Examination on Deep Learning Approach to Nuclear Proliferation Risk Modeling Philseo Kim a , Man-Sung Yim a a Department of Nuclear and Quantum Engineering, KAIST, Daejeon 34141, Republic of Korea * Corresponding author: msyim@kaist.ac.kr 1. Introduction Why do countries develop nuclear weapons while others do not/cannot? This question has been a major concern for scholars in the recent 20 years. Several studies have examined what factors cause nuclear proliferation. However, while scholars have focused on finding significant determinants of proliferation risk, they have not sought to enhance the ability to predict proliferation levels in a country. As Bell [1] pointed out, variables identified as significant determinants of nuclear proliferation have failed to offer a strong prediction ability. Therefore, this study has sought to examine the possibility of improving the ability to predict and classify proliferation levels of a country. We apply a deep learning algorithm, specifically Multilayer Perceptron (MLP) to improve the prediction efficiency in classifying a proliferation level of a country. This preliminary research will have implications for the study of proliferation and nuclear security. 2. Methods and Results In this section, we discuss some of the techniques used to develop neural networks and the results of the model. 2.1 Dataset The proliferation timeline of a country defined by Bleek [2] is used in this study. Bleek [2] updated the proliferation behavior of a country from prior works. Bleek [2] defined proliferation level as 4: No interest, Explore, Pursue, and Acquire. He presented that 31 countries have at least explored the nuclear program. Therefore, in this preliminary study, we only use 30 countries’ proliferation timelines from 1945-2000 (We did not include the West Germany case in this study). A number of studies have identified the determinants that could explain the cause of proliferation, such as domestic environment, economic capability, and nonproliferation norms. In this preliminary study, about 37 features identified as significant determinants from the previous studies were first selected to support developing a deep learning algorithm [3-7]. The specific model features are described in Table I. Table I: Model Features Category Variables Economic Capability GDP per capita GDP Industrial Capacity CINC Nuclear Fuel Cycle Capability Uranium ore production Uranium conversion capability Uranium enrichment capability Uranium oxide fuel fabrication capability Mixed oxide fuel fabrication capability Wet spent fuel storage facility Dry spent fuel storage facility Spent fuel reprocessing capability Zirconium alloy processing capability Zircaloy tube fabrication capability Heavy water production capability Power reactor capacity Number of fast reactors Number of heavy water research reactors Number of graphite research reactors Number of light water research reactors Number of other types of research reactors Latency Level Nuclear Assistance Sensitive Nuclear Assistance Civilian Nuclear Assistance Cumulative total number of civilian nuclear assistance Security/ Threat Environment Disputes Rivalry Allies Percentage of democratic countries Domestic Environment Polity score Trade ratio Liberalization IAEA membership Transactions of the Korean Nuclear Society Virtual Autumn Meeting October 21-22, 2021