deeponet based on the univer- sal approximation theorem of operators.Nature machine intelligence, 3(3):218–229,

    work page 2021

  19. [19]

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic mod- els.Machine Intelligence Research, 22(4):730–751,

    [Luet al., 2025 ] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic mod- els.Machine Intelligence Research, 22(4):730–751,

    work page 2025

  20. [20]

    arXiv preprint arXiv:2505.18190 , year=

    [Maet al., 2025 ] Yuezhou Ma, Haixu Wu, Hang Zhou, Huikun Weng, Jianmin Wang, and Mingsheng Long. Phy- sense: Sensor placement optimization for accurate physics sensing.arXiv preprint arXiv:2505.18190,

    work page internal anchor Pith review arXiv 2025

  21. [21]

    A., Ross, Z

    [Rahmanet al., 2022 ] Md Ashiqur Rahman, Zachary E Ross, and Kamyar Azizzadenesheli. U-no: U-shaped neu- ral operators.arXiv preprint arXiv:2204.11127,

    work page arXiv 2022

  22. [22]

    Variational inference with normalizing flows

    [Rezende and Mohamed, 2015] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. InInternational conference on machine learning, pages 1530–1538. PMLR,

    work page 2015

  23. [23]

    U-net: Convolutional networks for biomedical image segmentation

    [Ronnebergeret al., 2015 ] Olaf Ronneberger, Philipp Fis- cher, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer- assisted intervention, pages 234–241. Springer,

    work page 2015

  24. [24]

    On conditional diffusion models for pde simula- tions.Advances in Neural Information Processing Sys- tems, 37:23246–23300,

    [Shysheyaet al., 2024 ] Aliaksandra Shysheya, Cristiana Di- aconu, Federico Bergamin, Paris Perdikaris, Jos ´e Miguel Hern´andez-Lobato, Richard Turner, and Emile Math- ieu. On conditional diffusion models for pde simula- tions.Advances in Neural Information Processing Sys- tems, 37:23246–23300,

    work page 2024

  25. [25]

    Denoising Diffusion Implicit Models

    [Songet al., 2020a ] Jiaming Song, Chenlin Meng, and Ste- fano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502,

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  26. [26]

    Score-Based Generative Modeling through Stochastic Differential Equations

    [Songet al., 2020b ] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Er- mon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456,

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  27. [27]

    Chapman and Hall/CRC,

    [Strogatz, 2024] Steven H Strogatz.Nonlinear dynamics and chaos: with applications to physics, biology, chemistry, and engineering. Chapman and Hall/CRC,

    work page 2024

  28. [28]

    A review of numerical meth- ods for nonlinear partial differential equations.Bulletin of the American Mathematical Society, 49(4):507–554,

    [Tadmor, 2012] Eitan Tadmor. A review of numerical meth- ods for nonlinear partial differential equations.Bulletin of the American Mathematical Society, 49(4):507–554,

    work page 2012

  29. [29]

    Factorized fourier neural operators

    [Tranet al., 2023 ] Alasdair Tran, Alexander Mathews, Lex- ing Xie, and Cheng Soon Ong. Factorized fourier neural operators. InThe Eleventh International Conference on Learning Representations,

    work page 2023

  30. [30]

    Elsevier,

    [Tuet al., 2023 ] Jiyuan Tu, Guan Heng Yeoh, Chaoqun Liu, and Yao Tao.Computational fluid dynamics: a practical approach. Elsevier,

    work page 2023

  31. [31]

    Pesanet: Physics-encoded spectral attention network for simulating pde-governed complex systems

    [Wanet al., 2025 ] Han Wan, Rui Zhang, Qi Wang, Yang Liu, and Hao Sun. Pesanet: Physics-encoded spectral attention network for simulating pde-governed complex systems. In James Kwok, editor,Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25, pages 7751–7759. International Joint Conferences on Artificial In...

    work page 2025

  32. [32]

    [Wanget al., 2025 ] Sifan Wang, Zehao Dou, Siming Shan, Tong-Rui Liu, and Lu Lu

    Main Track. [Wanget al., 2025 ] Sifan Wang, Zehao Dou, Siming Shan, Tong-Rui Liu, and Lu Lu. Fundiff: Diffusion models over function spaces for physics-informed generative modeling. arXiv preprint arXiv:2506.07902,

    work page arXiv 2025

  33. [33]

