radiacode/train/vega_ml

ML for Isotope Identification

A machine learning system for identifying radioactive isotopes from gamma-ray spectra captured by Radiacode scintillation detectors.

Project Status

✅ Completed: Synthetic gamma spectra generation system
✅ Completed: Vega ML model architecture (CNN-FCNN hybrid)
✅ Completed: Training pipeline with GPU support
✅ Completed: Inference engine
✅ Completed: Realistic CsI(Tl) background model
✅ Completed: Hybrid training (measured + synthetic background)
✅ Completed: Web dashboard (FastAPI + Chart.js)
🔲 Next: Retrain model with realistic background
🔲 Future: Real-time inference on Radiacode devices

---

Overview

This project builds a neural network that identifies radioactive isotopes from gamma spectra. Since collecting real spectra requires radioactive sources and is expensive/regulated, we generate synthetic training data based on realistic physics models.

Target Hardware

• Training: NVIDIA RTX 5060 Ti GPU (Blackwell, requires PyTorch 2.7+ with CUDA 12.8)
• Inference: Radiacode 101, 102, 103, 103G, 110 scintillation detectors

Data Format

• Input: 1D spectrum (1023 energy channels, 20-3000 keV, normalized to max)
• Output: Multi-label isotope classification (82 isotopes) with activity estimation (Bq)

---

Quick Start

Installation

# Create virtual environment
python -m venv .venv
source .venv/bin/activate  # Linux/Mac

# Install dependencies
pip install numpy scipy pillow

# Install PyTorch (nightly for RTX 5090/Blackwell support)
pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu128

Generate Synthetic Data

# Generate 10 test samples (default)
python -m vega_ml.synthetic_spectra.generate_spectra --num_samples 10 --output_dir data/synthetic

# With measured background for hybrid training (recommended)
python -m vega_ml.synthetic_spectra.generate_spectra \
  --num_samples 50000 \
  --output_dir data/synthetic \
  --measured_background /path/to/background_24h.npy

Train the Model

# Quick test run (5 epochs, small dataset)
python -m vega_ml.training.vega.run_training --test

# Full training
python -m vega_ml.training.vega.run_training \
  --data-dir data/synthetic \
  --model-dir models \
  --epochs 100 --batch-size 64

Run Inference

# Run inference on synthetic data
python -m vega_ml.inference.run_inference --model models/vega_best.pt --data data/synthetic

---

Vega Model Architecture

Vega is a CNN-FCNN hybrid model optimized for gamma spectrum isotope identification, based on research showing 99%+ accuracy on similar tasks.

Architecture Details

Component	Configuration
Input	1023 energy channels
CNN Backbone	3 ConvBlocks [64, 128, 256 channels]
Kernel Size	7 (captures spectral features)
FC Layers	[512, 256] with dropout
Output Heads	Dual: Classification (82 isotopes) + Regression (activity)
Total Parameters	34.5M
Activation	LeakyReLU + BatchNorm

Training Features

• Mixed Precision (AMP): Faster training on modern GPUs
• Multi-task Learning: Simultaneous isotope ID + activity estimation
• Loss Function: BCE (classification) + Huber (regression)
• LR Scheduling: ReduceLROnPlateau with early stopping

---

Synthetic Spectra Generation

Realistic Background Model

The background continuum uses a realistic CsI(Tl) shape calibrated against real Radiacode 103 measurements, not a simple exponential:

• Asymmetric hump at ~110 keV (sigma_left=55 keV, sigma_right=50 keV) — the dominant low-energy scatter peak characteristic of CsI(Tl) detectors
• Compton tail: 0.45exp(-E/240) + 0.04exp(-E/700) — realistic high-energy falloff
• Noise floor at 0.8% of peak — prevents zero-count channels

This replaces the previous simple exponential A*exp(-0.002*E) which failed to reproduce the characteristic CsI(Tl) response.

Hybrid Training with Measured Background

When a measured background file (background_24h.npy) is available, the generator blends it with the synthetic model:

• 70% measured background shape (scaled to target CPS)
• 30% synthetic continuum (for robustness against measurement artifacts)
• Stochastic isotope peaks (K-40, radon, thorium) are still added on top with random activity levels

This is controlled by the --measured_background CLI argument or the MEASURED_BACKGROUND_PATH environment variable.

