Validation of a Machine Learning-Based Nuclear Forensics Methodology for the Discrimination of a Chemically Separated Plutonium Sample from Low-Enriched Uranium

• Published in: Nuclear science and engineering Vol. 198; no. 9; pp. 1817 - 1829
• Main Authors: O'Neal, Patrick J., Martinson, Sean P., Chirayath, Sunil S.
• Format: Journal Article
• Language: English
• Published: Taylor & Francis 01.09.2024
• Subjects: fuel burnup; machine learning; Plutonium forensics; reactor-type; time since irradiation
• ISSN: 0029-5639; 1943-748X
• DOI: 10.1080/00295639.2023.2271711

When the foundation of a method is simulated data, it is paramount that the method is validated with data from physical samples when possible. This study presents the results of validating a recently developed nuclear forensics methodology with a new low-burnup plutonium sample, chemically separated from low-enriched uranium irradiated in thermal neutron flux. The nuclear forensics methodology uses machine learning models to discriminate the reactor type of origin, fuel burnup, and time since irradiation (TSI) of chemically separated plutonium samples. The machine learning models use intra-elemental isotope ratios of cesium, samarium, europium, and plutonium as features; the isotopic ratio data for training the models were generated through fuel burnup simulations of various nuclear reactor types. The methodology predicted the reactor type and fuel burnup of the plutonium sample successfully. Initial difficulties to predict the TSI were resolved with the inclusion of a new intra-elemental isotope ratio of cerium.