keffdata/pwr-smr-2026-01-community

⚛️ KEFF Data | PWR-SMR-2026-01 Dataset (Community Edition)

Bridging academia and industry with high-fidelity synthetic simulation data for next-generation nuclear AI models.

1. Overview

The PWR-SMR-2026-01 dataset is a high-fidelity Monte Carlo neutron transport dataset generated for static reactor states of a generic Small Modular Reactor (SMR). It is designed strictly for research, model development, design-space exploration, benchmarking, and offline anomaly detection workflows.

This "Community Edition" contains 1,518 unique simulation cases representing varying operational states, core perturbations, and anomaly regimes.

Quick Specifications

Parameter	Specification
Simulation Engine	OpenMC (Continuous-energy Monte Carlo)
Cross-Section Library	ENDF/B-VIII.0
Total Simulation Cases	1,518 unique static reactor states
Histories per Case	12,500,000 neutron histories
Target Architecture	17x17 PWR Fuel Assembly SMR Lattice
Dataset Size	~90+ GB (Uncompressed HDF5 files)

---

2. Anomaly Classes & Data Distribution

To ensure robust training for multi-variable deep learning architectures, the 1,518 unique statepoints are distributed across three independent safety-critical operational anomalies and a nominal baseline[cite: 291, 297]:

Anomaly Class	Cases (N)	Perturbation Range	Primary Physics Mechanism
Control Rod (CR) Misalignment	555	0–200 cm insertion depth	Control rod worth: negative reactivity insertion
Fuel Temperature (Doppler)	481	950–1200 K	Doppler broadening of U-238 resonance capture
Coolant Void Fraction	481	0–35% void fraction	Under-moderated regime density feedback
Nominal Baseline	1	Reference state	Unperturbed core benchmark (keff = 1.26660$)

Verification & Validation (V&V) Baselines

Every simulation case in this community release has been rigorously audited against industrial safety standards to guarantee physical integrity before deployment[cite: 228, 241]: Source Convergence: Fundamental-mode spatial distribution verified with a Shannon entropy drift of 0.0419% (Passing industry threshold of < 0.1%)[cite: 66, 300]. Statistical Reliability: Adjusted for a Lag-1 autocorrelation coefficient of 0.62, raising the conservative industrial noise floor to ±56 pcm[cite: 85, 92, 300]. Code-to-Benchmark Accuracy: Verified against standard reference benchmarks for fresh PWR fuel pins with a stable code bias of -339.7 pcm, safely within the standard ±1000 pcm acceptance limit[cite: 51, 184, 300].

3. Dataset Structure

The dataset provides raw output directly from the OpenMC engine to preserve maximum physical fidelity. To protect proprietary methodologies, source XML generation files (metadata/) are excluded from this open-access tier.

For each simulation case, you will find:

• statepoint.250.h5: The complete state of the simulation at the final batch, containing k-effective eigenvalues, source site distributions, and runtime metadata.
• tallies.h5: Spatial and energy-dependent scoring matrices (flux, fission rates, heating) across the reactor geometry.
• manifest.json: Master cryptographic signature file (SHA-256) for verifying data integrity.

---

4. Anomaly Detection Regimes (Machine Learning Guidance)

This dataset operates in two distinct detectability regimes based on the separability index (Z), where the corrected statistical uncertainty floor is ±56 pcm.

$$Z=\frac{\mathrm{\mid }\mathrm{\Delta }\rho \mathrm{\mid }}{{\sigma }_{corrected}}Z\; =\; \backslash frac\{|\backslash Delta\backslash rho|\}\{\backslash sigma\_\{corrected\}\}$$

• Regime A: Macro-Anomalies (Z > 3.0)
  • Examples: Major control rod drops (≥50 cm), severe voiding (>20%).
  • Approach: Single-frame models (CNNs, MLPs) can detect these anomalies in individual snapshots due to a strong signal-to-noise ratio.
• Regime B: Micro-Perturbations (Z < 3.0)
  • Examples: Temperature drifts (+200 K), minor voiding (<20%), slight rod misalignments (<50 cm).
  • Approach: The signal is partially masked by the statistical noise floor. Requires temporal models (LSTMs, Transformers), trend-based analysis, or denoising autoencoders.

---

5. How to Use (Python Snippet)

Due to the large file sizes (90+ GB), we recommend using the huggingface_hub library to download specific files or the entire dataset locally.

from huggingface_hub import snapshot_download

# Download the entire dataset to your local machine
local_dir = snapshot_download(
    repo_id="keffdata/pwr-smr-2026-01-community",
    repo_type="dataset",
    local_dir="./keff_dataset"
)
print(f"Dataset successfully downloaded to {local_dir}")

6. Licensing & Commercial Use

Phase 1 Open-Access Framework To accelerate global nuclear AI research, KEFF Data provides this Community Edition dataset under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. It is entirely free for academic, educational, and non-commercial R&D workflows.

Enterprise & Commercial Licensing Any commercial entity, corporate R&D division, or national laboratory wishing to utilize this dataset, its underlying methodology, or its outputs for proprietary development, commercial product training, or operational benchmarking must obtain explicit written authorization and a commercial license.

For enterprise licensing structures and access to our production-tier datasets, contact: contact@keffdata.com

7. Citation

If you use this dataset in your academic research or non-commercial projects, please cite it using the official DOI:

DOI: 10.57967/hf/8914

@dataset{aslan2026keff,
  author       = {Caglayan Aslan},
  title        = {PWR-SMR-2026-01: High-Fidelity Monte Carlo Dataset for Nuclear Anomaly Detection},
  publisher    = {Hugging Face},
  year         = {2026},
  month        = {March},
  doi          = {10.57967/hf/8914},
  url          = {[https://doi.org/10.57967/hf/8914](https://doi.org/10.57967/hf/8914)},
  note         = {KEFF Data - Document ID: RED-PAPER-PH1-SMR}
}