---
license: mit
---

This repository contains pre-trained weights for the large-sized GyroSwin variant introduced in [GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations](https://arxiv.org/abs/2510.07314).

The model was pre-trained on turbulent gyrokinetics simulations generated via the numerical code GKW.
For instructions on the data generation process see the `data_generation` directory of our [github repository](https://github.com/ml-jku/neural-gyrokinetics).

The large variant of GyroSwin was trained on 241 simulations of turbulent gyrokinetics simulations, amounting to 44,585 training samples, on the task of next-step prediction. In total the model consists of 998.3M parameters.

## Example usage

We provide an example on how to use the model for inference in our [github repository](https://github.com/ml-jku/neural-gyrokinetics).

## Citation

If you found our work useful, please consider citing it:

    @inproceedings{paischergyroswin,
        title={GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations},
        author={Paischer, Fabian and Galletti, Gianluca and Hornsby, William and Setinek, Paul and Zanisi, Lorenzo and Carey, Naomi and Pamela, Stanislas and Brandstetter, Johannes},
        booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}
    }