Scientists say they have developed a new artificial intelligence system capable of predicting uranium enrichment levels with extraordinary precision while identifying different types of nuclear material using gamma spectroscopy data.

The findings, based on tests of three machine-learning models, XGBoost, CatBoost, and LightGBM, could offer the UN nuclear watchdog valuable new tools to strengthen its safeguards under the Nuclear Non-Proliferation Treaty (NPT), according to researchers at the University of Sharjah.

The study, published in the journal Annals of Nuclear Energy, found that all three models performed perfectly in identifying uranium categories, recording 100% accuracy, precision, recall, and F1-score.

The research introduces a dual-task machine learning framework that simultaneously predicts uranium-235 (235U) enrichment levels and classifies nuclear material types using gamma spectroscopy data.

The three gradient boosting models were trained on structured datasets derived from IAEA benchmark simulations of high-purity germanium (HPGe) detector outputs.

“The regression component achieved state-of-the-art performance, with XGBoost yielding a mean squared error (MSE) of 0.0019, a mean absolute error (MAE) of 0.0281, and a coefficient of determination (R²) of 1.0000,” the authors write. “LightGBM and CatBoost followed with R² values of 0.99999 and 0.9971, respectively.”

Among the three models, XGBoost delivered the strongest performance in predicting enrichment levels, achieving an almost negligible prediction error and a perfect R² score of 1.0000.

The authors report that even LightGBM attained nearly identical accuracy while offering faster computation times. As for CatBoost, they note that it demonstrated excellent performance and proved particularly robust with smaller datasets, maintaining stable and consistent results.

In fact, the researchers maintain that all three classification models achieved perfect scores, recording 100% accuracy, precision, recall, and F1-score across the highly enriched uranium (HEU), low-enriched uranium (LEU), and natural uranium (NU) categories.

They point out that their system would prove critical to improving nuclear safeguards by providing faster and more accurate analysis of nuclear materials, potentially enhancing the ability of UN nuclear inspectors and regulators to monitor compliance with international non-proliferation agreements.

“These results confirm the robustness and generalizability of gradient-boosted decision tree algorithms in the highly sensitive domain of nuclear safeguards,” the authors explain. “The proposed framework demonstrates the potential of machine learning as a decision-support tool for selected safeguards activities involving gamma spectroscopy.”

Advancing UN nuclear watchdog’s safeguards

The authors’ methodology addresses two critical analytical tasks: (1) regression to predict the percentage of uranium-235 (235U) enrichment, and (2) classification to identify nuclear material as natural uranium (NU), low-enriched uranium (LEU), or highly enriched uranium (HEU).

“These objectives are particularly relevant to those IAEA verification activities that involve non-destructive assay (NDA), where inspectors may need to simultaneously determine quantitative enrichment levels and qualitative material classifications to support safeguards verification and anomaly detection,” the authors note.

To simulate real-world safeguards conditions, the study draws on structured gamma spectroscopy datasets representing measurements obtained using high-purity germanium (HPGe) detectors, the industry standard for isotopic assay.

The proposed framework employs three state-of-the-art gradient boosting algorithms, XGBoost, CatBoost, and LightGBM, each selected for its robustness on structured data and suitability for small, high-dimensional nuclear datasets.

“These models were trained independently for regression and classification tasks using engineered and isotope-derived features and evaluated across metrics including mean squared error (MSE), mean absolute error (MAE), coefficient of determination (R²), accuracy, precision, recall, and F1-score,” the authors explain.

XGBoost leads the field

Although additional validation under operational conditions would be required before deployment within IAEA safeguards, the authors’ analysis demonstrates excellent predictive performance on benchmark datasets.

The results identify XGBoost as the most accurate and generalizable model. Its fine-grained tree construction, combined with robust regularization and carefully optimized hyperparameters, appears to have enabled it to capture complex nonlinear relationships within the spectral and isotopic feature space.

In safeguards applications involving gamma spectroscopy, such predictive accuracy could provide valuable decision support for enrichment estimation, subject to further validation under real-world inspection conditions.

CatBoost also delivered strong performance, although it exhibited slightly greater residual variance. The model achieved an MSE of 2.7757, MAE of 0.7580, and a coefficient of determination (R²) of 0.9971. The higher error, particularly at lower enrichment levels, may reflect the model’s sensitivity to data sparsity and the characteristics of its ordered boosting approach when trained on smaller datasets.

Nevertheless, CatBoost attained perfect classification accuracy (100%), achieving precision, recall, and F1-scores of 1.00 across all material categories. Its symmetric tree architecture and stable gradient estimation enabled the model to distinguish effectively between enrichment classes despite imbalances in class distribution.

LightGBM combined exceptional computational efficiency with high predictive accuracy. In the regression task, it achieved an MSE of 0.0119, MAE of 0.0389, and an R² score of 0.99999, closely approaching the performance of XGBoost.

Achieving such performance while requiring substantially less training time and memory underscores LightGBM’s potential for use in low-latency safeguards applications. The classification model likewise achieved 100% accuracy, with balanced precision, recall, and F1-score across HEU, LEU, and NU categories.

From promise to practice

While further validation under real-world nuclear inspection conditions remains necessary, the scope of the study is limited to the evaluation of machine-learning algorithms using benchmark gamma spectroscopy datasets representative of uranium gamma spectroscopy measurements.

As a result, the findings have yet to be assessed under the operational realities of IAEA safeguards, where verification activities are governed by safeguards agreements, facility-specific arrangements, inspection objectives, measurement uncertainty requirements, and a range of legal and operational considerations.

Consequently, the results should be interpreted as evidence of the technical potential of AI-assisted spectral analysis rather than as proof of enhanced safeguards effectiveness.

The authors emphasize that future research should test these methods on representative operational datasets and in collaboration with safeguards stakeholders to better evaluate their suitability within existing verification frameworks.