pith. sign in

arxiv: 2606.01470 · v1 · pith:V42YIPE2new · submitted 2026-05-31 · ⚛️ physics.flu-dyn · cs.AI· cs.LG

Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence

Payel Mukhopadhyay , Stefan S. Nixon , Romain Watteaux , Michael McCabe , Alberto Bietti , Kyunghyun Cho , Cristiana Diaconu , Irina Espejo Morales
show 15 more authors
David Fouhey Siavash Golkar Tom Hehir Shirley Ho Jake Kovalic Geraud Krawezik Francois Lanusse Tanya Marwah Rudy Morel Mariel Pettee Helen Qu Jeff Shen Hadi Sotoudeh Stuart B. Dalziel Miles Cranmer
This is my paper

Pith reviewed 2026-06-28 16:05 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn cs.AIcs.LG
keywords Rayleigh-Taylor instabilityfoundation modeldirect numerical simulationlaboratory experimentmixing growth ratezero-shot generalizationinitial conditionsstable stratification
0
0 comments X

The pith

A foundation model finetuned on few DNS runs of Rayleigh-Taylor instability predicts the higher experimental mixing growth rates zero-shot on laboratory data.

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

The paper examines whether physics foundation models trained solely on simulations can generalize to real laboratory experiments, using the Rayleigh-Taylor instability as a test case. After finetuning on three or fewer direct numerical simulation realizations, the Walrus model recovers essential RTI dynamics in extended predictions. Applied without further training to sliding-barrier laboratory measurements, it shifts from the lower DNS-like growth rate to the higher experimental band of alpha around 0.06-0.07. This supplies independent evidence that differences in initial conditions largely explain the long-standing simulation-experiment discrepancy. The same model also transfers zero-shot to stable stratification, a regime absent from its training data, by correctly reducing mixing-layer growth.

Core claim

After finetuning on three or fewer DNS realizations, the Walrus foundation model for continuum dynamics recovers key Rayleigh-Taylor instability physics over long rollouts. When applied zero-shot to sliding-barrier laboratory data, the model leaves the DNS-like regime and enters the observed experimental growth band for alpha (~0.06-0.07), having seen no experimental samples. These results supply independent, data-driven evidence that initial conditions play a crucial role in the longstanding sim-experiment gap in alpha. The model additionally generalizes zero-shot to stable stratification, a buoyancy regime absent from training, and correctly slows mixing-layer growth.

What carries the argument

The Walrus foundation model for continuum dynamics, finetuned on a small number of DNS realizations and then applied zero-shot to experimental time series to predict mixing evolution.

If this is right

  • The model can predict laboratory mixing behavior without ever seeing experimental training samples.
  • Initial conditions are a primary driver of the discrepancy between simulation and experiment values of the mixing growth rate alpha.
  • The same model can handle buoyancy regimes such as stable stratification that were never present in its training data.
  • Foundation models trained exclusively on simulations can be deployed on sparse and noisy laboratory settings for fluid instabilities.

Where Pith is reading between the lines

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

  • Similar few-shot finetuning on DNS could be used to probe initial-condition sensitivity in other fluid instabilities where simulation-experiment gaps persist.
  • If the zero-shot transfer generalizes, it offers a route to test whether other experimental artifacts beyond initial conditions contribute to observed discrepancies.
  • The efficiency with three or fewer DNS runs suggests that adaptation costs remain low even when extending the approach to additional laboratory configurations.

Load-bearing premise

The sliding-barrier laboratory data can be treated as directly comparable to idealized DNS without unaccounted differences in boundary conditions, measurement noise, or other experimental artifacts.

What would settle it

Measuring that the model's predicted alpha on the sliding-barrier laboratory data remains near the DNS value of ~0.02 instead of entering the 0.06-0.07 experimental band would falsify the transfer claim.

Figures

Figures reproduced from arXiv: 2606.01470 by Alberto Bietti, Cristiana Diaconu, David Fouhey, Francois Lanusse, Geraud Krawezik, Hadi Sotoudeh, Helen Qu, Irina Espejo Morales, Jake Kovalic, Jeff Shen, Kyunghyun Cho, Mariel Pettee, Michael McCabe, Miles Cranmer, Payel Mukhopadhyay, Romain Watteaux, Rudy Morel, Shirley Ho, Siavash Golkar, Stefan S. Nixon, Stuart B. Dalziel, Tanya Marwah, Tom Hehir.

