README.md

---
license: cc-by-nc-4.0
pretty_name: CTF4Nuclear MSFR
language:
- en
size_categories:
- 10K<n<100K
tags:
- scientific-ml
- nuclear-engineering
- molten-salt-reactor
- physics
- benchmark
- surrogate-modelling
- time-series
- pde
---

# CTF4Nuclear — Molten Salt Fast Reactor (MSFR) benchmark

A spatio-temporal benchmark for reduced-order and surrogate modelling of
**Molten Salt Fast Reactor (MSFR)** multiphysics simulations. The dataset
consists of snapshot matrices from coupled neutronics + thermal-hydraulics
simulations on a 2D axial-symmetric slice of the EVOL MSFR geometry,
discretised on a 3 880-node mesh. Each snapshot is a state vector of
length **19 400** = 5 fields × 3 880 nodes, with the fields concatenated
in the order

```
[qPrompt, qDecay, T, Ux, Uz]   # prompt power, decay power, temperature, two velocity components
```

Snapshots are saved every Δ*t* = 0.05 s of simulated physical time.

This dataset is the public training-and-config component of the
**CTF4Nuclear** Common Task Framework. The accompanying test sets are
held out and used to score submissions to the public leaderboard at
[`huggingface.co/spaces/ctf4science/ctf4nuclear-msfr-leaderboard`](https://huggingface.co/spaces/ctf4science/ctf4nuclear-msfr-leaderboard).

## Quick start

```python
from huggingface_hub import snapshot_download

# 1. Download into the layout the ctf4science loader expects
snapshot_download(
    repo_id="ctf4science/ctf4nuclear-msfr",
    repo_type="dataset",
    local_dir="ctf4science/data/msfr",
)

# 2. Load any of the 9 evaluation pairs
from ctf4science.data_module import load_dataset
train_list, init_data = load_dataset("msfr", pair_id=1, transpose=True)
```

The `ctf4science` Python package is open source at
[`github.com/CTF-for-Science/ctf4science`](https://github.com/CTF-for-Science/ctf4science).
A ready-to-run *Baseline Last* notebook is provided in the leaderboard
repo under
[`baselines/baseline_last.ipynb`](https://huggingface.co/spaces/ctf4science/ctf4nuclear-msfr-leaderboard/tree/main/baselines).

## Files

| File | Shape | Used in (pairs / metrics) |
|---|---|---|
| `train/X1train.npz`  | 2000 × 19400 | pair 1 (E1, E2) |
| `train/X2train.npz`  | 2000 × 19400 | pair 2 (E3), pair 3 (E4) |
| `train/X3train.npz`  | 2000 × 19400 | pair 4 (E5), pair 5 (E6) |
| `train/X4train.npz`  |  500 × 19400 | pair 6 (E7, E8) |
| `train/X5train.npz`  |  500 × 19400 | pair 7 (E9, E10) |
| `train/X6train.npz`  | 1500 × 19400 | pair 8, pair 9 (parametric) |
| `train/X7train.npz`  | 1500 × 19400 | pair 8, pair 9 (parametric) |
| `train/X8train.npz`  | 1500 × 19400 | pair 8 (parametric) |
| `train/X9train.npz`  |  500 × 19400 | pair 8 (initialization for E11) |
| `train/X10train.npz` | 1500 × 19400 | pair 9 (parametric) |
| `train/X11train.npz` |  500 × 19400 | pair 9 (initialization for E12) |
| `nodes.npy`          | 3880 × 2     | (x, z) mesh coordinates in metres |
| `msfr.yaml`          | —            | benchmark configuration (pairs, metrics, matrix metadata) |
| `croissant.json`     | —            | Croissant 1.0 RAI metadata |

Each `.npz` stores a single float32 array under key `X` with shape
`(n_timesteps, 19400)`. The held-out test files `X1test.npz`–`X9test.npz`
referenced in `msfr.yaml` are **not** released — they are used by the
leaderboard's scoring backend only.

## Evaluation pairs and metrics

The benchmark defines **9 evaluation pairs** producing **12 metric
scores** (E1–E12) across three task families:

| Task family | Metrics | Description |
|---|---|---|
| Short-time forecasting | E1, E7, E9, E11, E12 | Relative L2 error over the first *k* = 20 forecast steps |
| Long-time forecasting  | E2, E4, E6, E8, E10  | Spectral L2 error on the last *k* = 200 steps (100 Fourier modes) |
| Reconstruction (denoising) | E3, E5            | Relative L2 error over the full denoised trajectory |

The composite score is the mean of E1–E12, with each metric clipped to
[−100, 100]. See [the `ctf4science` documentation](https://github.com/CTF-for-Science/ctf4science/blob/main/docs/source/evaluation_metrics.rst)
for the formal definitions.

## Data generation

Trajectories were produced by the `msfrPimpleFoam` OpenFOAM v2312 solver
(Aufiero et al., 2014) on the EVOL MSFR geometry (Brovchenko et al.,
2013). Four simulations span a 1D sweep of the momentum-source base
level (pump velocity) from 10 to 15. The solver couples
RANS computational fluid dynamics (realisable *k*–ε) with multi-group
diffusion neutronics. All fields are z-score normalised per field with
global standard deviations (means and std are available from the
dataset authors).

## Citation

```bibtex
@misc{riva2026msfr,
  title  = {A spatio-temporal benchmark for reduced-order and surrogate
            modelling of Molten Salt Fast Reactor (MSFR) multiphysics
            simulations},
  author = {Riva, Stefano and Introini, Carolina and Cammi, Antonio},
  year   = {2026},
  url    = {https://huggingface.co/datasets/ctf4science/ctf4nuclear-msfr},
}

@article{aufiero2014openfoam,
  title   = {Development of an OpenFOAM model for the Molten Salt Fast
             Reactor transient analysis},
  author  = {Aufiero, Manuele and Cammi, Antonio and Geoffroy, Olivier
             and Losa, Mario and Luzzi, Lelio and Ricotti, Marco E. and
             Rouch, Herv\'e},
  journal = {Chemical Engineering Science},
  volume  = {111},
  pages   = {390--401},
  year    = {2014},
  doi     = {10.1016/j.ces.2014.03.003},
}

@misc{brovchenko2013evol,
  title  = {Optimization of the pre-conceptual design of the MSFR},
  author = {Brovchenko, Mariya and Merle Lucotte, Elsa and Rouch, Herv\'e
            and others},
  year   = {2013},
  url    = {https://www.janleenkloosterman.nl/reports/evol_d22_201309.pdf},
}
```

## License

[Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/).

## Mirrors

- **Hugging Face** (this repo) — primary, fastest for `snapshot_download`
- [Open Science Framework](https://osf.io/6rzhm/) — original distribution, full tarball

## Contact

Stefano Riva — `stefano.riva@autodesk.com`