High Bias Machine Learning For Sensitivity Limits Of Antineutrino-Based Safeguards

Antineutrino detectors have potential for use as independent, tamper-proof tools for reactor monitoring using the antineutrino flux from reactor cores. Also, since the quantity and energies of these antineutrinos are dependent on the reactor inventory, the isotopic composition of the fuel can be verified based on any inconsistencies in the antineutrino spectrum. The count rate needed for reasonable verification, however, relies heavily on the absolute efficiency of the antineutrino detector as well as the strength of the statistical inference. In this work, we test the statistical limitations of a simulated Advanced Fast Reactor-100 (AFR-100) antineutrino-based safeguards system to detect special nuclear material being removed from the reactor core. After simulating 12 different diversion scenarios in which 1 significant quantity of plutonium (1 SQ is defined as 8 kg by the International Atomic Energy Agency (IAEA)) was pulled from various reactor core locations, expected antineutrino spectra along with the corresponding variances were created for 3-month collection periods. The diversion scenarios vary by three different factors, including the assembly location, diverted fuel replacement, and reactor power manipulation. Based on sampling from the diversion scenario antineutrino variances, a series of support vector machine models were applied to evaluate how well each of these models could interpret ‘unseen’ events, which are diversion scenarios withheld from the model training set but completely fill the model test set. Regardless of the model, the diversion predicting power, or safeguards power, was inaccurate on a case-by-case basis. However, by generating a large enough sample size, the models converged to a robust, high bias, lower limit safeguards power that would have previously been considered zero using statistical analysis methods. While these converged values were still much lower than the 0.2 safeguards power limit established by the IAEA for low-probability events, they do indicate the importance of simulations to identify which diversion scenarios are most at risk of proliferation.