Pith. sign in

REVIEW 5 major objections 5 minor 64 references

OmniFluids: Physics Pre-trained Modeling of Fluid Dynamics

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A neural operator pre-trained purely on the governing equations—no simulation data—outperforms data-hungry baselines on 2D and 3D turbulence benchmarks while running 10–100× faster than direct numerical simulation.

desk verdict Physics-only pretraining for neural operators is a real step forward, but the 3D long-horizon claims are ahead of the evidence. read the letter →

arxiv 2506.10862 v2 pith:OAP2EVAN submitted 2025-06-12 physics.flu-dyn cs.LG

classification physics.flu-dyncs.LG
keywords neuraloperatorphysics-informedlearningpre-trainedmodelfluiddynamicsturbulenceFourierdistillationfew-shotfine-tuning
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

OmniFluids aims to show that a neural operator for fluid dynamics can be pre-trained from the governing equations alone, with no simulation data, and still beat data-hungry baselines on 2D and 3D benchmarks. The training pipeline has three stages: physics-only pre-training on a Crank-Nicolson pseudo-spectral residual, distillation to a coarse-grid student, and few-shot fine-tuning. The paper reports that this recipe outperforms PINO, FNO, FFNO, FactFormer, CNO, DeepONet, and DPOT in accuracy, turbulence statistics, and long-horizon stability, while running 10–100 times faster than direct numerical simulation. It also recovers unknown viscosity and forcing fields from two noisy frames of the 2D Navier-Stokes equations. If these results hold, a unified CFD surrogate can be built without large labeled datasets.

What carries the argument

The load-bearing components are the mixture of operators (MoO), the multi-frame decoder, and factorized Fourier layers. MoO keeps a bank of expert operator branches selected by a routing network on the PDE parameters, then collapses them into one personalized operator for inference. The multi-frame decoder predicts $K$ future states in one forward pass, which makes the physics loss cheap to evaluate in parallel; the physics loss itself is a mean-squared Crank-Nicolson pseudo-spectral residual (Eq. 11) over all $K$ frames. Factorized Fourier layers decompose the spectral transform into separable 1D transforms along each axis, reducing parameter cost from $O(d^3)$ to $O(d)$ in 3D.

What would settle it

A decisive test would be to pre-train OmniFluids on the same equations and then evaluate the physics residual of its long-horizon rollouts against DNS: if the predicted fields drift from the reference while the residual stays near zero, the training signal is not sufficient. Concretely, one could run the model on a 2D KSE or NSE test case to $t = 10$ and check whether the relative $\ell^2$ error grows while the Crank-Nicolson residual of the predicted trajectory remains small, or whether the kinetic energy spectrum diverges from the reference at high wavenumbers.

Watch

Extended reading notes

Core claim

The paper's central claim is that a physics-only pre-training objective—minimizing the Crank-Nicolson time-discretized PDE residual over $K$ parallel frames—is sufficient to learn a transferable fluid operator. After distillation to a coarse grid and few-shot fine-tuning, the resulting student operator predicts 2D KSE, 2D NSE, and 3D NSE flows with lower relative $\ell^2$ error, better kinetic energy spectra, and higher temporal correlation than state-of-the-art data-driven and physics-informed baselines, while preserving turbulence statistics and running orders of magnitude faster than DNS. The same differentiable operator solves inverse problems, inferring viscosity and external forcing from two noisy frames.

Load-bearing premise

The whole approach rests on the assumption that minimizing a short-horizon, discretized physics residual (over only $t \in [0, 0.2]$) drives the network toward the true fluid solution operator, rather than toward one of the many near-zero-residual trajectories that chaotic equations admit.

Editorial extensions

If this is right

  • A physics-only pre-trained operator can be adapted to new Reynolds numbers and forcing conditions with as few as 2–10 trajectories.
  • Turbulence statistics such as the kinetic energy spectrum and vorticity or velocity PDFs are preserved over long horizons, matching DNS in the tested regimes.
  • The same framework supplies an efficient inverse solver, recovering viscosity and forcing from two noisy frames.
  • Inference on coarse grids yields 10–100× speedup over DNS at comparable accuracy, eliminating the need for fine spatiotemporal discretization.
  • Multi-frame parallel prediction removes the sequential time-stepping bottleneck of physics-informed training, enabling fully parallel loss computation.