Figure 1
Figure 1. Figure 1: Overview. (A) Walrus is pretrained on a broad set of continuum-dynamics simulations and then finetuned on a small number of RTI direct numerical simulations (DNS) in the Boussinesq regime (approximately incompressible, with small density differences driving buoyancy). From a total of 5 DNS realizations, we use 3 for finetuning, 1 for val￾idation (used to monitor training progress and select the best model)… view at source ↗
Figure 2
Figure 2. Figure 2: Results from a model trained natively at 1283 are provided for completeness in Appendix 8.4. DNS t = 10 t = 40 t = 60 t = 90 Walrus ( W 3 D DNS ) [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Energy spectra. Spectra of the density and velocity components at t = 70, in the self-similar regime, comparing DNS and W3D DNS (evaluated at 1283 ; original DNS at 2563 ). Walrus matches the spectral shape across the inertial range, supporting faithful reproduction of the kinetic-energy distribution across scales. The z-direction is per￾pendicular to the interface; x and y are the two horizontal direction… view at source ↗
Figure 4
Figure 4. Figure 4: Mixing growth rate. Left: concentration-derived mixing width h(t) computed from Eq. 1. Right: corre￾sponding growth coefficient α(t) computed from h(t) and its time derivative (Eq. 5) for W3D DNS and the held-out DNS test case S5. W3D DNS captures both the integrated mixing-layer growth and the late-time self-similar growth regime of the DNS reference. For completeness, we note that Walrus predictions exhi… view at source ↗
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sample efficiency of 3D RTI finetuning. First three panels: z-averaged kinetic-energy spectra in the x-y planes of the held-out test case S5, computed at a matched rollout time of t = 70 in the self-similar regime for three independent finetuning runs starting from the same pretrained checkpoint Wpre and using 1, 2, or 3 training realizations from {S1, S2, S3}. In each case, the spectrum of the resulting f… view at source ↗
Figure 7
Figure 7. Figure 7: Initial conditions: DNS vs. sliding-barrier experiment. Concentration fields for a representative DNS realization (left) and a sliding-barrier experimental sample (right). The DNS interface carries short-wavelength per￾turbations characteristic of idealized numerical initialization, whereas the experimental interface carries large-scale structure set by barrier motion, structural vibration, and molecular d… view at source ↗
Figure 8
Figure 8. Figure 8: Zero-shot transfer from idealized DNS to experimental RTI data. (a) Validation on held-out 2D DNS slices. The black curve shows the mean growth coefficient α(t) computed over the held-out 2D DNS slices, with the shaded band denoting the spread across slices. The blue curve, labeled W2D DNS(DNS), shows the corresponding rollout of the DNS-specialized model W2D DNS initialized from DNS input frames. The hori… view at source ↗
Figure 9
Figure 9. Figure 9: α(t) for held-out test set experimental data: zero-shot transfer and experimentally adapted predic￾tions. Left: rollout of W2D DNS on the three held-out experimental samples, without any experimental finetuning. Right: rollout of W2D DNS+Exp, obtained after a second finetuning stage on the two remaining experimental samples and eval￾uated on the same three held-out cases. In both panels, black denotes the … view at source ↗
Figure 10
Figure 10. Figure 10: Zero-shot transfer of Walrus finetuned on unstratified DNS to stratified Rayleigh-Taylor flow. (a) Ini￾tial horizontally averaged concentration profiles for the unstratified and stratified cases. (b) Evolution of the mean concentration profile in the stratified reference DNS. (c) Corresponding zero-shot rollout from W3D DNS, finetuned only on unstratified RTI DNS. Stable stratification confines the partia… view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative agreement for Wnative DNS,128 at native 1283 resolution and t = 60. Comparison between the held-out native 1283 DNS and the corresponding prediction of Wnative DNS,128, obtained by finetuning Wpre directly on native 1283 RTI data. This setting is separate from the downsampled 2563 → 1283 setup used in the main text. Direct native-1283 finetuning yields good local reconstruction while again pre… view at source ↗
Figure 12
Figure 12. Figure 12: Rayleigh-Taylor flow structure for Wnative DNS,128 at native 1283 resolution. Vertical cross-sections of the vertical velocity component vz at multiple rollout times for the held-out native 1283 direct numerical simulation (DNS, top row) and the corresponding prediction of Wnative DNS,128 (bottom row). The slices correspond to an x-z plane taken at the midplane of the domain in the y direction. In this na… view at source ↗
Figure 13
Figure 13. Figure 13: Breakdown of the Fourier Neural Operator (FNO) on 3D RTI rollout. Comparison between the held￾out DNS trajectory (top row) and the corresponding autoregressive FNO rollout (bottom row) at four representative times. While the DNS develops the expected RTI mixed layer and bubble morphology, the FNO prediction quickly departs from physically plausible evolution and forms large slab-like structures that do no… view at source ↗
Figure 14
Figure 14. Figure 14: Breakdown of the Tensorized Fourier Neural Operator (TFNO) on 3D RTI rollout. Comparison between the held-out DNS trajectory (top row) and the corresponding autoregressive TFNO rollout (bottom row) at four representative times. After an initially plausible state, the TFNO rollout collapses into highly unphysical artifacts and fails to sustain the development of a realistic RTI mixed layer. This illustrate… view at source ↗
Figure 15
Figure 15. Figure 15: Breakdown of ConvNeXt-UNet on 3D RTI rollout. Comparison between the held-out DNS trajectory (top row) and the corresponding autoregressive ConvNeXt-UNet rollout (bottom row) at four representative times. Although the early state remains roughly plausible, the rollout soon develops spurious artifacts that are incompatible with the expected RTI evolution. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Evolution of the α value for DNS and W3D DNS, shown along with a shade of blue band which shows the randomness associated to patch jittering. The blue band is created by taking 10 predicted outputs of WDNS on the same initial conditions of the test set. As noted in the main text, Walrus architecture has some inherent randomness in its output predictions, which is a consequence of the patch jittering techn… view at source ↗
Figure 17
Figure 17. Figure 17: Experimental setup. The polycarbonate barrier (light grey) is removed at velocity [PITH_FULL_IMAGE:figures/full_fig_p031_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Mean concentration evolution for W2D DNS initialized from DNS and experimental frames. Space-time maps of the horizontally averaged concentration profile c¯(t, z) (defined in Sec. 2) for W2D DNS initialized from 2D DNS frames (left) and the same model evaluated zero-shot on experimental initial conditions (right). Supplying experimen￾tal frames shifts the rollout away from the DNS-like mixing regime towar… view at source ↗
Figure 19
Figure 19. Figure 19: Robustness of zero-shot experimental transfer to the number of input frames. Zero-shot predictions on the held-out experimental samples for three separate DNS-specialized 2D models, all initialized from Wpre and then finetuned on simulated 2D RTI slices using context lengths L = 1, L = 2, and L = 3, respectively. In each panel, black denotes the experimental data and gold the corresponding Walrus predicti… view at source ↗
Figure 20
Figure 20. Figure 20: shows zero-shot rollouts of Wpre when initialized from 2D DNS input (top row) and from experimental input (bottom row). In neither case does the pretrained model produce a physically meaningful RTI evolution. On 2D DNS input, it does not recover the expected RTI morphology. On experimental input, it likewise fails to generate a coherent late-time RTI rollout. This matters because it shows that the experim… view at source ↗
Figure 21
Figure 21. Figure 21: Rollouts of the 2D DNS-specialized model W2D DNS on 2D DNS and experimental input. Top row: rollout of W2D DNS initialized from 2D DNS input. Bottom row: zero-shot rollout of W2D DNS initialized from experimental input. Columns show representative times. After specialization to RTI on simulated 2D DNS slices, the model reproduces DNS-like growth in the top row, but develops markedly stronger mixing and in… view at source ↗
Figure 22
Figure 22. Figure 22: Early experimental evolution and the effect of second-stage finetuning. Top: held-out experimental concentration fields during the barrier-release transient and early instability growth. Bottom: corresponding rollouts of the experimentally adapted model W2D DNS+Exp. Relative to the zero-shot results discussed in the main text, the second finetuning stage improves agreement with the release-driven early-ti… view at source ↗
Figure 23
Figure 23. Figure 23: Dominant early-time DMD mode in experiment, W2D DNS+Exp, and the idealized DNS baseline. The leading dynamic mode decomposition (DMD) mode of the early-time vorticity field highlights the large-scale anisotropic structure associated with experimental release. This structure is pronounced in the experimental data (top) but is not captured by the idealized DNS baseline (bottom), whose dominant mode remains … view at source ↗
read the original abstract

