---
license: mit
tags:
- physics
- fluid-dynamics
- scientific-machine-learning
- foundation-models
- rayleigh-taylor-instability
- walrus
pretty_name: Rayleigh Taylor Instability 128^3 Boussinesq
---
# Rayleigh-Taylor Instability (RTI) Boussinesq Regime Dataset

**One-line description:** Three-dimensional miscible Rayleigh-Taylor instability simulations in the incompressible Boussinesq regime, provided at effective `128^3` resolution by downsampling native `256^3` DNS trajectories.

**Longer description:** This dataset contains 3D direct numerical simulation (DNS) trajectories of Rayleigh-Taylor instability (RTI) in the incompressible Boussinesq regime. The native simulations were generated at `256^3` resolution with the **TurMix3D** code and are provided here in downsampled `128^3` form. The dataset is intended for scientific machine learning tasks such as ML surrogate modeling. The dataset is formatted as "The Well" benchmark format.

**Associated paper:** https://arxiv.org/abs/2606.01470

**Original native-resolution `256^3` dataset:** https://huggingface.co/datasets/pmukhop/rti-dataset-boussinesq_256x3

**Stratified RTI dataset:** https://huggingface.co/datasets/pmukhop/rti-stratified-data


**Domain experts:** Stefan S. Nixon, University of Cambridge; Romain Watteaux, CEA DAM; Stuart B. Dalziel, University of Cambridge.

**Code or software used to generate the data:** TurMix3D.
**TurMix3D reference:** [Romain Watteaux PhD thesis](https://theses.hal.science/tel-00669707/document)

## About the data

This dataset contains five statistically independent three-dimensional DNS realizations of miscible RTI in the incompressible Boussinesq regime. The fields are:

* normalized concentration field
* normalized velocity components

The data are stored as HDF5 trajectory files and are organized into fixed train, validation, and test splits:

```text id="vk4pxs"
data/
  train/
    *.h5
  valid/
    *.h5
  test/
    *.h5
stats.yaml
```

## Initial conditions

The five DNS trajectories use idealized, statistically independent initial interface perturbations. The initial displacement field is constructed in Fourier space by exciting modes in a narrow annular band around a mean wavenumber `k0`. The modes in this annulus are assigned random phases, and conjugate symmetry is enforced so that the inverse Fourier transform gives a real-valued physical interface perturbation.

The perturbation spectrum is controlled by three dimensionless parameters:

* perturbation Reynolds number `Re`
* spectral bandwidth `B`
* initial steepness `S`

For the native `256^3` simulations, these parameters are `{Re, B, S} = {7, 0.3, 0.1}`. Each of the five DNS realizations uses a different random initialization of the perturbation spectrum.

These simulations are different from the RTI simulations in **the Well**. In particular, the Well RTI examples use radially log-normal profiles in Fourier space for initialization, whereas this dataset uses annular narrow-band perturbations with random phases.

## Equations

The simulations are in the incompressible miscible Boussinesq regime with Atwood number `At = 0.0625`.

The concentration field is normalized. The velocity field is normalized by the free-fall velocity scale `sqrt(At g H)`, with `g = 0.02` and `H` the domain length.

The governing equations are:

```text id="2aoy0y"
∂t ρ + ∇ · (ρ u) = 0
```

```text id="jdp1vx"
∂t(ρ u) + ∇ · (ρ u u) = -∇p + ∇ · τ + ρ g
```

```text id="vj3j5t"
∇ · u = -κ ∇ · (∇ρ / ρ)
```

with deviatoric stress tensor

```text id="e5n08i"
τ = ρ ν [ ∇u + (∇u)^T - (2/3)(∇ · u) I ]
```

where:

* `ρ` is density
* `u` is velocity
* `p` is pressure
* `g` is gravity
* `κ` is molecular diffusivity
* `ν` is kinematic viscosity
* `I` is the identity matrix


## Citation

If you use this dataset, the native-resolution version of this dataset, or any associated datasets, please cite the associated paper:

```bibtex id="pzpooh"
@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}, 
}
```