Reading between the lines

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

  • If the residual objective is truly sufficient, the method could extend to other PDE families (e.g., reacting flows or magnetohydrodynamics) without data, provided a pseudo-spectral discretization is available.
  • The MoO routing on PDE parameters suggests a path toward conditional foundation models where one checkpoint covers a continuum of physical regimes, though the current test suite spans only KSE and NSE.
  • A stress test worth running is probing sensitivity to the pretraining horizon: if extending the training window beyond $t \in [0, 0.2]$ degrades rather than improves long-horizon error, the claim that the short-horizon residual encodes the operator would need qualification.
  • The inverse-problem success depends on the student operator being differentiable and accurate on coarse grids; a natural extension is joint estimation of missing equation terms, which the paper lists as future work.
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

5 major / 5 minor

Summary. The paper proposes OmniFluids, a neural operator architecture combining a mixture-of-operators (MoO), a multi-frame decoder, and factorized Fourier layers, and trains it in three stages: physics-only pre-training (minimizing a Crank-Nicolson pseudo-spectral PDE residual), operator distillation to a coarse-grid student, and few-shot fine-tuning on 2-10 trajectories. Experiments cover 2D KSE, 2D incompressible NSE, and 3D incompressible NSE, comparing against PINO, FNO, FFNO, FactFormer, CNO, DeepONet, and DPOT, as well as DNS. The authors report lower long-horizon errors, better turbulence statistics, 10-100x speedups over DNS, and successful recovery of viscosity and forcing from two noisy frames. The central claim is that physics-only pretraining yields a broadly transferable fluid operator with minimal data.

Significance. If established, the paper would be a meaningful step toward data-free foundation models for CFD, addressing a well-recognized bottleneck of existing neural operators that require large labeled datasets. The three-stage training pipeline is a sensible combination of existing ideas (physics-informed losses, knowledge distillation, fine-tuning), and the architectural components are individually ablated. The authors provide code and data availability, which supports reproducibility. However, the current evidence is not conclusive: several comparisons are not apples-to-apples, the DNS baseline is under-resolved, and the long-horizon extrapolation from a short-horizon training loss is not analyzed. The work is potentially significant, but the claims outrun what the experiments currently demonstrate.

major comments (5)
  1. [3D velocity-pressure NSE system; Supplementary Note B.2] The 3D comparison is not matched in spectral resolution: OmniFluids uses 64 modes per axis while PINO uses only 32 (Supplementary Note B.2: 'OmniFluids employs 64 modes per axis in 3D settings; however, for PINO... we use only 32 modes'). Since the number of Fourier modes directly controls representable frequency content, the reported superiority over PINO in Figs. 2c and 5 may be an artifact of this asymmetry. The authors should either run PINO with the same number of modes (if feasible) or clearly justify why the comparison is still fair despite different spectral truncations.
  2. [2D vorticity-velocity NSE system; 3D velocity-pressure NSE system; Figs. 4g and 5f] The claimed 10-100x speedups over 'DNS' are computed against DNS run at the same coarse spatial resolution as the neural operator (e.g., 128x128 for 2D NSE and 128^3 for 3D NSE) and with the largest stable time step. At these resolutions the DNS is under-resolved and likely inaccurate, so 'maintaining comparable accuracy' does not compare against a reliable solver. Please compare against a converged/resolved DNS, or clearly state that the speedup is relative to an under-resolved reference, and report errors against the high-resolution ground truth.
  3. [Physics-only pre-training; Eq. (11)] The physics loss L_PDE in Eq. (11) is minimized over t in [0,0.2], but the central evaluation extends predictions to t=5 (KSE) or t=10 (NSE), i.e., 25-50x beyond the training horizon. The paper provides no analysis of why a short-horizon Crank-Nicolson residual should produce a stable long-horizon operator, nor any evidence that low-residual solutions correspond to true trajectories (e.g., comparison of invariant measures, Lyapunov spectra, or sensitivity of the learned operator to the training horizon length). This is a load-bearing gap for the 'pure physics pre-training' claim and should be addressed with targeted experiments or analysis.
  4. [Results; Figs. 2-5; Appendix Data Table 1] All headline accuracy numbers and error-evolution curves appear to come from a single training run, and the test sets are small (10 trajectories for the 2D cases, 4 for the 3D case), with no error bars or multiple seeds. The claims 'consistently outperforms' and 'stable' are therefore not statistically substantiated. Please report means and standard deviations over at least 3-5 independent seeds and, where feasible, evaluate on a larger test set.
  5. [3D incompressible velocity-pressure NSE; Eq. (3)] The 3D NSE benchmark in Eq. (3) contains a non-standard linear damping term 0.1u. The paper describes the setup as '3D turbulence' with Re up to 3000, but this Rayleigh-friction-like term can substantially suppress small-scale fluctuations and reduce chaoticity, which may explain the unusually long correlation above 0.8. The authors should either remove or justify this term, and demonstrate that the simulated regime is genuinely turbulent (e.g., via an integral-scale Reynolds number, energy cascade, or comparison to standard forced NSE).
minor comments (5)
  1. [Eq. (11)] The notation 'u_0' is used ambiguously in the definition of L_PDE; clarify that it denotes the input initial condition provided to the multi-frame decoder.
  2. [Figs. 3e, 4g, 5f] The speed-accuracy plots use marker size to represent relative l2 error, but no legend with numeric values is provided; add a legend or annotate the markers.
  3. [Supplementary Note B.2] The temperature hyperparameter tau in Eq. (9), the number of experts P, and the number of layers/channels for the teacher and student are not fully specified; include these in the hyperparameter description for reproducibility.
  4. [Inverse Problem; Fig. 6] The inverse experiments report a single run per noise level with no confidence intervals over random initializations of the forcing field; add multiple runs to quantify variability.
  5. [3D NSE results] The text contains the typo 'decor-related' where 'decorrelated' is intended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: OmniFluids' physics-only pretraining, distillation, fine-tuning, and inverse inference each reduce to independent first-principles or empirical steps, not to their own outputs.

