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

arXiv:2606.01470 · physics.flu-dyn, cs.AI, cs.LG · Submitted 2026-05-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We're looking at the paper: "Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence".

Mira: Here's the gist of the paper. Foundation models, pretrained on large, diverse datasets and adapted via finetuning, exhibit a key property of transfer: representations learned during pretraining generalize to new settings, reducing reliance on task-specific data.

Kai: This paradigm is being tested in the physical sciences regarding its applicability to laboratory experiments.

Mira: The study focuses on the Rayleigh-Taylor instability (RTI), where a discrepancy exists between idealized Direct Numerical Simulations (DNS) and laboratory experiments concerning the late-time mixing growth rate, alpha.

Kai: Standard machine learning models struggle with RTI, and the significant gap—laboratory rates being roughly three times higher than DNS values—serves as a stringent test for whether foundation models trained solely on simulations can generalize to sparse, noisy laboratory settings.

Paper summary: Kai: So to recap this paper on "Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence," the main question is whether physics foundation models can be usefully deployed on laboratory experiments at all.

Mira: They test this question using the Rayleigh-Taylor instability, which is this fluid instability where a lighter fluid sits above a heavier one and it erupts into complexity. The core thesis is that if models trained only on idealized simulations can generalize to real laboratory conditions they were never trained on, then their usefulness for actual physical discovery is fundamentally limited.

Lev: That limitation directly relates to the concept of simto-real transfer, which is how well a model performs when moving from simulation data to experimental data without any direct training on that experimental data.

Kai: The paper claims that by finetuning a foundation model, they can make it cross this divide. They show that models trained on idealized simulations can predict laboratory behavior and even unseen physical regimes, simply by having the right initial conditions or through careful adaptation.

Mira: What matters is that this isn't just about reproducing what was in the training set; it’s about generalising to new physical regimes that were never shown. They test this on RTI because it's a compelling system where the discrepancy between simulation and experiment is well-known, specifically regarding the late-time mixing growth rate alpha.

Lev: It matters because if we can get this transfer working robustly, it suggests that foundation models could be a new tool for exploring physical sciences where full simulations are too expensive or difficult to run.

Kai: Right, so they are looking at whether these models can bridge the gap between idealized simulation and the messy reality of lab experiments by learning physics from simulations and then adapting that learning.

Conclusion: Mira: So looking at the paper "Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence," it boils down to testing if we can trust these large models when they are applied outside their training environment, specifically in physical systems like fluid turbulence.

Kai: The authors suggest that the success here comes from finding this emergent transfer property. It means the model isn't just recalling simulation results; it’s learning a representation of physics that applies broadly.

Lev: What does this mean practically is that if we can get these models to work across different physical regimes, it changes how we think about using them for experimental science—it moves them from being just sophisticated data fitting tools to potential instruments for generating new physical insights.

Mira: So the implication is that we might be able to use foundation models as a way to quickly explore complex physical spaces where traditional simulation methods are too slow or computationally heavy.

Kai: It points toward a future where these models could help us bridge the gap between theoretical predictions and what we actually observe in the lab by learning from both worlds simultaneously, which is what they call simto-real transfer.

Payel Mukhopadhyay, Stefan S. Nixon, Romain Watteaux, Michael McCabe, Alberto Bietti, Kyunghyun Cho, Cristiana Diaconu, Irina Espejo Morales, David Fouhey, Siavash Golkar, Tom Hehir, §hirlie Ho3,4,5,9,, Èraud Krawezik ´3,9,, Franc¸ois Lanusse7,9,, Tanya Marwah9,, Rudy Morel3,9,, Mariel Pettee8,9,, Helen Qu3,9,, Jeff Shen5,9,, Ádi Sotoudeh1, Øud B. Dalziel1, Íles Cranmer1, Geraud Krawezik ´3,9,, Franc¸ois Lanusse7,9,, Tanya Marwah9,, Rudy Morel3,9,, Mariel Pettee8,9,, Helen Qu3,9,, Ádi Sotoudeh1, Øud B. Dalziel1, Íles Cranmer1

University of Cambridge · CEA DAM/DIF · Flatiron Institute · New York University · Princeton University · Yale University

physics.flu-dyn, cs.AI, cs.LG

Submitted: 2026-05-31

Updated: 2026-11-04

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 100/100

The gist: Foundation models, pretrained on large, diverse datasets and adapted via finetuning, exhibit a key property of transfer…

Key concepts

Rayleigh-Taylor Instability (RTI)
RTI is a fluid instability where a lighter fluid accelerates into a heavier one, causing small initial density perturbations to grow into chaotic, multiscale mixing. It is a fundamental phenomenon observed in everything from tabletop flows to supernova explosions.
Simulation-Experiment Gap
This refers to the long-standing discrepancy where idealized Direct Numerical Simulations (DNS) of RTI predict a slower growth rate ($\alpha \approx 0.02$) than what is measured in actual laboratory experiments ($\alpha \approx 0.06-0.07$). The paper investigates if foundation models can bridge this gap.
Foundation Model Transfer
This involves taking a large model pretrained on broad simulation data and adapting it through finetuning to perform well on a new, sparse, or noisy task—in this case, transferring knowledge from simulations to real laboratory conditions without extensive new experimental training data.

