Speaker
Abstract
Laser-Induced Breakdown Spectroscopy (LIBS) can serve as a rapid detection method for quantifying uranium (U) concentration in molten salt during the electrorefining of spent nuclear fuel. However, its quantification accuracy is low due to limitations in data analysis methods, a problem that becomes particularly pronounced in the detection of high-concentration ions. Herein, this study focuses on solid LiCl-KCl-UO2Cl2 molten salt with a U content of 0-20 wt% as the research object. By integrating machine learning, the modeling approach of the LIBS quantitative analysis model is gradually optimized, with a comprehensive evaluation of the performance enhancement effects of data preprocessing algorithms, feature extraction algorithms, and modeling algorithms. The results indicate that the optimal modeling scheme is Channel Baseline Correction (CBC)-Channel Internal Standard (CIS)-Competitive Adaptive Reweighted Sampling (CARS) feature extraction-Partial Least Squares Regression (PLSR), which achieves a prediction set coefficient of determination (R2p) > 0.99 and enables the accurate quantification of high-concentration U in molten salt. The optimized modeling flow improved the prediction performance (evaluated by CV-RMSEp) by 82.32% relative to the baseline ULR model without preprocessing. The machine learning-enhanced LIBS modeling scheme proposed in this study is expected to overcome the technical challenges of precise quantitative analysis for high-concentration uranium and other ions in liquid molten salt systems.
摘要
激光诱导击穿光谱（LIBS）作为一种定量检测手段，可应用于乏燃料电解精炼过程中熔盐内铀（U）离子浓度的快速检测，但受限于数据分析方法的不足，其定量准确性偏低，在高浓度离子检测场景中该问题尤为突出。为此，本研究以固体LiCl-KCl-UO2Cl2熔盐中的U离子 (含量0-20 wt.%)为研究对象，提出了一种机器学习（ML）增强的LIBS建模流程，采用偏最小二乘回归（PLSR）算法建立定量模型，全面评估数据预处理算法、特征提取算法及建模算法对模型性能的提升效果。结果表明，采用通道基线校正（CBC）-通道内标（CIS）-竞争性自适应重加权采样（CARS）特征提取-偏最小二乘回归（PLSR）的建模流程为最佳建模方案，模型的预测集决定系数（R2p）> 0.99，实现了对熔盐中高浓度U的准确定量，与未进行预处理的传统ULR模型相比，预测性能提升了82.32%。本研究提出的机器学习增强型LIBS建模流程，有望攻克液态熔盐体系中高浓度铀及其他离子精准定量分析的技术难题.
| 关键词 | 激光诱导击穿光谱，高浓度铀，机器学习，熔盐，定量分析 |
|---|---|
| Keywords | LIBS, High concentration uranium, Machine learning, Molten salt, Quantitative analysis |