full rationale

The derivation chain is self-contained. OmniFluids' central physics-only pretraining is driven by Eq. (11), a Crank-Nicolson pseudo-spectral residual evaluated on the network's own multi-frame outputs; this is a first-principles consistency condition, not a fit to the quantities later predicted. No parameter is fitted to a subset of the test data and then renamed a prediction. The downstream adaptation stages—operator distillation (Eq. (12)) and few-shot fine-tuning (LData)—are standard compression and supervised alignment steps: the student is trained to match the teacher, and fine-tuning uses at most 10 trajectories, which does not reduce the long-horizon evaluation to the training signal. The inverse experiments (Eq. (13)) optimize viscosity and forcing against two observed frames (t=0 and t=1) and then forecast t=2 to t=10; future frames are not used in the optimization, so the forecast is not a refitted target. The paper's self-citations (e.g., Refs. [31,59] and several same-group architecture papers) occur only in related-work or background contexts and are not load-bearing for the claim that PDE residuals alone yield a transferable operator. The acknowledged limitations—known governing equations, structured grids, no explicit conservation constraints—are scope conditions, not circularity indicators. The main vulnerability, that a short-window residual (t in [0,0.2]) supports long-horizon stability to t=5–10, is a correctness and robustness question, not a circularity one.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on numerical discretization, well-posedness, and learnability assumptions rather than on fitted physical constants. The only fitted quantities appear in inverse experiments and are outputs, not inputs, of the method. No new physical entities are introduced.

free parameters (5)
  • Routing temperature tau = not reported
    Equation (9) uses a temperature-scaled softmax for the mixture of operators; the value is chosen by hand and not stated in the paper.
  • Number of spectral modes per axis = 64 for 3D, maximum for 2D
    Affects spectral resolution and parameter count. The 3D choice of 64 modes versus PINO's 32 modes influences the comparison.
  • Number of experts and layers = P not specified; layers 6-12 depending on task
    Architecture capacity choices selected by validation, not fitted to data.
  • Number of parallel frames K = 100 (KSE), 50 (2D NSE), 40 (3D NSE)
    Multi-frame decoder output length, chosen in the training setup and used in the physics loss.
  • Distillation time factor M = not reported, stated as M >> 1
    Coarse temporal stride for the student model in Eq. (12); its value is not given.
assumptions (4)
  • domain assumption Well-posedness of the KSE, 2D NSE, and 3D NSE on periodic domains for the sampled parameter ranges
    The physics loss and reference data assume unique smooth-enough solutions over t in [0, 10]. Global regularity for 3D NSE is not established in general, though forced viscous flows on finite time horizons are standard numerical practice.
  • domain assumption The Crank-Nicolson pseudo-spectral residual in Eq. (11) is a faithful and sufficient supervisor for learning the solution operator
    This residual is the only training signal in pretraining. If the discretized residual admits spurious low-loss trajectories, the pretrained model may not recover the true operator.
  • domain assumption A neural operator exists and is learnable that maps initial fields and PDE parameters to future frames at the chosen resolutions
    This is the standard operator-learning hypothesis; the paper does not prove approximation or trainability guarantees.
  • standard math Fourier pseudo-spectral differentiation on periodic domains is numerically reliable for the benchmark PDEs
    The pseudo-spectral method and FFT are used for spatial derivatives in the loss and in DNS reference generation, which is standard but still a modeling choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OmniFluids: Physics Pre-trained Modeling of Fluid Dynamics." pith.science (2026). https://pith.science/paper/OAP2EVAN

