ASSAS-PWR-1300 — Hugging Face mirror

ASTEC simulation outputs for severe-accident research and machine-learning surrogate modelling of a simplified generic Western-type 1300 MWe pressurized water reactor (PWR), produced within ASSAS — Artificial intelligence for the Simulation of Severe AccidentS.

This Hugging Face repository is a distribution mirror of the official RADAR KIT collection. All dataset DOIs belong to the original RADAR records. This mirror does not mint or replace any DOI.

Scope	Official identifier
Head dataset / entire collection	10.35097/bp03ve8f24y4w17m
Individual scenario / simulation record	One separate RADAR DOI per archived record; see radar_doi in scenarios.csv
Hugging Face mirror	Helmholtz-AI-Energy/ASSAS-PWR-1300

The supplied RADAR page currently describes the head record as in preparation, with no attached files; its DOI identifies the collection metadata record and does not by itself confirm publication.

The official RADAR landing page is the reference for the collection's archival metadata and citation. Scenario records are parts of that collection; their individual DOIs identify the specific data used. UUIDs in mirrored filenames are ASSAS upload identifiers, not DOIs.

Project and purpose

ASSAS develops a proof-of-concept severe-accident simulator and explores faster execution through code optimisation and AI surrogate models. Its project overview describes the generic PWR simulator and an openly available training database. Work Package 3, led by KIT, generates training databases; the software architecture explains how reference-code outputs support surrogate training and validation.

This collection contains simulation results, rather than measurements from operating reactors. Uses include time-dependent surrogate modelling, forecasting, and analysis of severe-accident progression. It includes loss-of-coolant accident (LOCA) and station-blackout (SBO) cases, with scenario names referring to contributing institutions, simulation indices, and extraction variants such as CESAR input/output. Use each record's metadata to establish its specific scope.

Contents

The local mirror manifest used to prepare this card contains 1,846 scenario records, totalling 7,425,408,288,264 bytes (approximately 7.43 TB, or 6.75 TiB) of simulation files. These figures describe the prepared inventory, not a guarantee of live upload completion. The size category above counts scenario files, rather than individual time steps.

Contributing institution / record rights holder	Scenario creators	LOCA records	SBO records	Total records
CIEMAT	Joan Fontanet Sáez	418	367	785
Karlsruhe Institute of Technology (KIT)	Anastasia Stakhanova; Fabrizio Gabrielli	236	229	465
Jožef Stefan Institute (JSI)	Jure Brence	389	0	389
ENEA	Marcello Savini	0	207	207
Total		1,043	803	1,846

LOCA means loss-of-coolant accident; SBO means station blackout. Types are taken from scenario names and descriptions in the source info.yaml files and recorded in the scenario_type column of scenarios.csv. Counts refer to separate upload records (UUIDs), including extraction variants and the three records awaiting RADAR review; they do not count unique accident trajectories. Each of the 1,846 records has exactly one of these two types in its name or description.

scenarios.csv maps every mirrored data path to its original scenario name, description, ASSAS UUID, DOI status, creators, rights holders, and SHA-256 checksum, with official scenario and head DOIs where recorded. Names, descriptions, and available step counts and durations come from the per-upload info.yaml files. DOI, authorship, and rights fields come from the saved RADAR metadata in the mirror manifest; the held-record list identifies exceptions. Empty step-count or duration fields mean that the source YAML did not provide the value.

DOI coverage: 1,843 records have individual RADAR DOIs in the saved metadata, each linked to the head collection. Three additional files in the mirror inventory are listed as held for review and have no recorded RADAR DOI. They are marked held_for_review_no_recorded_doi in scenarios.csv, with empty DOI and parent-DOI fields; they must not be represented as DOI-backed archival records until their RADAR metadata is available. The exceptions are:

• SBO_KIT_CESAR_in_ou_s189 (73f7e5bf-8561-4e56-82fe-028d6b57c537).
• SBO_KIT_CESAR_in_ou_s190 (8029d61e-f931-4536-b518-8d86752952b7).
• CIEMAT-SBO-early-PD_149 (1f6344c8-7af1-47fe-92f4-ef485adac4ee).