    Diffphycon: A generative approach to control complex physical sys- tems.Advances in Neural Information Processing Sys- tems, 37:4090–4147,

    [Weiet al., 2024 ] Long Wei, Peiyan Hu, Ruiqi Feng, Haodong Feng, Yixuan Du, Tao Zhang, Rui Wang, Yue Wang, Zhi-Ming Ma, and Tailin Wu. Diffphycon: A generative approach to control complex physical sys- tems.Advances in Neural Information Processing Sys- tems, 37:4090–4147,

    work page 2024

  34. [34]

    Transolver: A Fast Transformer Solver for PDEs on General Geometries

    [Wuet al., 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,

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    Koopman neural operator as a mesh-free solver of non-linear partial differential equations.Journal of Computational Physics, 513:113194,

    [Xionget al., 2024 ] Wei Xiong, Xiaomeng Huang, Ziyang Zhang, Ruixuan Deng, Pei Sun, and Yang Tian. Koopman neural operator as a mesh-free solver of non-linear partial differential equations.Journal of Computational Physics, 513:113194,

    work page 2024

  36. [36]

    Geodiff: A geo- metric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923, 2022

    [Xuet al., 2022 ] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. arXiv preprint arXiv:2203.02923,

    work page arXiv 2022

  37. [37]

    Phympgn: Physics-encoded message passing graph network for spatiotemporal pde systems.arXiv preprint arXiv:2410.01337,

    [Zenget al., 2024 ] Bocheng Zeng, Qi Wang, Mengtao Yan, Yang Liu, Ruizhi Chengze, Yi Zhang, Hongsheng Liu, Zi- dong Wang, and Hao Sun. Phympgn: Physics-encoded message passing graph network for spatiotemporal pde systems.arXiv preprint arXiv:2410.01337,

    work page arXiv 2024

  38. [38]

    A survey of sparse rep- resentation: algorithms and applications.IEEE access, 3:490–530,

    [Zhanget al., 2015 ] Zheng Zhang, Yong Xu, Jian Yang, Xuelong Li, and David Zhang. A survey of sparse rep- resentation: algorithms and applications.IEEE access, 3:490–530,

    work page 2015

  39. [39]

    Deciphering and integrating invariants for neural operator learning with various physical mechanisms.National Sci- ence Review, 11(4):nwad336,

    [Zhanget al., 2024 ] Rui Zhang, Qi Meng, and Zhi-Ming Ma. Deciphering and integrating invariants for neural operator learning with various physical mechanisms.National Sci- ence Review, 11(4):nwad336,

    work page 2024

  40. [40]

    Monte carlo neural pde solver for learning pdes via probabilistic representation.IEEE Transactions on Pattern Analysis and Machine Intelligence,

    [Zhanget al., 2025 ] Rui Zhang, Qi Meng, Rongchan Zhu, Yue Wang, Wenlei Shi, Shihua Zhang, Zhi-Ming Ma, and Tie-Yan Liu. Monte carlo neural pde solver for learning pdes via probabilistic representation.IEEE Transactions on Pattern Analysis and Machine Intelligence,

    work page 2025

  41. [41]

    epochs Learning rate 1D Burgers 24 3001×10 −4 2D Darcy 24 5001×10 −4 2D Poisson 24 5001×10 −4 2D NS 10 8001×10 −4 Table 3: Training hyperparameters for different cases

    Case Batch size Num. epochs Learning rate 1D Burgers 24 3001×10 −4 2D Darcy 24 5001×10 −4 2D Poisson 24 5001×10 −4 2D NS 10 8001×10 −4 Table 3: Training hyperparameters for different cases. We optimize all models using AdamW with weight decay1×10 −4. We adopt a learning-rate schedule with 10-epoch warmup followed by cosine decay from1×10 −4 to6×10 −5. We ...

    work page 2020

  42. [42]

    Bothfanduhave resolution 128×128

    to the resulting solution field, consistent with [Huanget al., 2024 ]. Bothfanduhave resolution 128×128. Burgers’ equation.We study the one-dimensional viscous Burgers’ equation onΩ = (0,1)with periodic boundary condi- tions and viscosityν= 0.01: ∂u(x, t) ∂t + ∂ ∂x u(x, t)2 2 =ν ∂2u(x, t) ∂x2 , x∈Ω, t∈(0, T].(23) The initial condition isu(x,0) =u 0(x). Fo...

    work page 2024