@misc{pith2026250610862,
  author       = {Pith},
  title        = {Pith review of: OmniFluids: Physics Pre-trained Modeling of Fluid Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OAP2EVAN}},
  note         = {Machine review of arXiv:2506.10862}
}
abstract

Computational fluid dynamics (CFD) drives progress in numerous scientific and engineering fields, yet high-fidelity simulations remain computationally prohibitive. While machine learning approaches offer computing acceleration, they typically specialize in single physical systems or require extensive training data, hindering their applicability in highly nonlinear and 3D flow scenarios. To overcome these limitations, we propose OmniFluids, a pure physics pre-trained model that captures fundamental fluid dynamics laws and adapts efficiently to diverse downstream tasks with minimal data. We develop a training framework combining physics-only pre-training, coarse-grid operator distillation, and few-shot fine-tuning. This enables OmniFluids to retain broad physics knowledge while delivering fast and accurate predictions. Architecturally, OmniFluids integrates a mixture of operators, a multi-frame decoder, and factorized Fourier layers, seamlessly incorporating physics-based supervision while allowing efficient and scalable modeling of diverse tasks. Extensive tests on a broad range of 2D and 3D benchmarks show that OmniFluids outperforms state-of-the-art AI-driven methods in terms of flow field prediction and turbulence statistics. It delivers 10--100$\times$ speedups over traditional solvers while maintaining a comparable accuracy and accurately identifies unknown physical parameters from sparse, noisy data. This work demonstrates the potential of training a unified CFD solver exclusively from physics knowledge, offering a new approach for efficient and generalizable modeling across complex fluid systems.

Figures

Figures reproduced from arXiv: 2506.10862 by the authors.

Figure 1
Figure 1. Schematic framework of OmniFluids. a, Teacher/student network architecture. OmniFluids employs a mixture of operators (MoO) backbone to flexibly handle diverse CFD tasks, alongside a multi-frame decoder that predicts K future steps in parallel. Factorized Fourier layers enable efficient and scalable modeling, especially in 3D. b, Training workflow for OmniFluids. (1) Physics-only pre-training. The teacher network is… view at source ↗
Figure 2
Figure 2. Multi-step error and predicted snapshots from PINO and OmniFluids across diverse CFD [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Predicted snapshots, multi-step errors, and inference speed comparison of baseline methods [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Predicted snapshots, flow statistics, temporal correlation, and inference speed of OmniFluids [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Predicted snapshots, flow statistics, temporal correlation, and inference speed of OmniFluids [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Inverse analysis of fluid systems with OmniFluids. a [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 58 canonical work pages

  1. [1]

    Analysis of a civil aircraft wing transonic shock buffet experiment

    L Masini, S Timme, and AJ Peace. Analysis of a civil aircraft wing transonic shock buffet experiment. J. Fluid Mech. , 884:A1, 2020

  2. [2]

    The quiet revolution of numerical weather prediction

    Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567):47–55, 2015

  3. [3]

    Ac- tive generation and magnetic actuation of microrobotic swarms in bio-fluids.Nat

    Jiangfan Yu, Dongdong Jin, Kai-Fung Chan, Qianqian Wang, Ke Yuan, and Li Zhang. Ac- tive generation and magnetic actuation of microrobotic swarms in bio-fluids.Nat. Commun., 10(1):5631, 2019

  4. [4]

    Fundamentals of computational fluid dynamics

    Harvard Lomax, Thomas H Pulliam, David W Zingg, and TA Kowalewski. Fundamentals of computational fluid dynamics. Appl. Mech. Rev., 55(4):B61–B61, 2002

  5. [5]

    Physics-informed machine learning.Nat

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nat. Rev. Phys. , 3(6):422–440, 2021

  6. [6]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    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. J. Comput. Phys. , 378:686–707, 2019

  7. [7]

    Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations.Science, 367(6481):1026–1030, 2020

    Maziar Raissi, Alireza Yazdani, and George Em Karniadakis. Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations.Science, 367(6481):1026–1030, 2020

