Machine learning for atmospheric radioxenon source discrimination based on isotopic activity concentrations

Monitoring the activities of radioxenon isotopes (^131mXe, ¹³³Xe, ^133mXe, ¹³⁵Xe) in the atmosphere is a key task of the International Monitoring System (IMS) under the Comprehensive Nuclear-Test-Ban Treaty (CTBT). Discriminating the source of radioxenon detected at monitoring stations is crucial and highly challenging. This paper investigates the performance of machine learning for radioxenon source discrimination. Based on simulated and measured data, a synthetic dataset of radioxenon isotopic activity concentrations was generated, comprising samples from nuclear explosion, reactor, and hybrid sources. This dataset was then used as training data for machine learning models. The performance of the trained models was evaluated on a series of randomly generated test datasets. Under an idealized low-MDC scenario, the radioxenon source discrimination accuracy reaches 88.6%, and the probability of correctly identifying nuclear explosion signals is as high as 94.2%, corresponding to a false negative rate of 5.8% and a false positive rate of 14.9%. A sensitivity analysis using more realistic MDC settings shows that the discrimination performance decreases substantially as the MDC increases.