Rapid nuclear forensics analysis via machine-learning-enabled laser-induced breakdown spectroscopy (LIBS)
Rapid nuclear forensics analysis via machine-learning-enabled laser-induced breakdown spectroscopy (LIBS)
复制标题
通过机器学习激光诱导击穿光谱 (LIBS) 进行快速核取证分析
DOI：
--
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
A. Dehayem
中科院分区：
文献类型：
--
作者：
K. H. Angeyo;B. Bhatt;A. Dehayem
核取证是一种分析方法，涉及对截获的核材料和放射性材料进行分析，以确定其核归属。目前核因子面临的关键挑战是缺乏合适的微量分析方法来直接、快速、微创地检测和量化核因子信号。激光诱导击穿光谱(LIBS)有可能在机器学习(ML)技术的帮助下克服这些限制。在这篇文章中，我们报告了支持ML的LIBS方法的发展，用于支持核安全的快速核因子分析和归属。385.464 nm、385.957 nm和386.592 nm的铀原子谱线被鉴定为铀的核磁共振信号，用于快速定性检测隐藏在有机结合剂和含铀矿物中的痕量铀。用LIBS测定铀的检出限为34 ppm。利用人工神经网络(ANN)和光谱特征选择技术，建立了纤维素和含铀矿物中痕量铀的多元定标策略：(1)铀线(348 nm~455 nm)；(2)铀线(380 nm~388 nm)；(3)隐蔽的铀峰(UV范围)。利用类别2的模型能够预测48 ppm的铀，相对误差预测(REP)为10%。利用细微铀峰，即3类铀峰的校正模型，可以预测由标准物质IAEA-RGU-1制备的球团中的铀，相对标准偏差为6%。这证明了人工神经网络对噪声LIBS光谱进行建模以进行痕量定量分析的能力。我们开发的定标模型预测含铀矿物中的铀浓度在54-677 ppm的范围内。利用肯尼亚不同地区采集的含铀样品的特征选择，对LIBS光谱(200-980 nm)进行主成分分析(PCA)，将其归类为与其地理来源相关的组。主成分分析表明，这些样品的分组主要是由于稀土元素，即Ce、Dy、Pr、Pm、Nd和Sm。因此，具有ML功能的LIBS在野外核素分析和隐蔽条件下NRM中铀的归属方面具有实用价值。
Nuclear forensics (NF) is an analytical methodology that involves analysis of intercepted nuclear and radiological materials (NRM) so as to establish their nuclear attribution. The critical challenge in NF currently is the lack of suitable microanalytical methodologies for direct, rapid, minimally invasive detection and quantification of NF signatures. Laser-induced breakdown spectroscopy (LIBS) has the potential to overcome these limitations with the aid of machine-learning (ML) techniques. In this paper, we report the development of ML-enabled LIBS methodology for rapid NF analysis and attribution in support of nuclear security. The atomic uranium lines at 385.464 nm, 385.957 nm, and 386.592 nm were identified as NF signatures of uranium for rapid qualitative detection of trace uranium concealed in organic binders and uranium-bearing mineral ores. The limit of detection of uranium using LIBS was determined to be 34 ppm. A multivariate calibration strategy for the quantification of trace uranium in cellulose and uranium-bearing mineral ores was developed using an artificial neural network (ANN, a feed forward back-propagation algorithm) and spectral feature selection: (1) uranium lines (348 nm to 455 nm), (2) uranium lines (380 nm to 388 nm), and (3) subtle uranium peaks (UV range). The model utilizing category 2 was able to predict the 48 ppm of uranium with a relative error prediction (REP) of 10%. The calibration model utilizing subtle uranium peaks, that is, category 3, could predict uranium in the pellets prepared from certified reference material (CRM) IAEA-RGU-1, with an REP of 6%. This demonstrates the power of ANN to model noisy LIBS spectra for trace quantitative analysis. The calibration model we developed predicted uranium concentrations in the uranium-bearing mineral ores in the range of 54–677 ppm. Principal component analysis (PCA) was performed on the LIBS spectra (200–980 nm) utilizing feature selection of the uranium-bearing samples collected from different regions of Kenya clustered into groups related to their geographic origins. The PCA loading spectrum revealed that the groupings of these samples were mainly due to rare earth elements, namely, cerium, dysprosium, praseodymium, promethium, neodymium, and samarium. ML-enabled LIBS therefore has utility in field NF analysis and attribution of uranium in NRM under concealed conditions.