README.md
scenarios.csv
NETCDF_STRUCTURE.md
netcdf_structure.json
describe_netcdf.py
data/<ASSAS-upload-UUID>.nc
samples/LOCA_KIT_init_sim_s983.nc

File format

The mirror uploader copies the original dataset.h5 bytes into data/<UUID>.nc without conversion or numerical modification. The .nc suffix alone does not establish NetCDF compatibility for every file. The sample supplied here opens as NetCDF-4 with Python's netCDF4 library; HDF5 tools can also be used to inspect the underlying storage.

Files contain hierarchical groups and multidimensional variables, rather than a single flat table. The sample includes groups for the primary and secondary circuits and other model components, with variables such as pressure, temperature, mass, and sensor values. Inspect the actual groups, dimensions, variable attributes, units, and time information before combining scenarios. Group layouts and available variables may differ between extraction variants. No common time grid, complete trajectory, or train/validation/test split is implied.

NetCDF structure and tensor shapes

The complete structure reference lists every group and all 456 variables individually, including dimension declarations, exact tensor shapes, data types, units, and descriptions. netcdf_structure.json additionally preserves every group and variable attribute, including extraction strategies, original units, and metadata fields. Both references are exhaustive for the supplied sample; other files may use different dimensions or variables. Use describe_netcdf.py to generate the same inventory for any selected file without loading its numerical tensors.

The hierarchy below is verified against the supplied sample. Groups organise physical subsystems; individual variables are separate tensors, so there is no file-wide feature tensor or batch dimension. The root contains attributes and groups, with no root-level variables.

/
├── other/{global,private,cesar,metadata}
├── primary/{wall,volume,junction,pipe}
├── secondary/{wall,volume,junction}
├── vessel/{general,mesh,face}
├── connection/{general,fission}
├── containment/{connection,wall,dome,pool,zone}
└── dimensions

T denotes the number of stored time steps in a variable’s group (T = 9 in this sample). Shapes below describe each variable, in its stored axis order; the variable count is not an extra tensor axis. Numerical variables in this sample use float32. Spatial/component counts are exact sample sizes, not guaranteed constants across scenarios. Dimension names are reproduced verbatim, including abbreviated names such as wall_secondar.

Group	Dimension axes	Sample tensor shape	Variables with these axes	Example variable	Example unit	Units across these variables
/other/global	(T)	(9,)	2	vessel_rupture_time	s	s
/other/global	(T, sensor)	(9, 780)	1	sensor_values	1	1
/other/private	(T)	(9,)	14	freqsave	Hz	1, Hz, Pa, s
/other/cesar	(T)	(9,)	8	macro_ts_begin	s	1, s
/other/cesar	(T, cesar_output)	(9, 2233)	1	variable_prime	1	1
/primary/wall	(T, wall_primary)	(9, 92)	8	Fp_power_wall_primary	W	K, W
/primary/volume	(T, volume_primary)	(9, 85)	13	T_gas_primary_volume	K	1, K, kg/m³, kg, Pa
/primary/junction	(T, junction_primary)	(9, 93)	3	V_liq_junction_ther	m/s	kg/s, m/s
/secondary/wall	(T, wall_secondar)	(9, 125)	5	Fp_power_secondar_wall	W	K, W
/secondary/volume	(T, volume_secondar)	(9, 73)	12	T_gas_secondar_volume	K	1, K, kg/m³, kg, Pa
/secondary/junction	(T, junction_secondar)	(9, 76)	4	V_liq_secondar_junction_ther	m/s	kg/s, m/s
/vessel/general	(T, mesh)	(9, 76)	3	m_magma_vessel	kg	kg
/vessel/general	(T, component_clad)	(9, 36)	4	T_comp_clad	K	1, J, K, kg
/vessel/general	(T, component_fuel)	(9, 36)	4	T_comp_fuel	K	1, J, K, kg
/vessel/general	(T, component_bono)	(9, 3)	3	T_comp_bono	K	J, K, kg
/vessel/general	(T, component_crod)	(9, 84)	3	T_comp_crod	K	J, K, kg
/vessel/general	(T, component_clad_crod)	(9, 84)	3	T_comp_clad_crod	K	J, K, kg
/vessel/general	(T, component_tguide)	(9, 84)	3	T_comp_tguide	K	J, K, kg
/vessel/general	(T, component_tinst)	(9, 36)	3	T_comp_tinst	K	J, K, kg
/vessel/general	(T, component_grid)	(9, 30)	3	T_comp_grid	K	J, K, kg
/vessel/general	(T, component_LP_wall)	(9, 36)	3	T_comp_LP_wall	K	J, K, kg
/vessel/general	(T, component_LP_corium)	(9, 7)	3	T_comp_LP_corium	K	J, K, kg
/vessel/general	(T)	(9,)	4	FP_A_heat	Bq	Bq, 1, kg
/vessel/mesh	(T, mesh)	(9, 76)	15	P_vessel	Pa	1, K, kg/m³, kg/s, kg, Pa
/vessel/face	(T, face)	(9, 140)	3	Q_m_liq_face	kg/s	kg/s, m/s
/connection/general	(T, connection_flow)	(9, 2)	9	Mwater_cum_connecti_flow	kg	J, kg/s, kg
/connection/general	(T, connection_break)	(9, 38)	11	Mwater_cum_connecti_break	kg	J, K, kg/s, kg, Pa
/connection/general	(T, connection_feedwater)	(9, 29)	3	Mwater_cum_connecti_feedwater	kg	kg/s, kg, W
/connection/general	(T, connection_heat)	(9, 4)	3	T_connecti_heat	K	K, Pa, W
/connection/general	(T, connection_source)	(9, 5)	1	Qsteam_connecti_source	kg/s	kg/s
/connection/general	(T, connection_mcci)	(9, 1)	19	T_connecti_mcci	K	J, K, kg/s, Pa, T
/connection/fission	(T)	(9,)	132	Q_fp_AA_connecti_flow	kg/s	kg/s, kg
/connection/fission	(T, connection_break)	(9, 38)	76	Q_fp_AB_connecti_break	kg/s	kg/s
/containment/connection	(T, conn)	(9, 68)	2	containment_liq_vel_conn	m/s	m/s
/containment/wall	(T, wall_containm, wall_profile)	(9, 62, 20)	1	containment_wall_temp	K	K
/containment/dome	(T)	(9,)	44	m_AL_aerosol_dome	kg	kg
/containment/zone	(T, zone)	(9, 18)	9	average_total_pressure	Pa	1, K, Pa
/containment/zone	(T)	(9,)	1	water_level	m	m
/dimensions	(T)	(9,)	1	time_points	s	s

