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, 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 arXiv: 2606.01470

Model description

This checkpoint is a finetuned version of Walrus, a physics foundation model for continuum dynamics. It was specialized to Rayleigh-Taylor instability (RTI) using idealized direct numerical simulation (DNS) data.

In the paper, this checkpoint is used for:

RTI emulation on held-out DNS The model is evaluated on held-out RTI DNS rollouts using physical diagnostics including morphology, mixing-layer growth, energy spectra, and global energy-budget terms.
Zero-shot transfer to stably stratified RTI The same DNS-finetuned model is initialized from stably stratified RTI conditions, a buoyancy regime absent from finetuning, and evaluated on whether it correctly slows and confines mixing-layer growth.

This checkpoint is not the checkpoint used for the zero-shot transfer to sliding-barrier laboratory experiments. The sliding-barrier laboratory data and corresponding checkpoint will be made available in the near future.

Repository contents

This repository contains:

model.safetensors: preferred model weights file.
coalesced.pth: PyTorch checkpoint, provided for compatibility with existing Walrus loading code.
extended_config.yaml: model and experiment configuration associated with the checkpoint.

If possible, use model.safetensors for loading the weights, as it avoids Python pickle deserialization.

Intended use

This checkpoint is intended for research use in scientific machine learning, fluid dynamics, and physics foundation models. It may be useful for:

reproducing the RTI DNS emulation results from the paper,
evaluating the model on held-out idealized RTI DNS rollouts,
studying transfer from unstratified RTI to stably stratified RTI,
analyzing learned representations for buoyancy-driven flow.

Training data

This checkpoint was finetuned on idealized RTI direct numerical simulation data in the Boussinesq regime. No laboratory experimental data were used to train this checkpoint.

Please see the paper for full details on the simulation setup, preprocessing, finetuning procedure, and evaluation protocols.

Evaluation

The model was evaluated using physically meaningful RTI diagnostics, including:

mixing-layer morphology,
mixing-layer growth coefficient,
kinetic-energy spectra,
global energy-budget terms,
zero-shot transfer to stably stratified RTI.

For the DNS emulation task, the model recovers characteristic RTI physics over long autoregressive rollouts. For the stratified transfer task, the DNS-finetuned model correctly slows and confines mixing-layer growth when initialized from stably stratified conditions, despite stable stratification being absent from finetuning.

Citation

If you use this checkpoint, please cite:

bibtex

@misc{mukhopadhyay2026emergenttransferphysicsfoundation,
      title={Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence}, 
      author={Payel Mukhopadhyay and Stefan S. Nixon and Romain Watteaux and Michael McCabe and Alberto Bietti and Kyunghyun Cho and Cristiana Diaconu and Irina Espejo Morales and David Fouhey and Siavash Golkar and Tom Hehir and Shirley Ho and Jake Kovalic and Geraud Krawezik and Francois Lanusse and Tanya Marwah and Rudy Morel and Mariel Pettee and Helen Qu and Jeff Shen and Hadi Sotoudeh and Stuart B. Dalziel and Miles Cranmer},
      year={2026},
      eprint={2606.01470},
      archivePrefix={arXiv},
      primaryClass={physics.flu-dyn},
      url={https://arxiv.org/abs/2606.01470}, 
}

Contact

For questions, please refer to the paper or contact pm858@cam.ac.uk.

pmukhop/rti-walrus-model

Walrus Rayleigh Taylor Instability (RTI) Checkpoint