  8. [8]

    PINNacle: A comprehensive benchmark of physics- informed neural networks for solving pdes

    Hao Zhongkai, Jiachen Yao, Chang Su, Hang Su, Ziao Wang, Fanzhi Lu, Zeyu Xia, Yichi Zhang, Songming Liu, Lu Lu, and Zhu Jun. PINNacle: A comprehensive benchmark of physics- informed neural networks for solving pdes. In Advances in Neural Information Processing Systems, volume 37, pages 76721–76774, 2024. 20

Show all 64 references
  1. [9]

    Physics-informed learning of governing equations from scarce data

    Zhao Chen, Yang Liu, and Hao Sun. Physics-informed learning of governing equations from scarce data. Nat. Commun., 12(1):6136, 2021

  2. [10]

    Discovery of partial differential equations from highly noisy and sparse data with physics-informed information criterion.Research, 6:0147, 2023

    Hao Xu, Junsheng Zeng, and Dongxiao Zhang. Discovery of partial differential equations from highly noisy and sparse data with physics-informed information criterion.Research, 6:0147, 2023

  3. [11]

    Characterizing possible failure modes in physics-informed neural networks

    Aditi Krishnapriyan, Amir Gholami, Shandian Zhe, Robert Kirby, and Michael W Mahoney. Characterizing possible failure modes in physics-informed neural networks. In Advances in Neural Information Processing Systems , volume 34, pages 26548–26560, 2021

  4. [12]

    Can physics-informed neural networks beat the finite element method?IMA J

    Tamara G Grossmann, Urszula Julia Komorowska, Jonas Latz, and Carola-Bibiane Schönlieb. Can physics-informed neural networks beat the finite element method?IMA J. Appl. Math. , 89(1):hxae011, 2024

  5. [13]

    Neural operators for accelerating scientific simulations and design.Nat

    Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Miguel Liu-Schiaffini, Jean Kossaifi, and Anima Anandkumar. Neural operators for accelerating scientific simulations and design.Nat. Rev. Phys., 6(5):320–328, 2024

  6. [14]

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators

    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. Nat. Mach. Intell. , 3(3):218–229, 2021

  7. [15]

    Learn- ing nonlinear operators in latent spaces for real-time predictions of complex dynamics in phys- ical systems

    KatianaKontolati, SomdattaGoswami, GeorgeEmKarniadakis, andMichaelDShields. Learn- ing nonlinear operators in latent spaces for real-time predictions of complex dynamics in phys- ical systems. Nat. Commun., 15(1):5101, 2024

  8. [16]

    DeepONet based preconditioning strategies for solving parametric linear systems of equations.SIAM J

    Alena Kopaničáková and George Em Karniadakis. DeepONet based preconditioning strategies for solving parametric linear systems of equations.SIAM J. Sci. Comput. , 47(1):C151–C181, 2025

  9. [17]

    Factorized Fourier neural operators

    Alasdair Tran, Alexander Mathews, Lexing Xie, and Cheng Soon Ong. Factorized Fourier neural operators. In International Conference on Learning Representations , 2023

  10. [18]

    Pre- dictionofturbulentchannelflowusingFourierneuraloperator-basedmachine-learningstrategy

    Yunpeng Wang, Zhijie Li, Zelong Yuan, Wenhui Peng, Tianyuan Liu, and Jianchun Wang. Pre- dictionofturbulentchannelflowusingFourierneuraloperator-basedmachine-learningstrategy. Phys. Rev. Fluids , 9(8):084604, 2024

  11. [19]

    Worrall, and Max Welling

    Johannes Brandstetter, Daniel E. Worrall, and Max Welling. Message passing neural PDE solvers. In International Conference on Learning Representations , 2022

  12. [20]

    PhyMPGN: Physics-encoded message passing graph network for spatiotemporal PDE systems

    Bocheng Zeng, Qi Wang, Mengtao Yan, Yang Liu, Ruizhi Chengze, Yi Zhang, Hongsheng Liu, Zidong Wang, and Hao Sun. PhyMPGN: Physics-encoded message passing graph network for spatiotemporal PDE systems. InInternational Conference on Learning Representations , 2025

  13. [21]

    Conservation-informed graph learning for spatiotemporal dynamics prediction

    Yuan Mi, Pu Ren, Hongteng Xu, Hongsheng Liu, Zidong Wang, Yike Guo, Ji-Rong Wen, Hao Sun, and Yang Liu. Conservation-informed graph learning for spatiotemporal dynamics prediction. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 , pa...

