Neutron Shielding Surrogate Model

Overview

This project develops a neural network surrogate model for neutron transport in multi-layer shielding configurations. The surrogate replaces expensive MCNP6 Monte Carlo simulations with fast neural network predictions, enabling rapid design iterations and parameter optimization.

Project Structure

.
├── 1_generate_mcnp_inputs.py      # Generate MCNP input decks
├── 2_run_mcnp_batch.py            # Run MCNP simulations in parallel
├── 3_train_surrogate_model.py     # Train neural network
├── 4_inference_and_benchmark.py   # Benchmark and demonstrate usage
├── 5_analysis_and_visualization.py # Comprehensive analysis
├── requirements.txt               # Python dependencies
├── manifest.csv                   # Parameter samples (generated)
├── training_data.csv              # MCNP results (generated)
├── best_model.pt                  # Trained model (generated)
├── scaler.pkl                     # Feature scaler (generated)
├── model_metadata.json            # Model configuration (generated)
└── plots/                         # Visualization outputs (generated)

Requirements

Software

• Python 3.8+
• MCNP6 (with appropriate licenses)
• CUDA-capable GPU (optional, for faster training)

Python Packages

numpy>=1.21.0
pandas>=1.3.0
scipy>=1.7.0
torch>=1.9.0
scikit-learn>=0.24.0
matplotlib>=3.4.0
seaborn>=0.11.0

Install with:

pip install -r requirements.txt

Quick Start

Step 1: Generate Training Data

# Generate MCNP input files (creates 2000 cases)
python 1_generate_mcnp_inputs.py

# Run MCNP simulations (modify MCNP path if needed)
# This will take several hours depending on your system
python 2_run_mcnp_batch.py

Note: Modify mcnp_exe path in 2_run_mcnp_batch.py to point to your MCNP6 installation.

Step 2: Train Surrogate Model

# Train neural network on MCNP results
python 3_train_surrogate_model.py

This will:

• Load and preprocess MCNP results
• Train a feedforward neural network
• Save the best model and scaler
• Generate training curves

Step 3: Benchmark and Use Model

# Benchmark accuracy and speed
python 4_inference_and_benchmark.py

# Run comprehensive analysis
python 5_analysis_and_visualization.py

Usage Examples

Making Predictions

from inference_and_benchmark import SurrogatePredictor

# Load trained model
predictor = SurrogatePredictor()

# Define shielding configuration
params = {
    'concrete_thick': 50.0,    # cm
    'lead_thick': 5.0,         # cm
    'poly_thick': 10.0,        # cm
    'source_energy': 2.0,      # MeV
    'source_intensity': 1e9    # particles/s (Caution: for 1E4 to 1E5 partcle histories use PC but for any larger particle histories will need to run on HPC)
}

# Predict flux
flux = predictor.predict(params)
print(f"Predicted flux: {flux:.2e}")

Parameter Optimization

import numpy as np

# Find minimum concrete thickness for target flux
target_flux = 1e-6
concrete_range = np.linspace(10, 100, 100)

for thickness in concrete_range:
    params['concrete_thick'] = thickness
    flux = predictor.predict(params)
    if flux < target_flux:
        print(f"Optimal thickness: {thickness:.1f} cm")
        break

Model Performance

Accuracy

• R² Score: ~0.95-0.99 (depending on training data quality)
• Mean Relative Error: 1-5%
• Predictions within MCNP uncertainty: >80% of cases

Speed

• Prediction time: ~0.1-1 ms per case
• MCNP time: ~60 seconds per case (1M particles)
• Speedup factor: ~10^5x

Customization

Adjusting Network Architecture

Edit 3_train_surrogate_model.py:

model = ShieldingSurrogate(
    input_dim=5,
    hidden_dims=[256, 128, 64, 32],  # Modify layer sizes
    dropout=0.3                       # Adjust dropout rate
)

Changing Parameter Ranges

Edit 1_generate_mcnp_inputs.py:

self.param_bounds = {
    'concrete_thick': (10, 150),   # Adjust ranges
    'lead_thick': (0, 30),
    # ... etc
}

Adding More Materials

Modify the materials dictionary in MCNPInputGenerator class.

Limitations and Considerations

1. Training Data Coverage: Model is only reliable within the parameter space used for training. Extrapolation beyond training ranges may be inaccurate.

2. Geometry Constraints: Current implementation is limited to 1D slab geometry. 2D/3D geometries require different input representations (e.g., CNNs for grid-based, GNNs for assembly-based).

3. Physics Fidelity: Surrogate captures average behavior but may miss edge cases or unusual physics regimes. Always validate critical designs with full MCNP.

4. Energy Dependence: Current model predicts total (energy-integrated) flux. For energy-dependent predictions, modify network to output multiple energy groups.

5. MCNP Statistical Uncertainty: Training data contains statistical noise from Monte Carlo. Using variance reduction in MCNP can improve surrogate accuracy.

Extension Ideas

Short-term (1-2 weeks)

• Multi-energy group predictions
• Angular flux distributions
• Uncertainty quantification using ensemble methods
• Active learning to sample parameter space efficiently

Medium-term (1-2 months)

• 2D cylindrical geometries
• Time-dependent problems (pulsed sources)
• Inverse problems (source/material identification)
• Physics-informed neural networks (PINNs)

Long-term (3+ months)

• 3D arbitrary geometries using graph neural networks
• Coupled neutron-photon transport
• Criticality eigenvalue problems
• Integration with CAD-to-MCNP workflows

Troubleshooting

MCNP Runs Failing

• Check MCNP installation and PATH
• Verify material definitions (atomic fractions should sum to 1.0)
• Increase NPS if statistical errors are too large

Training Not Converging

• Increase training data (generate more MCNP cases)
• Adjust learning rate or network architecture
• Check for NaN values in training data
• Use gradient clipping if loss explodes

Poor Predictions

• Verify parameter is within training range
• Check MCNP statistical uncertainties in training data
• Try log-transforming features with large dynamic ranges
• Collect more training data in poorly-performing regions

Citation

If you use this code in your research, please cite:

@software{neutron_shielding_surrogate,
  title = {Neural Network Surrogate for Neutron Shielding Calculations},
  author = {Kavish Imam},
  year = {2026},
  url = {https://github.com/yourusername/neutron-surrogate}
}

License

MIT License - see LICENSE file for details

Contact

For questions or issues:

• Open an issue on GitHub
• Email: kavishimam@gmail.com