The sample contains 440 numerical variables and 16 character-array metadata variables (456 variables in total). Most physical quantities have shape (T, N), while global quantities have shape (T,). For example, primary/volume/P_primary_volume has shape (9, 85); containment/wall/containment_wall_temp has shape (9, 62, 20), with time, containment-wall, and wall-profile axes. Vessel component variables use distinct component dimensions rather than a single shared component axis.

/dimensions/time_points contains the stored time-coordinate values, with shape (T,). The time dimension is also declared within individual data groups. Inspect coordinate and unit attributes in the selected file when aligning time-dependent tensors; matching array lengths alone do not establish compatible coordinates.

Character metadata variables use NetCDF S1 arrays with axes (metadata_entry_count, string_length). In this sample all metadata entry counts are zero: shapes are (0, 4), (0, 7), or (0, 50). These empty label arrays do not imply that their associated numerical tensors are empty. primary/pipe contains only metadata arrays; containment/pool is empty in this sample.

For a model input, an individual (T, N) variable can be expanded to (T, N, 1); stacking F compatible variables along a new feature axis produces (T, N, F). Add a batch axis only after explicitly selecting or aligning scenarios. These are derived model-input shapes, not tensors stored in the files.

Use this recursive inspection to obtain exact shapes from any selected file without reading all tensor values:

from netCDF4 import Dataset

def show_structure(group):
    for name, variable in group.variables.items():
        path = f"{group.path.rstrip('/')}/{name}"
        print(path, variable.dimensions, variable.shape, variable.dtype)
    for child in group.groups.values():
        show_structure(child)

with Dataset(path) as ds:  # path to a downloaded .nc file
    show_structure(ds)

Units

Units are recorded per variable, using the source attribute unit (singular). The overview table gives the example variable’s unit and every unit found among variables with the same axes in that group. Shared tensor shapes do not imply shared units. The complete reference lists each variable’s unit symbol and original unit annotation.

