Female is a minority in STEM (Science, Technology, Engineering, Math) fields. It is encouraging and grateful that there is a solidarity among young female scientists, such as the young woman scientist camp (2020 YWS Camp.). And, thanks for the award at this program.

Research Content

For nuclear fuel cycle facilities, it is indispensable to implement a safeguards system that can be monitored and verified by the International Atomic Energy Agency (IAEA). Current safeguards technologies mainly rely on a mass balance method. Due to unique features of pyroprocessing, there is uncertainty to apply traditional safeguards methods to meet safeguard-ability requirements in commercializing pyroprocessing, and new approaches and technologies have been suggested to enhance existing safeguards. Process monitoring (PM) is one candidate. By employing PM, it is possible to indirectly track a flow of special nuclear materials. Various types of signals can be produced in (near) real time from a variety of sensors installed in specific locations at the facility. Though the signal alone do not imply information about the operation state, the process can be diagnosed based on the statistical experience by collecting data and developing data library. On this wise, machine learning approach can be applied.

In this study, the feasibility of using PM based on machine learning for improving safeguards-ability of pyroprocessing was examined. Cathode potential in electrorefining was selected as a target signal to investigate the operation state (normal vs. off-normal). Cathode potential data was produced through a series of experiments in lab-scale. Surrogate materials, which have close standard reduction potentials leading to feasible environment for codeposition, were employed in the experiments. The composition of deposition was analyzed using inductively coupled plasma atomic emission spectroscopy (ICP-OES) to classify input data. Before learning, data preprocessing was conducted in terms of complicity and quantity to prepare appropriate data for learning. Both neural network (NN) and recurrent neural network (RNN) were used in learning. To optimize the learning, various variables were tested. As a results, two-layer RNN with optimizer AMSGrad showed the best result, presenting more than 80% of classification accuracy. To advance current result, further research is planned. Considering special characteristics of safeguards, which discourages the necessary of off-normal operation to acquire off-normal data, a classification model using only normal data (machine learning without negative data) will be developed. If safeguards-ability can be improved with this method, this method would be applied not only to electrorefining, but also to overall facility including other unit processes.