Machine Learning-Enhanced Laser-Induced Breakdown Spectroscopy for Quantitative Analysis of High-Concentration Uranium in Molten Salt

Laser-induced breakdown spectroscopy (LIBS) exhibits significant potential for the rapid analysis of uranium (U) in molten salts during the electrorefining of spent nuclear fuel. However, its quantitative accuracy, particularly for high-concentration U, is still constrained by inadequacies in data analysis methods. Herein, we proposed a machine-learning (ML)-enhanced LIBS modeling flow, encompassing steps of data preprocessing, feature extraction, and model training, which achieved accurate quantification of high-concentration U (up to 20 wt%) in solidified LiCl-KCl-UO 2 Cl 2 molten salt. The partial least squares regression (PLSR) algorithm was employed to build the quantitative model, resulting in better performance than the univariate linear regression (ULR) model. To further enhance predictive accuracy, the channel baseline correction (CBC) and channel internal standard (CIS) preprocessing methods were developed. Finally, two feature extraction algorithms, the competitive adaptive reweighted sampling (CARS) and uninformative variables elimination (UVE), were evaluated. The results demonstrated that the integration of CBC, CIS, CARS, and PLSR constituted the optimal modeling flow, achieving a low cross-validation root mean square error of prediction (CV-RMSE p ) of 0.3092 wt%. The optimal modeling flow improved the prediction performance (evaluated by CV-RMSE p ) by 57.42% relative to the best conventional model (internal standard combined with ULR) and by 82.32% relative to the ULR model without spectral preprocessing. The ML-enhanced LIBS modeling flow proposed in this work is expected to overcome the obstacle of precise quantification of high-concentration U and other ions in liquid molten salt systems.