Unit	Original annotation(s)	Variables in sample
Bq	[Bq]	1
1	[-]	22
Hz	[Hz]	1
J	[J]	18
K	[K]	33
kg/m³	[kg/m3]	5
kg/s	[kg/s]	160
kg	[kg]	138
m/s	[m/s]	8
m	[m]	1
not stored	not stored	16
Pa	[Pa]	21
s	[s]	19
s	not stored	1
T	[T]	5
W	[W]	7

Tables use standard unit symbols, with 1 for dimensionless. This changes display notation only; no numerical conversion is applied. The source labels second and seconds both appear as s here and remain separate rows in the catalogue to preserve their distinct original annotations. The 16 character metadata arrays have no stored unit; this is distinct from an explicit dimensionless unit. /dimensions/time_points uses seconds. When stacking features with different physical units, retain per-feature units and choose any normalisation explicitly.

Small NetCDF sample

samples/LOCA_KIT_init_sim_s983.nc is a byte-for-byte copy of one original stored simulation file, not synthetic data or a newly truncated trajectory. It is 1,825,372 bytes (about 1.83 MB) and contains 9 stored time steps in its primary-volume variables. This short record demonstrates file access; it should not be treated as representative of full accident duration or outcome.

• Original scenario name: LOCA_KIT_init_sim_s983.
• Creators: Anastasia Stakhanova and Fabrizio Gabrielli (KIT).

Download and read

Install huggingface_hub, netCDF4, and numpy, then download just the sample:

from huggingface_hub import hf_hub_download
from netCDF4 import Dataset

path = hf_hub_download(
    repo_id="Helmholtz-AI-Energy/ASSAS-PWR-1300",
    repo_type="dataset",
    filename="samples/LOCA_KIT_init_sim_s983.nc",
)
with Dataset(path) as ds:
    print(ds.data_model)  # NETCDF4 for this sample
    print(list(ds.groups))
    volume = ds.groups["primary"].groups["volume"]
    pressure = volume.variables["P_primary_volume"]
    print(pressure.dimensions, pressure.shape)
    print({key: pressure.getncattr(key) for key in pressure.ncattrs()})
    first_steps = pressure[:3, :]  # Read a small slice.

For local access, set path = "samples/LOCA_KIT_init_sim_s983.nc". To retrieve another scenario, find its path_in_repo in scenarios.csv and use that path as filename. Download selected files to avoid retrieving the entire multi-terabyte collection. These hierarchical scientific files are intended for direct scientific-library access.

Interpretation and limitations

Simulation results depend on the source model, initial and boundary conditions, numerical settings, and extracted variables. Scenario labels and descriptions are preserved from the upload metadata and do not constitute a uniform parameter specification. Missing metadata should not be inferred from a filename alone. A stored record can cover only part of a trajectory.

When constructing evaluation splits, keep related simulation variants and time slices together to reduce leakage between training and evaluation. Validate units, missing values, time coordinates, and physical consistency for the intended analysis. The project targets research and simulator development; these data alone do not establish validation for operational safety decisions.

License, attribution, and citation

The saved RADAR metadata specifies Creative Commons Attribution 4.0 International (CC BY 4.0) for all 1,846 records. Retain attribution to the creators and rights holders of the scenarios used, link the license, and indicate any modifications in derived work.

Cite the official RADAR records, including when downloading through this mirror. Cite the head DOI for the collection and the individual scenario DOI(s) for the records used. Mention the Hugging Face repository as the access location; it does not replace the official DOI citation.

Collection citation using the saved RADAR metadata:

Kieckhefen, Paul. ASSAS-PWR-1300. Karlsruhe Institute of Technology. Dataset. https://doi.org/10.35097/bp03ve8f24y4w17m

Use the current RADAR citation export for publication-year and version details. The saved head metadata identifies Paul Kieckhefen as collection creator and Anastasia Stakhanova, Fabrizio Gabrielli, Jure Brence, Joan Fontanet Sáez, and Marcello Savini as producers. Scenario-specific creators are listed in scenarios.csv.

Funding and sources

ASSAS received funding from the Euratom Research and Training Programme under grant agreement 101059682; see the European Commission project record.

This card was prepared from the ASSAS project website, the per-scenario info.yaml files, the mirror manifest, and saved RADAR head and scenario metadata associated with the official collection. The RADAR landing-page text supplied by the depositor confirms the head DOI, related scenario DOIs, contributors, rights holders, and license. It shows the head record as “In Bearbeitung” (in preparation) and reports no attached files, preventing archival/publication/review of that empty head package. Thus, the head DOI is the official RADAR identifier recorded for the collection, while publication and DOI resolution are not asserted here. The supplied page text and saved metadata are snapshots; live status was not independently checked. Collection counts describe the local manifest snapshot used on 19 October 2026.