Machine learning-assisted laser-induced breakdown spectroscopy for monitoring molten salt compositions of small modular reactor fuel under varying laser focus positions

Next-generation advanced nuclear reactors based on molten salts are interested to apply machine learning (ML) technology to minimize human error and realize effective autonomous operation. Owing to harsh environments with limited access to molten salts, laser-induced breakdown spectroscopy (LIBS) has been investigated as a possible option for remote online monitoring. However, the height of molten salts is easily fluctuated by vibration. In addition, the level of molten salts could change during normal operation through the insertion of a controlling structure. While these uncertainties should be considered, their effects have not been studied yet. In this study, LIBS has been actively coupled with ML to automate the online monitoring of difficult-to-access molten salt systems. To practically apply a prediction model with ML, we intentionally defocus the measurement by manipulating the sample position. This study investigates the focusing and defocusing spectra of Sr and Mo as fission products for constructing the two prediction models using partial least squares and artificial neural network methods. For each method, the prediction models trained with focusing spectra only or focusing and defocusing spectra simultaneously are constructed and compared to each other. While the prediction model using only focusing spectra resulted in a root mean square error of prediction (RMSEP) of 0.1943-0.2175 wt%, a prediction model using both spectra led to approximately 10 times enhanced RMSEP (0.0210-0.0316 wt%). This study implies that not only focusing data but also defocusing data are needed to construct the prediction model while considering its practical usage in a real system, especially in the complex processes of the nuclear industry.