Whether physics foundation models can be usefully deployed on laboratory experiments remains an open question for scientific machine learning (ML). We test this question on the Rayleigh-Taylor instability (RTI), a ubiquitous and demanding fluid instability seen from tabletop flows to supernova explosions, in which small perturbations at a density interface grow into chaotic, multiscale mixing as a lighter fluid accelerates into a heavier one. Standard ML models struggle with RTI, and despite over a century of theoretical, numerical, and experimental work, it carries an unresolved discrepancy between simulation and experiment: the late-time mixing growth rate, $\alpha$, measured in most laboratory experiments ($\sim$ 0.06-0.07), is roughly three times the value from idealized direct numerical simulations (DNS, $\sim$ 0.02). The gap's origin remains debated. These properties make RTI a stringent test for a question that matters well beyond RTI: can foundation models trained only on simulations generalise to sparse, messy, and noisy laboratory settings? We finetune Walrus, a foundation model for continuum dynamics, on three or fewer DNS realizations and recover key RTI physics over long rollouts. Applied zero-shot to sliding-barrier laboratory data, the finetuned model leaves the DNS-like regime and enters the observed growth band, having never seen a single experimental sample. These results provide independent, data-driven evidence that initial conditions play a crucial role in the longstanding sim-experiment gap in $\alpha$. The model also generalises zero-shot to stable stratification, a buoyancy regime absent from training, correctly slowing mixing-layer growth. Together, our results show that foundation models can generalise well beyond their training data, predicting laboratory behavior and unseen physical regimes, opening new ways to probe longstanding simulation-experiment gaps.

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

3 major / 2 minor

Summary. The paper claims that a foundation model (Walrus) finetuned on three or fewer DNS realizations of Rayleigh-Taylor instability recovers key physics over long rollouts; when applied zero-shot to sliding-barrier laboratory data, it produces a late-time mixing growth rate α in the experimental band (~0.06-0.07) rather than the DNS value (~0.02), supplying independent evidence that initial conditions explain the sim-experiment gap. The model also generalizes zero-shot to stable stratification, a regime absent from training.

Significance. If the central transfer result holds under rigorous controls, the work would demonstrate that physics foundation models trained solely on simulation can usefully predict laboratory behavior in a demanding, multiscale fluid instability, offering a new route to probe longstanding simulation-experiment discrepancies. The zero-shot generalization to an unseen buoyancy regime and the data-driven support for the role of initial conditions would be notable strengths for the field.

major comments (3)
  1. [§4] §4 (Laboratory data preparation): The claim that the observed α shift arises purely from differences in initial conditions requires that the laboratory perturbation spectrum, interface profile, and effective Atwood number are extracted and supplied without unaccounted mismatch; the manuscript provides no quantitative side-by-side comparison of these quantities between the sliding-barrier data and the DNS training set, which is load-bearing for the interpretation.
  2. [§5.1] §5.1 (Alpha extraction protocol): The reported entry into the experimental α band (0.06-0.07) is central to the headline result, yet the text does not demonstrate that the identical growth-rate measurement protocol (including any spatial averaging or time-window choices) is applied to both the model rollout and the laboratory reference data; any discrepancy here would undermine the direct comparison.
  3. [§5.3] §5.3 (Boundary-condition controls): The laboratory setup includes walls and end effects absent from the idealized DNS; without explicit tests showing that these do not alter the effective growth rate when the model is driven by the lab initial conditions, the attribution of the α change solely to initial conditions remains at risk.
minor comments (2)
  1. [Abstract] The abstract states 'three or fewer' DNS realizations but the main text should state the exact count used for the primary finetuning run to aid reproducibility.
  2. [Figures] Figure captions for the growth curves should explicitly label the DNS baseline, experimental reference band, and model prediction with consistent line styles and include uncertainty estimates where available.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful and constructive report. The points raised highlight important aspects of rigor for the laboratory transfer claim. We respond to each major comment below and indicate the revisions planned.

read point-by-point responses

  1. Referee: [§4] §4 (Laboratory data preparation): The claim that the observed α shift arises purely from differences in initial conditions requires that the laboratory perturbation spectrum, interface profile, and effective Atwood number are extracted and supplied without unaccounted mismatch; the manuscript provides no quantitative side-by-side comparison of these quantities between the sliding-barrier data and the DNS training set, which is load-bearing for the interpretation.

