논문 상세보기
Nuclear Fuel Cycle-Related R&D Activities Classification Using Deep Learning
This study aims to classify R&D activities related to the nuclear fuel cycle using the deep learning methodology. First, R&D data of the Republic of Korea were collected from the National Science & Technology Information Service (NTIS) for the years 2021, 2022, and 2023. We use keywords such as ‘nuclear,’ ‘uranium,’ ‘plutonium,’ and ‘thorium’ to find nuclear-related R&D projects in the NTIS database. Among the numerous R&D projects found through keyword searches, overlapping and medical-related R&D projects were excluded. Finally, 495 R&D projects conducted in 2021, 430 R&D projects conducted in 2022, and 296 R&D projects conducted in 2023 were obtained for analysis. After that, Safeguards experts determine whether the R&D projects are subject to declaration under the AP. The values of the content validity index (CVI) and content validity ratio (CVR) were used to verify whether the experts’ judgments were valid. The 1,218 collected and labeled data were then divided 8:2 into training and test datasets to see if deep learning could be applied to classify nuclear fuel cycle-related R&D activities. We use the Python and TensorFlow packages, including RNN, GRU, and CNN methods. First, the collected text information was preprocessed to remove punctuation marks and then tokenized to make it suitable for deep learning. After 20 epochs of training to classify the nuclear fuel cycle-related R&D activities, the RNN model achieved 97.30% accuracy and a 5.85% error rate on the validation dataset. The GRU model achieved 96.53% accuracy and a 9.06% error rate on the validation dataset. In comparison, the CNN model achieved 94.61% accuracy and a 2.57% error rate on the validation dataset. When applying the test dataset to each model, the RNN model had a test accuracy of 83.20%, the GRU test accuracy of 82.79%, and the CNN model had a test accuracy of 85.66% for the same dataset. This study applied deep learning models to labeled data judged by various experts, and the CNN model showed the best results. In the future, this study will continue to develop an optimum deep learning model that can classify nuclear fuel cycle-related R&D activities to achieve the purpose of safeguards measures from open-source data such as papers and articles.