Features

• 82 isotopes with accurate gamma emission lines
• Realistic physics: Gaussian peaks, Poisson noise, Compton continuum, CsI(Tl) background shape
• Multiple detector models: Radiacode 101, 102, 103, 103G, 110 with correct FWHM and energy ranges
• Configurable variation: Activity levels, measurement durations, isotope combinations
• Decay chains: Uranium-238, Thorium-232 chains with secular equilibrium

Sample Distribution (v3)

Type	Proportion	Description
Background only	15%	Environmental background only
Single calibration	20%	One check source + background
Single medical	8%	Medical isotope + background
Single industrial	5%	Industrial source + background
Uranium chain	10%	U-238 + daughters in equilibrium
Thorium chain	10%	Th-232 + daughters in equilibrium
NORM	7%	Naturally occurring radioactive material
Fallout	5%	Cs-137 + Cs-134 signature
Mixed	10%	Random 2-3 isotope mixes
Complex mix	5%	4-6 isotopes from various categories
Weak source	5%	Near-detection-limit sources

---

Project Structure

train/vega_ml/
├── README.md                    # This file
├── agents.md                    # AI agent context documentation
├── .gitignore                   # Git ignore rules
│
├── synthetic_spectra/           # Spectrum generation package
│   ├── __init__.py
│   ├── config.py                # Detector configurations (Radiacode 101-110)
│   ├── generator.py             # Main generation logic (SpectrumConfig)
│   ├── generate_spectra.py      # CLI batch generation (v1)
│   ├── generate_spectra_v3.py   # CLI batch generation (v3, parallel)
│   ├── ground_truth/
│   │   ├── isotope_data.py      # 82 isotopes database
│   │   └── decay_chains.py      # Decay chain definitions
│   └── physics/
│       └── spectrum_physics.py  # Physics calculations + realistic CsI(Tl) background
│
├── training/                    # Training infrastructure
│   └── vega/                    # Vega model package
│       ├── __init__.py
│       ├── isotope_index.py     # Isotope ↔ index mapping
│       ├── model.py             # VegaModel architecture + VegaLoss
│       ├── dataset.py           # PyTorch Dataset/DataLoader
│       ├── train.py             # Training loop & utilities
│       └── run_training.py      # CLI training script
│
├── inference/                   # Inference engine
│   ├── vega_inference.py        # VegaInference class
│   └── run_inference.py         # CLI inference script
│
├── models/                      # Saved model checkpoints
│   ├── vega_best.pt             # Best validation loss
│   ├── vega_final.pt            # Final epoch
│   └── vega_history.json        # Training metrics
│
└── data/                        # Generated data (git-ignored)
    └── synthetic/
        └── spectra/

---

Technical Details

Detector Specifications

Model	Crystal	FWHM @ 662 keV	Energy Range	Channels
Radiacode 101	CsI(Tl)	9.0%	20-3000 keV	1024
Radiacode 102	CsI(Tl)	9.5%	20-3000 keV	1024
Radiacode 103	CsI(Tl)	8.4%	20-3000 keV	1024
Radiacode 103G	GAGG(Ce)	7.4%	20-3000 keV	1024
Radiacode 110	CsI(Tl)	8.4%	20-3000 keV	1024

Note: Only the first 1023 channels are used (channel 1023 is an overflow bin).

Physics Model

• Peak shape: Gaussian with FWHM scaling as sqrt(E/662) for scintillators
• Expected counts: lambda = A * t * I * epsilon * T
• Noise: Poisson counting statistics
• Background: Realistic CsI(Tl) continuum (asymmetric hump + Compton tail) + environmental isotope peaks (K-40, radon daughters, thorium daughters)
• Hybrid mode: Measured background can be blended with synthetic (70/30 ratio) for maximum realism

Isotope Categories

• Natural background (K-40, Ra-226, Rn-222)
• Decay chains (U-238, Th-232, U-235)
• Calibration sources (Am-241, Cs-137, Co-60, Ba-133, Eu-152)
• Medical isotopes (Tc-99m, F-18, I-131, Ga-68)
• Industrial sources (Ir-192, Se-75)
• Reactor fallout (Cs-134, Cs-137, Sr-90)

---

Development

Dependencies

numpy>=1.24.0
scipy>=1.10.0
pillow>=9.0.0
scikit-learn>=1.3.0
torch>=2.0.0

GPU Support

For Blackwell GPUs (RTX 50-series, sm_120), use PyTorch 2.7+ with CUDA 12.8:

pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu128

For AI Agents

See agents.md for comprehensive documentation on system architecture, physics model details, and configuration options.

---

TODO

• Push to repository - Initial commit with generation system
• Create PyTorch DataLoader for training
• Implement CNN-FCNN model architecture (Vega)
• Create training script with logging
• Implement inference module
• Realistic CsI(Tl) background model
• Hybrid training with measured background
• Retrain model with realistic background
• Add model evaluation metrics & confusion matrix
• Implement real-time inference on Radiacode devices

---

License

[TBD]

---

Acknowledgments

• Radiacode for device specifications
• IAEA Nuclear Data Services for isotope data
• NNDC at Brookhaven National Laboratory
• Wang et al. research on CNN-FCNN for gamma spectroscopy