Terminology

Summary

Foundation models, pretrained on large, diverse datasets and adapted via finetuning, exhibit a key property of transfer: representations learned during pretraining generalize to new settings, reducing reliance on task-specific data. p. 0 This paradigm is being tested in the physical sciences regarding its applicability to laboratory experiments. p. 0 The study focuses on the Rayleigh-Taylor instability (RTI), where a discrepancy exists between idealized Direct Numerical Simulations (DNS) and laboratory experiments concerning the late-time mixing growth rate, alpha. p. 0 Standard machine learning models struggle with RTI, and the significant gap—laboratory rates being roughly three times higher than DNS values—serves as a stringent test for whether foundation models trained solely on simulations can generalize to sparse, noisy laboratory settings. p. 0

The researchers utilize Walrus [23], a foundation model for continuum dynamics pretrained on broad simulation data where RTI was excluded (Wpre). p. 1 This model is then finetuned on three or fewer DNS realizations of RTI to obtain W3D DNS. p. 0 The findings demonstrate several critical capabilities:

  1. Zero-Shot Transfer to Experimental Conditions: Walrus transfers zero-shot to sliding-barrier laboratory data, entering the experimentally observed growth band near alpha about 0.07 without any experimental training data. p. 1 This provides evidence that initial conditions are crucial in bridging the sim-experiment gap in alpha. p. 3

  2. Generalization to New Physical Regimes: Walrus also generalizes zero-shot to stably stratified RTI, correctly slowing mixing-layer growth by capturing buoyancy-driven flow physics absent from its training data.

  3. Sim-to-Real Transfer Capability: The pretraining-finetuning paradigm successfully achieves sim-to-real transfer and generalizes to qualitatively new physical regimes in a data-limited setting.

The results strongly suggest that foundation models can transfer learned physics from idealized simulation to real experimental conditions, even when the two occupy different physical regimes. p. 1 Specifically, the study shows that when initialized with experimental frames instead of DNS frames, the model shifts its prediction toward the higher- alpha growth regime characteristic of laboratory data. p. 0 This contrast reveals two key aspects: first, that the model has learned self-similarity physics (producing DNS-like growth from DNS initial conditions); and second, that this learned physics is not tied to a specific training distribution; when given experimental initial frames carrying large-scale structure absent in training, the model extrapolates correctly to a different self-similar regime. p. 3

Furthermore, the study details a two-stage finetuning process: an initial stage (using a higher learning rate) that establishes direct transfer from DNS to experiment, and a second, lighter finetuning stage (with a significantly smaller learning rate) that improves agreement with the barrier-release-driven early-time structure by exposing the model to experimental evolution. p. 3 This final adapted model, W2D DNS+Exp, qualitatively captures the complex transient behavior of real experiments. p. 3 The overall conclusion is that this pretraining-finetuning paradigm offers a promising, purely data-driven route for sim-to-real transfer in physical systems where laboratory data is scarce and idealized simulation falls short. p. 3 Limitations noted include the resolution of the DNS used being below that of highest-fidelity RTI benchmarks, and the need to fully understand how Walrus represents flow to confirm its emulation without traditional fluid dynamical modeling.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adapt foundation models to laboratory settings without experimental training data by finetuning on a small number of idealized simulation realizations, demonstrating that the transfer results provide independent, data-driven evidence that initial conditions play a crucial role in the longstanding sim-experiment gap in α.

  2. Enable zero-shot generalization to unseen physical regimes by showing that foundation models finetuned on idealized simulation can generalise well beyond their training conditions, predicting both real laboratory behavior and qualitatively new physical regimes, such as correctly predicting the confinement of a mixing layer under stably stratified conditions.

  3. Improve predictive accuracy for late-time growth rates in turbulent flows by using initial experimental frames to guide the model into the observed growth band, as the zero-shot rollout enters the experimentally observed growth band near α ≈ 0.07, without the model ever having seen experimental data of any kind.

  4. Enhance robustness to input noise and variability by employing patch-jittering during training, which is shown to make the alpha prediction robust to the randomness caused by the jitter (shown as a blue band).

  5. Reduce sample inefficiency for task-specific adaptation by showing that finetuning on a single 3D RTI realization already yields excellent spectral emulation, meaning models can achieve strong pointwise and local reconstruction with very little task-specific data.

  6. Bridge the sim-to-real divide by using a two-stage finetuning process where the model first specializes to RTI using simulation data and then makes a minimal adjustment with only two experimental samples to capture complex initial condition structures.

Sources

Related papers