    Authors: We agree that explicit quantitative comparisons are required to support the interpretation. In the revised manuscript we will add a new panel or table in §4 that directly compares (i) the initial perturbation power spectra obtained via consistent Fourier analysis, (ii) the initial interface thickness and shape profiles, and (iii) the effective Atwood numbers between the three DNS training realizations and the sliding-barrier laboratory data. Extraction methods will be documented to confirm no unaccounted mismatch. revision: yes

  2. Referee: [§5.1] §5.1 (Alpha extraction protocol): The reported entry into the experimental α band (0.06-0.07) is central to the headline result, yet the text does not demonstrate that the identical growth-rate measurement protocol (including any spatial averaging or time-window choices) is applied to both the model rollout and the laboratory reference data; any discrepancy here would undermine the direct comparison.

    Authors: We acknowledge that the protocol must be shown to be identical. The growth rate α is obtained from the same definition of mixing-layer width h(t) and the same linear regression in the self-similar regime, using identical time windows and spatial averaging. In the revision we will state the exact protocol (time interval, averaging domain, fitting procedure) explicitly in §5.1 and confirm its uniform application to both model rollouts and laboratory data, accompanied by supplementary h(t) curves for verification. revision: yes

  3. Referee: [§5.3] §5.3 (Boundary-condition controls): The laboratory setup includes walls and end effects absent from the idealized DNS; without explicit tests showing that these do not alter the effective growth rate when the model is driven by the lab initial conditions, the attribution of the α change solely to initial conditions remains at risk.

    Authors: This concern is valid. The laboratory measurements are taken from the central region, and the model is driven by lab initial conditions on a periodic domain approximating that region. We have not performed explicit wall-boundary tests. In the revision we will add a limitations paragraph in §5.3 discussing possible wall influence and, where computationally feasible, include a sensitivity test imposing approximate no-slip or damping boundaries to quantify any effect on α. We maintain that the primary change is driven by initial conditions, but will clarify the boundary assumptions. revision: partial

Circularity Check

0 steps flagged

No significant circularity; central result is empirical generalization to external lab data

full rationale

The paper's load-bearing claim is that a model finetuned only on DNS, when fed sliding-barrier laboratory inputs zero-shot, produces an alpha value that falls inside the independently measured experimental band (0.06-0.07). This alpha is extracted from the model's rollout on new data and compared to an external experimental reference; it is not fitted to the target alpha, not defined in terms of itself, and not justified by a self-citation chain. The derivation chain therefore remains self-contained against external benchmarks, with no reduction of the reported prediction to the model's own inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the premise that the pre-trained Walrus model encodes transferable continuum physics and that the chosen laboratory dataset differs from DNS primarily through initial conditions rather than other unmodeled factors.

free parameters (1)
  • finetuning dataset size
    The decision to use three or fewer DNS realizations is a modeling choice that directly affects the reported generalization performance.
axioms (1)
  • domain assumption Walrus has learned sufficiently general representations of fluid dynamics from its pre-training corpus to support effective finetuning and zero-shot transfer.
    This assumption underpins the entire finetuning and application procedure described in the abstract.

pith-pipeline@v0.9.1-grok · 5959 in / 1309 out tokens · 38458 ms · 2026-06-28T16:05:10.305255+00:00 · methodology

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

96 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stef...

    2021

  2. [2]

    Pdebench: An extensive benchmark for scientific machine learning.Advances in Neural Information Processing Systems, 35:1596–1611, 2022

    Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKinlay, Francesco Alesiani, Dirk Pfl ¨uger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning.Advances in Neural Information Processing Systems, 35:1596–1611, 2022

    2022

  3. [3]

    Dalziel, Drummond Buschman Fielding, Daniel Fortunato, Jared A

    Ruben Ohana, Michael McCabe, Lucas Thibaut Meyer, Rudy Morel, Fruzsina Julia Agocs, Miguel Beneitez, Marsha Berger, Blakesley Burkhart, Stuart B. Dalziel, Drummond Buschman Fielding, Daniel Fortunato, Jared A. Goldberg, Keiya Hirashima, Yan-Fei Jiang, Rich Kerswell, Suryanarayana Maddu, Jonah M. Miller, Payel Mukhopadhyay, Stefan S. Nixon, Jeff Shen, Roma...

    2024

  4. [4]

    Realpdebench: A benchmark for complex physical systems with real-world data

    Peiyan Hu, Haodong Feng, Hongyuan Liu, Tongtong Yan, Wenhao Deng, Tianrun Gao, Rong Zheng, Haoren Zheng, Chenglei Yu, Chuanrui Wang, Kaiwen Li, Zhi-Ming Ma, Dezhi Zhou, Xingcai Lu, Dixia Fan, and Tailin Wu. Realpdebench: A benchmark for complex physical systems with real-world data. InThe Fourteenth Inter- national Conference on Learning Representations, ...

    2026

  5. [5]

    Towards multi-spatiotemporal-scale generalized PDE modeling

    Jayesh K Gupta and Johannes Brandstetter. Towards multi-spatiotemporal-scale generalized PDE modeling. Transactions on Machine Learning Research, 2023

    2023

  6. [6]

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

    Zhongkai Hao, Jiachen Yao, Chang Su, Hang Su, Ziao Wang, Fanzhi Lu, Zeyu Xia, Yichi Zhang, Songming Liu, Lu Lu, et al. Pinnacle: A comprehensive benchmark of physics-informed neural networks for solving pdes. arXiv preprint arXiv:2306.08827, 2023

    work page arXiv 2023

  7. [7]

    AirfRANS: High fidelity com- putational fluid dynamics dataset for approximating reynolds-averaged navier–stokes solutions

    Florent Bonnet, Jocelyn Ahmed Mazari, Paola Cinnella, and patrick gallinari. AirfRANS: High fidelity com- putational fluid dynamics dataset for approximating reynolds-averaged navier–stokes solutions. InThirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022

    2022

  8. [8]

    Artur Toshev, Gianluca Galletti, Fabian Fritz, Stefan Adami, and Nikolaus A. Adams. Lagrangebench: A la- grangian fluid mechanics benchmarking suite. InThirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023

    2023

  9. [9]

    Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, Andr ´as Hor ´anyi, Joaqu ´ın Mu˜noz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, Adrian Simmons, Cornel Soci, Saleh Abdalla, Xavier Abellan, Gianpaolo Balsamo, Peter Bechtold, Gionata Biavati, Jean Bidlot, Massimo Bonavita, Giovanna Chiara, Per Dahlgren, Dick Dee, Michail...

