P. O’Neal S.S. Chirayath, “Quantifying the Effect of Measurement Uncertainty on Separated Plutonium Attribution Methodology”, INMM/ESARDA Joint Annual Meeting, 22-26 May 2023, Vienna, Austria.
Ongoing work at Texas A&M University has produced a nuclear forensics methodology that can attribute a separated plutonium sample’s reactor of origin, fuel burnup, and time since irradiation (TSI). The parameter attribution is performed using two models trained with machine learning, a classification model for the reactor-type and a regression model for fuel burnup. The TSI is calculated analytically using the predicted reactor-type and fuel burnup. Sets of intra-element isotope ratios are the features used in both models. The training data used in producing the models is sampled from a library of MCNP fuel burnup simulations that have been performed for a set of reactors of interest. For the validation of the model performance multiple irradiation campaigns were conducted to produce physical samples that could be attributed. These campaigns involved irradiating uranium samples of various initial enrichment levels at the High Flux Isotope Reactor (HFIR) and Missouri University Research Reactor (MURR). Subsequently the the plutonium produced was separated and the isotopic concentrations determined. The use of simulated data for the model production, and then physical data in the use of the model introduces an unavoidable amount of incongruency, there will always be differences between these two. Additionally, both sets of data introduce their own sources of uncertainty, and characterizing this uncertainty is important for judging the ultimate capabilities of this attribution methodology. A study is being performed to use the validation data’s measurement uncertainties to study the effect that variance in the input can have on model predictions. To do this, a set of test data will be produced for each validation plutonium sample by sampling each isotope ratio from a normal distribution with the measured mean and variance for that isotope ratio, predictions will be made with this data set to find how the predictions change with the natural variation in the measured isotope ratios values. The prediction distribution can be compared against a similar distribution produced by a legacy forensics methodology that utilizes a maximum likelihood statistical method to attribute rather than machine learning models.