Citation:
P. O’Neal, “Using Machine Learning Techniques to Verify Mixed Plutonium Sample Attributes”, 63rd Annual Meeting of the Institute of Nuclear Materials Management (INMM), Virtual Conference, 25-27 July 2022.
Abstract:
An area of interest within the nuclear nonproliferation community is developing robust methods to identify information regarding the origin of an unknown sample of special nuclear material (SNM) by using its own physical characteristics. These methods can be of use in deterring the theft or smuggling of SNM, as well as in verifying declared activities taking place at facilities under nuclear safeguards monitoring. Work at Texas A&M University has focused on producing a nuclear forensics methodology that can attribute a plutonium sample’s reactor-type of origin, fuel burnup, and time since irradiation (TSI). This methodology should use a set of intra-elemental isotope ratios as the forensic signature for analysis. These efforts have yielded two preliminary methodologies, one that focused on using a maximum likelihood calculation to perform the attribution and another that utilized models trained using machine learning. Both methodologies lack the ability to attribute mixed or spoofed plutonium samples, where plutonium samples originating in two different reactors are combined. A modification to the machine learning methodology is underway which will add the capacity to identify spoofed plutonium samples. The current machine learning methodology utilizes a support vector machine (SVM) classifier to identify reactor-type, followed by a burnup quantification using gaussian process regression (GPR) models. Last, an analytic calculation is used to determine TSI using the predicted reactor-type and fuel burnup. The SVM classifier and GPR model are trained using data produced with MCNP burnup simulations. The updated version will feature a new classifier which has been trained using a data set that contain mixed plutonium data and will therefore be capable of identifying samples which have similar characteristics. It will be determined if an SVM classification method is still the optimal method for this type of discrimination. A sensitivity analysis will determine if there is a minimum composition threshold that needs to be reached for the classifier to identify a mixed reactor-type.