    1999

  10. [10]

    David Neelin, David Randall, Sara Shamekh, Mark A Taylor, Nathan Urban, Janni Yuval, Guang Zhang, and Michael Pritchard

    Sungduk Yu, Walter Hannah, Liran Peng, Jerry Lin, Mohamed Aziz Bhouri, Ritwik Gupta, Bj ¨orn L¨utjens, Jus- tus Christopher Will, Gunnar Behrens, Julius Busecke, Nora Loose, Charles I Stern, Tom Beucler, Bryce Harrop, Benjamin R Hillman, Andrea Jenney, Savannah Ferretti, Nana Liu, Anima Anandkumar, Noah D Brenowitz, 18 Emergent transfer of a physics found...

    2023

  11. [11]

    EA- GLE: Large-scale learning of turbulent fluid dynamics with mesh transformers

    Steeven JANNY , Aur´elien B ´en´eteau, Madiha Nadri, Julie Digne, Nicolas THOME, and Christian Wolf. EA- GLE: Large-scale learning of turbulent fluid dynamics with mesh transformers. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  12. [12]

    Bubbleformer: Forecasting boiling with transformers

    Sheikh Md Shakeel Hassan, Xianwei Zou, Akash Dhruv, and Aparna Chandramowlishwaran. Bubbleformer: Forecasting boiling with transformers. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2026

    2026

  13. [13]

    LIPS - learning industrial physical sim- ulation benchmark suite

    Milad Leyli-abadi, Antoine Marot, J ´erˆome Picault, David Danan, Mouadh Yagoubi, Benjamin Donnot, Seif- Eddine Attoui, Pavel Dimitrov, Asma Farjallah, and Clement Etienam. LIPS - learning industrial physical sim- ulation benchmark suite. InThirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022

    2022

  14. [14]

    Turbulence in focus: Benchmarking scaling behavior of 3d volumetric super-resolution with BLASTNet 2.0 data

    Wai Tong Chung, Bassem Akoush, Pushan Sharma, Alex Tamkin, Ki Sung Jung, Jacqueline Chen, Jack Guo, Davy Brouzet, Mohsen Talei, Bruno Savard, Alexei Y Poludnenko, and Matthias Ihme. Turbulence in focus: Benchmarking scaling behavior of 3d volumetric super-resolution with BLASTNet 2.0 data. InThirty-seventh Conference on Neural Information Processing Syste...

    2023

  15. [15]

    A public turbulence database cluster and applications to study lagrangian evolution of velocity increments in turbulence.Journal of Turbulence, 9:N31, January 2008

    Yi Li, Eric Perlman, Minping Wan, Yunke Yang, Charles Meneveau, Randal Burns, Shiyi Chen, Alexander Szalay, and Gregory Eyink. A public turbulence database cluster and applications to study lagrangian evolution of velocity increments in turbulence.Journal of Turbulence, 9:N31, January 2008

    2008

  16. [16]

    Flowbench: A large scale benchmark for flow simulation over complex geometries, 2024

    Ronak Tali, Ali Rabeh, Cheng-Hau Yang, Mehdi Shadkhah, Samundra Karki, Abhisek Upadhyaya, Suriya Dhak- shinamoorthy, Marjan Saadati, Soumik Sarkar, Adarsh Krishnamurthy, Chinmay Hegde, Aditya Balu, and Baskar Ganapathysubramanian. Flowbench: A large scale benchmark for flow simulation over complex geometries, 2024

    2024

  17. [17]

    Poseidon: Efficient foundation models for PDEs

    Maximilian Herde, Bogdan Raonic, Tobias Rohner, Roger K ¨appeli, Roberto Molinaro, Emmanuel de Bezenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for PDEs. InThe Thirty-eighth Annual Confer- ence on Neural Information Processing Systems, 2024

    2024

  18. [18]

    Multiple physics pretraining for spatiotemporal surrogate models

    Michael McCabe, Bruno R ´egaldo-Saint Blancard, Liam Holden Parker, Ruben Ohana, Miles Cranmer, Al- berto Bietti, Michael Eickenberg, Siavash Golkar, Geraud Krawezik, Francois Lanusse, Mariel Pettee, Tiberiu Tesileanu, Kyunghyun Cho, and Shirley Ho. Multiple physics pretraining for spatiotemporal surrogate models. InThe Thirty-eighth Annual Conference on ...

    2024

  19. [19]

    Mahoney, and Amir Gholami

    Shashank Subramanian, Peter Harrington, Kurt Keutzer, Wahid Bhimji, Dmitriy Morozov, Michael W. Mahoney, and Amir Gholami. Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior. InAdvances in Neural Information Processing Systems, 2023

    2023

  20. [20]

    Morph: Pde foundation models with arbitrary data modality.arXiv preprint arXiv:2509.21670, 2025

    Mahindra Singh Rautela, Alexander Most, Siddharth Mansingh, Bradley C Love, Ayan Biswas, Diane Oyen, and Earl Lawrence. Morph: Pde foundation models with arbitrary data modality.arXiv preprint arXiv:2509.21670, 2025

    work page arXiv 2025

  21. [21]

    Physix: A foundation model for physics simulations,

    Tung Nguyen, Arsh Koneru, Shufan Li, et al. Physix: A foundation model for physics simulations.arXiv preprint arXiv:2506.17774, 2025

    work page arXiv 2025

  22. [22]

    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 Forty-first International Conference on Machine Learning, 2024

    2024

  23. [23]

    Michael McCabe, Payel Mukhopadhyay, Tanya Marwah, Bruno Regaldo-Saint Blancard, Francois Rozet, Cris- tiana Diaconu, Lucas Meyer, Kaze W. K. Wong, Hadi Sotoudeh, Alberto Bietti, Irina Espejo, Rio Fear, Siavash Golkar, Tom Hehir, Keiya Hirashima, Geraud Krawezik, Francois Lanusse, Rudy Morel, Ruben Ohana, Liam Parker, Mariel Pettee, Jeff Shen, Kyunghyun Ch...

