Researchers from the University of Sharjah (UAE) have presented an artificial intelligence system that, using gamma spectroscopy data - the analysis of a substance's radiation - simultaneously predicts the degree of uranium enrichment and identifies the type of nuclear material. The work was published in the journal Annals of Nuclear Energy.
The development is based on three machine learning models: XGBoost, CatBoost and LightGBM. All are built on gradient boosting: the algorithm combines many simple decision trees and successively improves them to increase accuracy. The models were trained on structured datasets from IAEA reference simulations for high-purity germanium detectors - standard equipment for isotope analysis.
The system solved two tasks at once: it predicted the percentage of uranium-235 enrichment (regression - forecasting a numerical value) and assigned the material to one of the categories - natural uranium, low-enriched or highly enriched.
In classification, all three models showed perfect results: precision, recall and the F1 score (a metric that accounts for both precision and recall) reached the maximum for all categories. In predicting the enrichment level, XGBoost proved best: its errors were minimal, and the coefficient of determination R2, which shows the share of explained data variance, reached its upper bound. For LightGBM and CatBoost, the R2 values were also very close to the maximum.
The authors note: LightGBM, at almost the same accuracy, works faster and requires less memory, while CatBoost is more robust on small datasets. The models were trained separately for each task using specially engineered features, including isotopic characteristics.
According to the researchers, such a system could help IAEA inspectors analyze nuclear materials faster and more accurately to verify compliance with the non-proliferation treaty. However, it is emphasized: the work was performed on reference simulations, not on real measurements.
Before introduction into nuclear safeguards practice, additional verification is needed in field conditions, where safeguards agreements, measurement accuracy requirements and legal norms apply. The authors themselves call the results evidence of the technical potential of AI spectrum analysis, not proof of improved safeguards effectiveness.