barella23/n-u-c-suppl-material

Dataset Description

Each dataset comes from a numerical simulation of a nuclear fission or MHD for fusion system. Most datasets produce spatio-temporal snapshot matrices of several coupled physical fields (e.g. temperature, velocity, power density) evaluated over a fixed spatial mesh; the pke-hpmr dataset is a low-dimensional ODE system coupling point-kinetics with a Lorenz attractor driving reactivity. The current dataset includes all the possible modelling strategies for nuclear systems: lumped models (0D), system codes (1D), and multi-physics simulations (2D, 3D).

Datasets currently included:

Dataset	Type	System	Fields / state variables
msfr	spatio-temporal	Molten Salt Fast Reactor (2D axisymmetric wedge, EVOL geometry)	prompt/decay power density, temperature, 2 velocity components
mhd-2cyls	spatio-temporal	Magneto-hydrodynamic channel flow slice (fusion blanket coolant, PbLi MHD)	temperature, pressure, 2 velocity components
dynasty	spatio-temporal	DYNASTY thermal–hydraulic model (1D natural circulation)	mass flow rate, temperatures
pke-hpmr	dynamical (ODE)	Point kinetics + Lorenz reactivity (HPMR)	log power, normalized precursors, Lorenz (x, y, z)

Repository Contents

assets/
  <dataset>.gif   # animated preview of each dataset (generated by visualize_data.ipynb)
data/
  <dataset>/
    <dataset>.yaml  # dataset configuration (train/test pairs, metrics, matrix metadata)
    nodes.npy       # spatial mesh node coordinates (spatio-temporal datasets only)
    train/          # training snapshot matrices (X*train.npz)
visualize_data.ipynb  # loads a train file per dataset and plots representative fields

For spatio-temporal datasets, each snapshot matrix X has shape (n_snapshots, n_fields * mesh_size): every row stacks all physical fields for one time step, in the order given by fields in the dataset's yaml config. nodes.npy gives the coordinates of each of the mesh_size spatial points.

For the dynamical pke-hpmr dataset, X has shape (n_snapshots, 6) with columns [log n, c1/c1*, c2/c2*, x, y, z] (physics-scaled point-kinetics state plus Lorenz variables). There is no spatial mesh.

Note: the test/ split is intentionally withheld from the public upload to preserve it as a hidden evaluation set, consistent with the CTF's goal of preventing overfitting to the test data.

Dataset Previews

Animated examples from the first training matrix of each dataset (assets/<dataset>.gif). For 2D meshes (msfr, mhd-2cyls), each frame shows a spatial snapshot of every field; for the 1D dynasty dataset, the GIF tracks the simulation time on the space–time evolution plots; for pke-hpmr, the GIF shows PKE state trajectories and the Lorenz attractor with a moving time marker. Color limits are fixed per field from the min/max over the full training matrix so the colorbar does not rescale between frames (where applicable).

pke-hpmr

Point kinetics with Lorenz-driven reactivity: log power and normalized precursor concentrations alongside the Lorenz attractor. Loaded from pair 1 (X1train.npz, Δt = 0.01).

dynasty

DYNASTY reduced-order model: mass flow rate (time series) and axial channel temperatures (space–time evolution).

msfr

Molten Salt Fast Reactor: prompt/decay power density, temperature, and velocity components on a 2D axisymmetric wedge mesh.

mhd-2cyls

Magneto-hydrodynamic channel flow: temperature, pressure, and velocity components in a fusion blanket coolant channel.

Visualizing the Data

visualize_data.ipynb is a quick sanity-check notebook: it auto-discovers every dataset under data/, loads one training file per dataset (pair 1 for pke-hpmr), and plots representative fields — spatial snapshots and mean-field time traces for mesh-based datasets, coupled state trajectories and a 3D Lorenz plot for pke-hpmr — confirming the released data is populated and non-trivial. The final cell exports the GIF previews above to assets/. Requires numpy, matplotlib, pyyaml, pillow, and jupyter.

Model Configuration Files

config-files-msfr/<model>/ contains the MSFR-specific yaml configuration files (hyperparameters and tuned parameters) used to produce the paper's results for each baseline method, collected from each model's third-party implementation. This makes the exact configurations used for benchmarking part of the reproducible supplementary material, independent of the model source code itself.

Results and Visualization

results/ contains the scores (msfr.csv) and the notebook/helper module (VisualizeScores.ipynb, visualize_scores_helpers.py) used to generate every figure and table in the paper. results/SensorSensitivityEIM/ contains a separate sensor-placement sensitivity study (EIM/DEIM-based) exploring how reconstruction accuracy depends on the number and choice of sensors.

Submitting to the Benchmark

The test/ splits are withheld and are not part of this repository. Final scores are therefore not self-reported: they are computed by the organisers by running a submitted model against the held-out test data through the evaluation pipeline. This is what makes this a common task framework rather than a dataset release, and it is the reason scores obtained here are comparable across entries.

What to submit

A submission consists of:

1. Model code that can be run through the openly available ctf4science interface, i.e. a run.py entry point that loads the training matrices for a given pair_id, fits the model, and returns a prediction matrix of the expected shape.

2. A configuration file per dataset, following the layout of config-files-msfr/<model>/. The minimal form is:

   dataset:
     name: msfr        # msfr | dynasty | mhd-2cyls | pke-hpmr
     pair_id: all      # or a specific pair, 1-9
   model:
     name: YourModel
     # model-specific hyperparameters below
   

3. A short description of the method, including the hyperparameter search space explored. Hyperparameters must be selected using the released training data only.

Rules

• Train on the released data only. Model selection and hyperparameter tuning must use the training matrices in data/<dataset>/train/. Any tuning against the test set invalidates the entry.
• One model per task is allowed. Each of the nine pairs may be fitted separately, but each configuration must be fixed before evaluation.
• Failed tasks score -100. A method that cannot produce a prediction for a task (e.g. through memory limits) receives the minimum score for that task, which enters the average.
• Report the observation design for the sparse state-estimation task: the measured field, the number of sensors, and whether sensor positions are prescribed or selected by the method.