    2026

  24. [24]

    PROSE-FD: A multimodal PDE foundation model for learning multiple operators for forecasting fluid dynamics

    Yuxuan Liu, Jingmin Sun, Xinjie He, Griffin Pinney, Zecheng Zhang, and Hayden Schaeffer. PROSE-FD: A multimodal PDE foundation model for learning multiple operators for forecasting fluid dynamics. InNeurips 2024 Workshop Foundation Models for Science: Progress, Opportunities, and Challenges, 2024. 19 Emergent transfer of a physics foundation model from si...

    2024

  25. [25]

    PDE-transformer: Efficient and versatile trans- formers for physics simulations

    Benjamin Holzschuh, Qiang Liu, Georg Kohl, and Nils Thuerey. PDE-transformer: Efficient and versatile trans- formers for physics simulations. InForty-second International Conference on Machine Learning, 2025

    2025

  26. [26]

    VICON: Vision in-context operator networks for multi-physics fluid dynamics prediction.Transactions on Machine Learning Research, 2026

    Yadi Cao, Yuxuan Liu, Liu Yang, Rose Yu, Hayden Schaeffer, and Stanley Osher. VICON: Vision in-context operator networks for multi-physics fluid dynamics prediction.Transactions on Machine Learning Research, 2026

    2026

  27. [27]

    Yeh, Jean Kossaifi, Kamyar Azizzadenesheli, and Anima Anandkumar

    Md Ashiqur Rahman, Robert Joseph George, Mogab Elleithy, Daniel Leibovici, Zongyi Li, Boris Bonev, Colin White, Julius Berner, Raymond A. Yeh, Jean Kossaifi, Kamyar Azizzadenesheli, and Anima Anandkumar. Pre- training codomain attention neural operators for solving multiphysics PDEs. InThe Thirty-eighth Annual Con- ference on Neural Information Processing...

    2024

  28. [28]

    Towards a foundation model for partial differential equations: Multi-operator learning and extrapolation, 2025

    Jingmin Sun, Yuxuan Liu, Zecheng Zhang, and Hayden Schaeffer. Towards a foundation model for partial differential equations: Multi-operator learning and extrapolation, 2025

    2025

  29. [29]

    DISCO: learning to DISCover an evolution operator for multi- physics-agnostic prediction

    Rudy Morel, Jiequn Han, and Edouard Oyallon. DISCO: learning to DISCover an evolution operator for multi- physics-agnostic prediction. InForty-second International Conference on Machine Learning, 2025

    2025

  30. [30]

    Test-time generalization for physics through neural operator splitting.ArXiv, abs/2602.00884, 2026

    Louis Serrano, Jiequn Han, Edouard Oyallon, Shirley Ho, and Rudy Morel. Test-time generalization for physics through neural operator splitting.ArXiv, abs/2602.00884, 2026

    work page arXiv 2026

  31. [31]

    Tadpole: Autoencoders as foundation models for 3d pdes with online learning, 2026

    Qiang Liu, Felix Koehler, Benjamin Holzschuh, and Nils Thuerey. Tadpole: Autoencoders as foundation models for 3d pdes with online learning, 2026

    2026

  32. [32]

    Ai for scientific discovery is a social problem.Patterns, 7(3):101497, March 2026

    Georgia Channing and Avijit Ghosh. Ai for scientific discovery is a social problem.Patterns, 7(3):101497, March 2026

    2026

  33. [33]

    D. H. Sharp. An overview of Rayleigh–Taylor instability.Physica D: Nonlinear Phenomena, 12:3–18, 1984

    1984

  34. [34]

    Cabot and Andrew W

    William H. Cabot and Andrew W. Cook. Reynolds number effects on Rayleigh–Taylor instability with possible implications for type ia supernovae.Nature Physics, 2:562–568, 2006

    2006

  35. [35]

    Instabilities and mixing in inertial confinement fusion.Annual Review of Fluid Mechanics, 57:197–225, 2025

    Ye Zhou, James D Sadler, and Omar A Hurricane. Instabilities and mixing in inertial confinement fusion.Annual Review of Fluid Mechanics, 57:197–225, 2025

    2025

  36. [36]

    K. I. Read. Experimental investigation of turbulent mixing by Rayleigh–Taylor instability.Physica D: Nonlinear Phenomena, 12(1-3):45–58, 1984

    1984

  37. [37]

    Snider and Malcolm J

    Dale M. Snider and Malcolm J. Andrews. Rayleigh–Taylor and shear driven mixing with an unstable thermal stratification.Physics of Fluids, 6(10):3324–3336, 1994

    1994

  38. [38]

    Ramaprabhu and M

    P. Ramaprabhu and M. J. Andrews. Experimental investigation of Rayleigh–Taylor mixing at small atwood numbers.Journal of Fluid Mechanics, 502:233–271, 2004

    2004

  39. [39]

    D. H. Olson and J. W. Jacobs. Experimental study of Rayleigh–Taylor instability with a complex initial pertur- bation.Physics of Fluids, 21(3):034103–034103–13, March 2009

    2009

  40. [40]

    D. L. Youngs. Modelling turbulent mixing by Rayleigh–Taylor instability.Physica D: Nonlinear Phenomena, 37:270–287, 1989

    1989

  41. [41]

    Dimonte and M

    G. Dimonte and M. Schneider. Turbulent Rayleigh–Taylor instability experiments with variable acceleration. Physical Review E, 54:3740–3743, 1996

