Linear Regression and Machine Learning for Nuclear Forensics of Spent Fuel from Six Types of Nuclear Reactors

The illicit trafficking of radioactive materials, especially weapon-grade uranium or plutonium, is a significant security threat. Nuclear forensics helps trace the illicit trafficking of radioactive materials. The present study develops the methods for the forensics of the possible origins of fuels irradiated in nuclear reactors, which are the most powerful sources producing radioactive materials, including plutonium. Three key factors are significant for irradiated fuel forensics, namely, initial ${}^{235}\mathrm{U}$ enrichment, burnup, and the type of irradiation nuclear reactors. The methods for the first two are determined based on experimental data of six nuclear-reactor technologies and are further verified using the neutron-transport-depletion coupling simulation of the two major commercial reactor technologies, a pressurized-water reactor (PWR) and a boiling-water reactor (BWR). In addition, three machine-learning techniques are applied to discriminate between a PWR and a BWR, which are quite similar in neutronic properties, with nice accuracy and generalization ability. In summary, the presently determined methods provide a reliable pathway to predict the origins of spent nuclear fuels.

Figure 1

Selected ratios of R₁= ${}^{235}\mathrm{U}{/}^{238}\mathrm{U}$ and R₂= ${}^{236}\mathrm{U}{/}^{238}\mathrm{U}$ versus real values of fuel enrichment for both PWR and BWR samples.

Figure 2

Schematics of the MLP used in the present work.

Figure 3

Illustration of the tenfold cross-validation method.

Figure 4

Trained model-based predictions versus real values of fuel enrichment for both a PWR and a BWR (upper), and AGR, RBMK, VVER-440, and VVER-1000 (lower) samples collected in the SFCOMPO-2.0 database.

Figure 5

Trained model-based predictions versus real values of burnup of fuel-irradiated PWR and BWR samples (upper) as well as AGR, RBMK, VVER-440, and VVER-1000 samples (lower) collected in the SFCOMPO-2.0 database.

Figure 6

MLP model-based prediction of a PWR and a BWR.

Figure 7

Confusion matrices of the LGR (left), SVM (middle), and MLP (right) models for the classification of a PWR and a BWR.

Figure 8

ROC curves of the MLP model. Dashed blue line (i.e., chance) indicates the ROC line of the random classification.

Figure 9

Model-based predictions versus real values of fuel enrichment for both a PWR and a BWR using simulation data.

Figure 10

Model-based predictions versus real values of burnup for both a PWR and a BWR using simulation data. SFCOMPO-2.0 data are included for comparison.