Radioactive hot-spot localisation and identification using deep learning.

Area of Science:

• Nuclear physics and engineering
• Computer science and artificial intelligence
• Radiological security and threat detection

Background:

• Detecting radioactive hot-spots and identifying radionuclides is a significant challenge for security sectors, particularly with radiological, nuclear, and explosive threats.
• Existing methods face limitations in accurately localizing, quantifying, and identifying unknown radioactive sources in complex scenarios.

Purpose of the Study:

• To develop and validate a Machine Learning (ML) based solution using artificial neural networks (NNs) for the detection, localization, quantification, and identification of radioactive sources.
• To create specialized NN models (RHLnet and RHIdnet) capable of processing radiological intensity counts and gamma spectra for accurate source characterization.

Main Methods:

• Development of RHLnet model for estimating the number, location, and activity of radioactive sources based on intensity counts and localization data.
• Development of RHIdnet model for identifying radionuclides by analyzing the gamma spectrum of detected sources.
• Training NN models using a comprehensive dataset of simulated data to ensure fast and accurate predictions.

Main Results:

• The proposed NN-based approach demonstrated fast and accurate predictions in both simulated and real-world scenarios with multiple sources.
• The system showed a low incidence of false negatives and false positives, indicating high reliability.
• The algorithm proved scalable for analyzing both very large regions and small, confined scenarios.

Conclusions:

• Artificial neural networks (NNs) are a powerful emerging tool in the nuclear field, offering novel solutions for radiological security.
• The developed RHLnet and RHIdnet models provide a reliable and scalable method for safeguarding human lives by improving radioactive source detection and identification.
• Further exploration of single and multiple isotope identification highlights the potential for continued advancements in nuclear security techniques.