    1996

  42. [42]

    S. B. Dalziel, P. F. Linden, and D. L. Youngs. Self-similarity and internal structure of turbulence induced by Rayleigh–Taylor instability.Journal of Fluid Mechanics, 399:1–48, 1999

    1999

  43. [43]

    Mueschke and Oleg Schilling

    Nicholas J. Mueschke and Oleg Schilling. Investigation of Rayleigh–Taylor turbulence and mixing using ex- perimentally measured initial conditions. I. comparison to experimental data.Physics of Fluids, 21(1):014106, 2009

    2009

  44. [44]

    Glimm, D

    J. Glimm, D. H. Sharp, T. Kaman, and H. Lim. New directions for Rayleigh–Taylor mixing.Philosophical Transactions of the Royal Society A, 371:20120183, 2013

    2013

  45. [45]

    M. J. Andrews and S. B. Dalziel. Small atwood number Rayleigh–Taylor experiments.Philosophical Transac- tions of the Royal Society A, 368:1663–1679, 2010

    2010

  46. [46]

    David L. Youngs. Rayleigh–Taylor mixing: direct numerical simulation and implicit large eddy simulation. Physica Scripta, 92(7):074006, 2017

    2017

  47. [47]

    Youngs, Andris Dimits, Stephan Weber, Martin Marinak, Stephan Wunsch, Christopher Garasi, Allen Robinson, Malcolm J

    Guy Dimonte, David L. Youngs, Andris Dimits, Stephan Weber, Martin Marinak, Stephan Wunsch, Christopher Garasi, Allen Robinson, Malcolm J. Andrews, Praveen Ramaprabhu, Alan C. Calder, Bruce Fryxell, Joseph Biello, L. Jon Dursi, Peter MacNeice, Kevin Olson, Paul Ricker, Robert Rosner, Frank Timmes, Henry Tufo, Yuan-Nan Young, and Michael Zingale. A compara...

    2004

  48. [48]

    D. L. Youngs. Numerical simulation of mixing by Rayleigh–Taylor and Richtmyer–Meshkov instabilities.Laser and Particle Beams, 12(4):725–750, 1994

    1994

  49. [49]

    Molecular mixing in Rayleigh–Taylor instability.Journal of Fluid Mechanics, 265:97–124, 1994

    PF Linden, JM Redondo, and DL Youngs. Molecular mixing in Rayleigh–Taylor instability.Journal of Fluid Mechanics, 265:97–124, 1994

    1994

  50. [50]

    Cook, William Cabot, and Paul L

    Andrew W. Cook, William Cabot, and Paul L. Miller. The mixing transition in Rayleigh–Taylor instability. Journal of Fluid Mechanics, 511:333–362, 2004

    2004

  51. [51]

    Gregory C. Burton. Study of ultrahigh atwood-number Rayleigh–Taylor mixing dynamics using the nonlinear large-eddy simulation method.Physics of Fluids, 23(4):045106, 2011

    2011

  52. [52]

    V . P. Statsenko, Yu. V . Yanilkin, and V . A. Zhmaylo. Direct numerical simulation of turbulent mixing.Philo- sophical Transactions of the Royal Society A, 371:20120216, 2013

    2013

  53. [53]

    Petersen

    Daniel Livescu, Tie Wei, and Mark R. Petersen. Direct numerical simulations of Rayleigh–Taylor instability. Journal of Physics: Conference Series, 318(8):082007, 2011

    2011

  54. [54]

    Livescu, J

    D. Livescu, J. R. Ristorcelli, M. R. Petersen, and R. A. Gore. New phenomena in variable-density Rayleigh– Taylor turbulence.Physica Scripta, T142:014015, 2010

    2010

  55. [55]

    D. L. Youngs. Application of MILES to Rayleigh–Taylor and Richtmyer–Meshkov mixing. InAIAA Paper 2003-4102, 2003

    2003

  56. [56]

    D. L. Youngs. The density ratio dependence of self-similar Rayleigh–Taylor mixing.Philosophical Transactions of the Royal Society A, 371:20120173, 2013

    2013

  57. [57]

    Rayleigh–Taylor and Richtmyer–Meshkov instability induced flow, turbulence, and mixing

    Ye Zhou. Rayleigh–Taylor and Richtmyer–Meshkov instability induced flow, turbulence, and mixing. ii.Physics Reports, 723–725:1–160, 2017

    2017

  58. [58]

    Cambridge University Press, 2024

    Ye Zhou.Hydrodynamic Instabilities and Turbulence: Rayleigh–Taylor, Richtmyer–Meshkov, and Kelvin– Helmholtz Mixing. Cambridge University Press, 2024

    2024

  59. [59]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede liu, Kaushik Bhattacharya, An- drew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2021

    2021

  60. [60]

    Multi-grid ten- sorized fourier neural operator for high- resolution PDEs.Transactions on Machine Learning Research, 2024

    Jean Kossaifi, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, and Anima Anandkumar. Multi-grid ten- sorized fourier neural operator for high- resolution PDEs.Transactions on Machine Learning Research, 2024

    2024

  61. [61]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation.CoRR, abs/1505.04597, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  62. [62]

    Rayleigh–Taylor and Richtmyer–Meshkov instability induced flow, turbulence, and mixing

    Ye Zhou. Rayleigh–Taylor and Richtmyer–Meshkov instability induced flow, turbulence, and mixing. i.Physics Reports, 720–722:1–136, 2017

    2017

  63. [63]

    Mohan, Nicholas Daniel, Michael Chertkov, and Daniel Livescu

    Arvind T. Mohan, Nicholas Daniel, Michael Chertkov, and Daniel Livescu. Spatio-temporal deep learning mod- els of 3D turbulence with physics informed diagnostics.Journal of Turbulence, 21(9–10):484–524, 2020