  14. [22]

    Scalable Transformer for PDE surrogate mod- eling

    Zijie Li, Dule Shu, and Amir Barati Farimani. Scalable Transformer for PDE surrogate mod- eling. In Advances in Neural Information Processing Systems , volume 36, pages 28010–28039, 2023. 21

  15. [23]

    A Transformer-based neural operator for large-eddy simulation of turbulence.Phys

    Zhijie Li, Tianyuan Liu, Wenhui Peng, Zelong Yuan, and Jianchun Wang. A Transformer-based neural operator for large-eddy simulation of turbulence.Phys. Fluids, 36(6), 2024

  16. [24]

    Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nat

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

  17. [25]

    Generative learning for forecasting the dynamics of high-dimensional complex systems.Nat

    Han Gao, Sebastian Kaltenbach, and Petros Koumoutsakos. Generative learning for forecasting the dynamics of high-dimensional complex systems.Nat. Commun., 15(1):8904, 2024

  18. [26]

    Physics-guided, physics-informed, andphysics-encoded neural networks and operators in scientific computing: Fluid and solid mechanics.J

    Salah A Faroughi, Nikhil M Pawar, Célio Fernandes, Maziar Raissi, Subasish Das, Nima K Kalantari, andSeyedKouroshMahjour. Physics-guided, physics-informed, andphysics-encoded neural networks and operators in scientific computing: Fluid and solid mechanics.J. Comput. Inf. Sci. ...

  19. [27]

    Learning the solution operator of parametric partial differential equations with physics-informed DeepONets

    Sifan Wang, Hanwen Wang, and Paris Perdikaris. Learning the solution operator of parametric partial differential equations with physics-informed DeepONets. Sci. Adv. , 7(40):eabi8605, 2021

  20. [28]

    Separable physics-informed DeepONet: Breaking the curse of dimensionality in physics-informed machine learning.Com- put

    Luis Mandl, Somdatta Goswami, Lena Lambers, and Tim Ricken. Separable physics-informed DeepONet: Breaking the curse of dimensionality in physics-informed machine learning.Com- put. Meth. Appl. Mech. Eng. , 434:117586, 2025

  21. [29]

    Physics-informed neural operator for learning partial differential equations.ACM/JMS J

    Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kam- yar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations.ACM/JMS J. Data Sci. , 1(3):1–27, 2024

  22. [30]

    Variational physics-informed neural operator (VINO) for solving partial differential equations.Comput

    Mohammad Sadegh Eshaghi, Cosmin Anitescu, Manish Thombre, Yizheng Wang, Xiaoying Zhuang, and Timon Rabczuk. Variational physics-informed neural operator (VINO) for solving partial differential equations.Comput. Meth. Appl. Mech. Eng. , 437:117785, 2025

  23. [31]

    Monte Carlo neural PDE solver for learning PDEs via probabilistic representation

    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 Trans. Pattern Anal. Mach. Intell. , 47(6):5059–5075, 2025

  24. [32]

    PDE-Net 2.0: Learning PDEs from data with a numeric-symbolic hybrid deep network.J

    Zichao Long, Yiping Lu, and Bin Dong. PDE-Net 2.0: Learning PDEs from data with a numeric-symbolic hybrid deep network.J. Comput. Phys. , 399:108925, 2019

  25. [33]

    Encoding physics to learn reaction–diffusion processes.Nat

    Chengping Rao, Pu Ren, Qi Wang, Oral Buyukozturk, Hao Sun, and Yang Liu. Encoding physics to learn reaction–diffusion processes.Nat. Mach. Intell. , 5(7):765–779, 2023

  26. [34]

    P2C2Net: PDE-preserved coarse correction network for efficient prediction of spatiotemporal dynamics

    Qi Wang, Pu Ren, Hao Zhou, Xin-Yang Liu, Zhiwen Deng, Yi Zhang, Ruizhi Chengze, Hong- sheng Liu, Zidong Wang, Jian-Xun Wang, Ji-Rong Wen, Hao Sun, and Yang Liu. P2C2Net: PDE-preserved coarse correction network for efficient prediction of spatiotemporal dynamics. In Advances in...

