Developing a machine learning-based verification method for subcritical nuclear tests with gamma spectroscopy measurements

Given rising tensions around nuclear testing, it is crucial to develop verification tools that can strengthen the nuclear test ban and the CTBT’s entry into force. Currently, there is no way to verify that states comply with the zero-yield standard, i.e., that they do not conduct supercritical nuclear tests at very low yields. In this work, we explore the potential of machine learning methods to distinguish between subcritical and supercritical tests using gamma spectra measurements one month after the test. Training and testing are based on thousands of spectral data generated with advanced computer simulations of very low-yield tests and gamma detectors. Preliminary results show that gradient-boosting methods can successfully distinguish between subcritical and supercritical tests with 97% accuracy. Other performance metrics also indicate a strong potential for this method to serve as a tool to detect violations of the zero-yield standard. Future work will assess the robustness of these performances against more variations in very low-yield tests and measurement parameters.