Validation of simulated training sets using a convolutional neural network for isotope identification in urban...

Area of Science:

• Nuclear physics and instrumentation
• Artificial intelligence and machine learning
• Public safety and security

Background:

• Real-time isotope identification in urban settings is vital for law enforcement threat assessment.
• Challenges exist in creating representative training datasets for neural networks due to variable urban conditions.
• Existing methods struggle with the wide range of isotopes, activities, and shielding encountered.

Purpose of the Study:

• To validate a method for generating realistic gamma spectra datasets without radioactive sources.
• To train and test a convolutional neural network (CNN) for isotope identification using simulated and real-world data.
• To assess the feasibility of using generated spectra for accurate isotope detection in uncontrolled environments.

Main Methods:

• Generated synthetic gamma spectra by combining simulated isotope signatures with background radiation from the SIGMA dataset.
• Utilized k-means clustering to extract a labeled dataset of 12,748 spectra from the SIGMA dataset.
• Trained a CNN classifier on the combined simulated and real-world spectral data.

Main Results:

• The CNN model achieved high prediction accuracy, with the lowest class accuracy at 96% for simulated data and 89.8% for real SIGMA dataset spectra.
• The method successfully identified isotopes within real-world gamma spectra from the SIGMA dataset.
• Validation confirmed the effectiveness of the simulated data generation technique.

Conclusions:

• The validated method for generating gamma spectra is effective for training isotope identification models.
• This approach significantly aids in developing robust AI tools for real-time threat detection in urban environments.
• Future work can expand the isotopic library and model complexity for enhanced security applications.