  27. [35]

    Multi-resolution partial differ- ential equations preserved learning framework for spatiotemporal dynamics.Commun

    Xin-Yang Liu, Min Zhu, Lu Lu, Hao Sun, and Jian-Xun Wang. Multi-resolution partial differ- ential equations preserved learning framework for spatiotemporal dynamics.Commun. Phys., 7(1):31, 2024

  28. [36]

    MultiPDENet: PDE-embedded learning with multi-time-stepping for accelerated flow simulation

    Qi Wang, Yuan Mi, Wang Haoyun, Yi Zhang, Ruizhi Chengze, Hongsheng Liu, Ji-Rong Wen, and Hao Sun. MultiPDENet: PDE-embedded learning with multi-time-stepping for accelerated flow simulation. InInternational Conference on Machine Learning , 2025. 22

  29. [37]

    Machine learning–accelerated computational fluid dynamics

    Dmitrii Kochkov, Jamie A Smith, Ayya Alieva, Qing Wang, Michael P Brenner, and Stephan Hoyer. Machine learning–accelerated computational fluid dynamics. Proc. Natl. Acad. Sci. , 118(21):e2101784118, 2021

  30. [38]

    A neural PDE solver with temporal stencil modeling

    Zhiqing Sun, Yiming Yang, and Shinjae Yoo. A neural PDE solver with temporal stencil modeling. In International Conference on Machine Learning , volume 202, pages 33135–33155, 2023

  31. [39]

    Graph neural PDE solvers with conservation and similarity-equivariance

    Masanobu Horie and Naoto Mitsume. Graph neural PDE solvers with conservation and similarity-equivariance. InProceedings of the 41st International Conference on Machine Learn- ing, pages 18785–18814, 2024

  32. [40]

    Learnable-differentiable finite volume solver for acceler- ated simulation of flows

    Mengtao Yan, Qi Wang, Haining Wang, Ruizhi Chengze, Yi Zhang, Hongsheng Liu, Zidong Wang, Fan Yu, Qi Qi, and Hao Sun. Learnable-differentiable finite volume solver for acceler- ated simulation of flows. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and...

  33. [41]

    Mesh-informed neural networks for operator learning in finite element spaces.J

    Nicola Rares Franco, Andrea Manzoni, and Paolo Zunino. Mesh-informed neural networks for operator learning in finite element spaces.J. Sci. Comput. , 97(2):35, 2023

  34. [42]

    Weak baselines and reporting biases lead to overopti- mism in machine learning for fluid-related partial differential equations.Nat

    Nick McGreivy and Ammar Hakim. Weak baselines and reporting biases lead to overopti- mism in machine learning for fluid-related partial differential equations.Nat. Mach. Intell. , 6(10):1256–1269, 2024

  35. [43]

    Foundation models for generalist medical artificial intelligence

    Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259–265, 2023

  36. [44]

    Bruinsma, Ana Lucic, Megan Stanley, Anna Allen, Johannes Brand- stetter, Patrick Garvan, Maik Riechert, Jonathan A

    Cristian Bodnar, Wessel P. Bruinsma, Ana Lucic, Megan Stanley, Anna Allen, Johannes Brand- stetter, Patrick Garvan, Maik Riechert, Jonathan A. Weyn, Haiyu Dong, Jayesh K. Gupta, Kit Thambiratnam, Alexander T. Archibald, Chun-Chieh Wu, Elizabeth Heider, Max Welling, Richard E. ...

  37. [45]

    DPOT: Auto-regressive denoising operator Transformer for large-scale PDE pre-training

    Zhongkai Hao, Chang Su, Songming Liu, Julius Berner, Chengyang Ying, Hang Su, Anima Anandkumar, Jian Song, and Jun Zhu. DPOT: Auto-regressive denoising operator Transformer for large-scale PDE pre-training. In International Conference on Machine Learning , pages 17616–17635, 2024

  38. [46]

    Multiple physics pretraining for spatiotemporal surrogate models

    Michael McCabe, Bruno Régaldo-Saint Blancard, Liam Parker, Ruben Ohana, Miles Cranmer, AlbertoBietti, MichaelEickenberg, SiavashGolkar, GeraudKrawezik, FrancoisLanusse, Mariel Pettee, Tiberiu Tesileanu, Kyunghyun Cho, and Shirley Ho. Multiple physics pretraining for spatiotemp...

