A comparison of machine learning methods to classify radioactive elements using prompt-gamma-ray neutron activation...

Area of Science:

• Nuclear security
• Analytical chemistry
• Machine learning

Background:

• Illicit radiological material detection is vital for nuclear security.
• Neutron-capture prompt-gamma activation analysis (PGAA) detects various radioactive materials.
• Current PGAA methods face challenges with long detection times and high false positive rates in nuclear forensics.

Purpose of the Study:

• Develop and evaluate machine learning algorithms for classifying radioactive elements from PGAA spectra.
• Compare the effectiveness of different algorithms, particularly for imbalanced datasets.
• Identify the optimal classifier for PGAA spectral data analysis.

Main Methods:

• Six machine learning algorithms were developed for classification.
• Performance was assessed using metrics like Precision, Recall, F1-score, Specificity, Confusion matrix, ROC-AUC, and Geometric Mean Score (GMS).
• Analysis focused on algorithms suitable for imbalanced datasets.

Main Results:

• Tree-based algorithms (Decision Trees, Random Forest, AdaBoost) outperformed Support Vector Machine and K-Nearest Neighbours.
• AdaBoost showed the highest recall and minimal false negatives for the minority class.
• The study identified AdaBoost as the preferred classifier for PGAA spectral data.

Conclusions:

• Machine learning, specifically AdaBoost, significantly enhances the classification of radioactive elements using PGAA data.
• This approach addresses limitations of traditional PGAA methods, improving accuracy and efficiency in nuclear security.
• AdaBoost's performance in handling imbalanced datasets makes it ideal for real-world nuclear forensics applications.