    2020

  64. [64]

    Long-term predictions of turbulence by implicit U-Net enhanced Fourier neural operator.Physics of Fluids, 35(7):075145, 2023

    Zhijie Li, Wenhui Peng, Zelong Yuan, and Jianchun Wang. Long-term predictions of turbulence by implicit U-Net enhanced Fourier neural operator.Physics of Fluids, 35(7):075145, 2023

    2023

  65. [65]

    Prediction of turbulent channel flow using Fourier neural operator-based machine-learning strategy.Physical Review Fluids, 9:084604, 2024

    Yunpeng Wang, Zhijie Li, Zelong Yuan, Wenhui Peng, Tianyuan Liu, and Jianchun Wang. Prediction of turbulent channel flow using Fourier neural operator-based machine-learning strategy.Physical Review Fluids, 9:084604, 2024

    2024

  66. [66]

    Adam Subel, Ashesh Chattopadhyay, Yifei Guan, and Pedram Hassanzadeh. Data-driven subgrid-scale modeling of forced Burgers turbulence using deep learning with generalization to higher Reynolds numbers via transfer learning.Physics of Fluids, 33(3):031702, 2021

    2021

  67. [67]

    Yifei Guan, Ashesh Chattopadhyay, Adam Subel, and Pedram Hassanzadeh. Stable a posteriori LES of 2D turbulence using convolutional neural networks: Backscattering analysis and generalization to higher Re via transfer learning.Journal of Computational Physics, 458:111090, 2022

    2022

  68. [68]

    Explaining the physics of transfer learning in data-driven turbulence modeling.PNAS Nexus, 2(3):pgad015, 2023

    Adam Subel, Yifei Guan, Ashesh Chattopadhyay, and Pedram Hassanzadeh. Explaining the physics of transfer learning in data-driven turbulence modeling.PNAS Nexus, 2(3):pgad015, 2023

    2023

  69. [69]

    P3d: Highly scalable 3d neural sur- rogates for physics simulations with global context

    Benjamin Holzschuh, Georg Kohl, Florian Redinger, and Nils Thuerey. P3d: Highly scalable 3d neural sur- rogates for physics simulations with global context. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  70. [70]

    Learning turbulent flows with generative models: Super-resolution, forecasting, and sparse flow reconstruction.arXiv preprint arXiv:2509.08752, 2025

    Vivek Oommen, Siavash Khodakarami, Aniruddha Bora, Zhicheng Wang, and George Em Karniadakis. Learning turbulent flows with generative models: Super-resolution, forecasting, and sparse flow reconstruction.arXiv preprint arXiv:2509.08752, 2025. 21 Emergent transfer of a physics foundation model from simulation to laboratory turbulenceA PREPRINT

    work page arXiv 2025

  71. [71]

    Fourier neural operator for large eddy simulation of compressible Rayleigh–Taylor turbulence.Physics of Fluids, 36(7):075165, 2024

    Tengfei Luo, Zhijie Li, Zelong Yuan, Wenhui Peng, Tianyuan Liu, Liangzhu Wang, and Jianchun Wang. Fourier neural operator for large eddy simulation of compressible Rayleigh–Taylor turbulence.Physics of Fluids, 36(7):075165, 2024

    2024

  72. [72]

    PhD thesis, University of Cambridge, 2018

    Christopher Brown.Exploring the effects of an obstruction on the evolution of the Rayleigh–Taylor Instability. PhD thesis, University of Cambridge, 2018

    2018

  73. [73]

    George, J

    E. George, J. Glimm, X.-L. Li, A. Marchese, and Z.-L. Xu. A comparison of experimental, theoretical, and nu- merical simulation Rayleigh–Taylor mixing rates.Proceedings of the National Academy of Sciences, 99(5):2587– 2592, 2002

    2002

  74. [74]

    Stuart B. Dalziel. Rayleigh–Taylor instability: experiments with image analysis.Dynamics of Atmospheres and Oceans, 20(1):127–153, 1993. American Geophysical Union Ocean Sciences Meeting

    1993

  75. [75]

    Ramaprabhu, Guy Dimonte, and M

    P. Ramaprabhu, Guy Dimonte, and M. J. Andrews. A numerical study of the influence of initial perturbations on the turbulent Rayleigh–Taylor instability.Journal of Fluid Mechanics, 536:285–319, 2005

    2005

  76. [76]

    Ramaprabhu, D

    Guy Dimonte, P. Ramaprabhu, D. L. Youngs, M. J. Andrews, and R. Rosner. Recent advances in the turbulent Rayleigh–Taylor instabilitya).Physics of Plasmas, 12(5):056301, May 2005

    2005

  77. [77]

    Fermi and J

    E. Fermi and J. von Neumann. Taylor instability of incompressible liquids. part 1. Taylor instability of an incompressible liquid. part 2. Taylor instability at the boundary of two incompressible liquids. Technical report, United States Atomic Energy Commission / Los Alamos Scientific Laboratory, 8 1953

    1953

  78. [78]

    D. L. Youngs. Numerical simulation of turbulent mixing by Rayleigh–Taylor instability.Physica D: Nonlinear Phenomena, 12:32–44, 1984

    1984

  79. [79]

    J. R. Ristorcelli and T. T. Clark. Rayleigh–Taylor turbulence: self-similar analysis and direct numerical simula- tions.Journal of Fluid Mechanics, 507:213–253, 2004

    2004

  80. [80]

    Th `ese de doctorat, ´Ecole normale sup´erieure de Cachan (ENS Cachan), 2011

    Romain Watteaux.D ´etection des grandes structures turbulentes dans les couches de m´elange de type Rayleigh– Taylor en vue de la validation de mod`eles statistiques turbulents bi-structure. Th `ese de doctorat, ´Ecole normale sup´erieure de Cachan (ENS Cachan), 2011. HAL thesis ID: tel-00669707

    2011

Showing first 80 references.