  39. [47]

    Building flexible machine learning models for scientific computing at scale

    Tianyu Chen, Haoyi Zhou, Ying Li, Hao Wang, Chonghan Gao, Rongye Shi, Shanghang Zhang, and Jianxin Li. Building flexible machine learning models for scientific computing at scale. In International Conference on Machine Learning , 2025

  40. [48]

    Deep transfer operator learning for partial differential equations under conditional shift.Nat

    Somdatta Goswami, Katiana Kontolati, Michael D Shields, and George Em Karniadakis. Deep transfer operator learning for partial differential equations under conditional shift.Nat. Mach. Intell., 4(12):1155–1164, 2022. 23

  41. [49]

    In-context operator learning with data prompts for differential equation problems.Proc

    Liu Yang, Siting Liu, Tingwei Meng, and Stanley J Osher. In-context operator learning with data prompts for differential equation problems.Proc. Natl. Acad. Sci. , 120(39):e2310142120, 2023

  42. [50]

    Learning spatiotemporal dynamics with a pretrained generative model

    Zeyu Li, Wang Han, Yue Zhang, Qingfei Fu, Jingxuan Li, Lizi Qin, Ruoyu Dong, Hao Sun, Yue Deng, and Lijun Yang. Learning spatiotemporal dynamics with a pretrained generative model. Nat. Mach. Intell. , 6(12):1566–1579, 2024

  43. [51]

    On flame propagation under conditions of stoichiometry.SIAM J

    Gregory I Sivashinsky. On flame propagation under conditions of stoichiometry.SIAM J. Appl. Math., 39(1):67–82, 1980

  44. [52]

    Nonlinear analysis of hydrodynamic instability in laminar flames-II

    Daniel M Michelson and Gregory I Sivashinsky. Nonlinear analysis of hydrodynamic instability in laminar flames-II. Numerical experiments.Acta Astronaut., 4(11-12):1207–1221, 1977

  45. [53]

    A vorticity-velocity method for the numerical solution of 3D incompressible flows

    Giulio Guj and Fulvio Stella. A vorticity-velocity method for the numerical solution of 3D incompressible flows. J. Comput. Phys. , 106(2):286–298, 1993

  46. [54]

    Navier-Stokes equations on thin 3D domains

    Geneviève Raugel and George R Sell. Navier-Stokes equations on thin 3D domains. I. Global attractors and global regularity of solutions.J. Am. Math. Soc. , 6(3):503–568, 1993

  47. [55]

    From zero to tur- bulence: Generative modeling for 3D flow simulation

    Marten Lienen, David Lüdke, Jan Hansen-Palmus, and Stephan Günnemann. From zero to tur- bulence: Generative modeling for 3D flow simulation. InInternational Conference on Learning Representations, 2024

  48. [56]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede liu, Kaushik Bhat- tacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. InInternational Conference on Learning Representations , 2021

  49. [57]

    Convolutional neural operators for robust and accurate learning of PDEs

    Bogdan Raonic, Roberto Molinaro, Tim De Ryck, Tobias Rohner, Francesca Bartolucci, Rima Alaifari, Siddhartha Mishra, and Emmanuel de Bézenac. Convolutional neural operators for robust and accurate learning of PDEs. InAdvances in Neural Information Processing Systems , volume 3...

  50. [58]

    Bayesian inverse problems for functions and applications to fluid mechanics.Inverse Probl., 25(11):115008, 2009

    Simon L Cotter, Massoumeh Dashti, James Cooper Robinson, and Andrew M Stuart. Bayesian inverse problems for functions and applications to fluid mechanics.Inverse Probl., 25(11):115008, 2009

  51. [59]

    DRVN (deep random vortex network): A new physics-informed machine learning method for simulating and inferring incompressible fluid flows.Phy

    Rui Zhang, Peiyan Hu, Qi Meng, Yue Wang, Rongchan Zhu, Bingguang Chen, Zhi-Ming Ma, and Tie-Yan Liu. DRVN (deep random vortex network): A new physics-informed machine learning method for simulating and inferring incompressible fluid flows.Phy. Fluids, 34(10), 2022

  52. [60]

    Layer normalization.arXiv preprint arXiv:1607.06450, 2016

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization.arXiv preprint arXiv:1607.06450, 2016

  53. [61]

    Bridging traditional and machine learning-based algorithms for solving PDEs: The random feature method

    Jingrun Chen, Xurong Chi, Weinan E, and Zhouwang Yang. Bridging traditional and machine learning-based algorithms for solving PDEs: The random feature method. J. Mach. Learn. , 1(3):268–298, 2022

  54. [62]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 , 2015. 24

  55. [63]

    PyTorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu ...

  56. [64]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInterna- tional Conference on Learning Representations , 2015. Acknowledgement: The work is supported by the National Natural Science Foundation of China (No. 62276269, No. 92270118), the Beijing Nat...

Pith tools

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