Forensic Science International 251 (2015) 61–68

Contents lists available at ScienceDirect

Forensic Science International journal homepage: www.elsevier.com/locate/forsciint

Raman spectroscopy of uranium compounds and the use of multivariate analysis for visualization and classification Doris Mer Lin Ho a,b,*, Andrew E. Jones c,**, John Y. Goulermas c, Philip Turner d, Zsolt Varga a, Lorenzo Fongaro a, Thomas Fangha¨nel a,e, Klaus Mayer a a

European Commission, Joint Research Centre (JRC), Institute for Transuranium Elements (ITU), Postfach 2340, 76125 Karlsruhe, Germany DSO National Laboratories, 20 Science Park Drive, Singapore 118230 The University of Liverpool, Department of Electrical Engineering and Electronics, Brownlow Hill, Liverpool L69 3GJ, United Kingdom d AWE, Aldermaston, Reading, Berkshire RG7 4PR, United Kingdom e Ruprecht-Karls-Universita¨t Heidelberg, Physikalisch-Chemisches Institut, Im Neuenheimer Feld, Germany b c

A R T I C L E I N F O

Article history: Received 3 September 2014 Received in revised form 3 March 2015 Accepted 5 March 2015 Available online 14 March 2015 Keywords: Nuclear forensics Raman spectroscopy Uranium ore concentrates Multi-variate analysis

A B S T R A C T

Raman spectroscopy was used on 95 samples comprising mainly of uranium ore concentrates as well as some UF4 and UO2 samples, in order to classify uranium compounds for nuclear forensic purposes, for the first time. This technique was selected as it is non-destructive and rapid. The spectra obtained from 9 different classes of chemical compounds were subjected to multivariate data analysis such as principal component analysis (PCA), partial least square-discriminant analysis (PLS-DA) and Fisher Discriminant Analysis (FDA). These classes were ammonium diuranate (ADU), sodium diuranate (SDU), ammonium uranyl carbonate (AUC), uranyl hydroxide (UH), UO2, UO3, UO4, U3O8 and UF4. Unsupervised PCA of full spectra shows fairly good distinction among the classes with some overlaps observed with ADU and UH. These overlaps are also reflected in the poorer specificities determined by PLS-DA. Higher values of sensitivities and specificities of remaining compounds were obtained. Supervised FDA based on reduced dataset of only 40 variables shows similar results to that of PCA but with closer clustering of ADU, UH, SDU, AUC. As a rapid and non-destructive technique, Raman spectroscopy is useful and complements existing techniques in multi-faceted nuclear forensics. ß 2015 The Authors. Published by Elsevier Ireland Ltd. This is an open access article under the CC BY-NCND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction Nuclear forensics or nuclear forensic science is a discipline that has emerged from the need to tackle illicit trafficking of nuclear material, with a view to provide hints on the history of the material [1–3]. Nuclear forensics is not limited to the analysis of special nuclear material, but also involves radioactive materials since there are several applications for them. For instance, radioisotopes are widely used in medicine (treatment, diagnostic or sterilization of medical supplies) or industry (gauging or radiography). Upon the confiscate of an unknown or lost radioactive material, the immediate tasks would be to understand the material, that is, ‘What is it?’ Additionally, questions such as ‘how and for what

* Corresponding author at: DSO National Laboratories, 20 Science Park Drive, Singapore 118230. Tel.: +65 6871 2884. ** Corresponding author. Tel.: +44 151 794 4564. E-mail addresses: [email protected] (D.M.L. Ho), [email protected] (A.E. Jones).

purpose was it produced?’, ‘Where was the material produced?’ and possibly ‘How it got to the position where it was found?’ are more questions demanding a right answer. The multi-disciplinary field tackles the questions by studying characteristic parameters of the nuclear material. These measurable parameters also referred as signatures, relate to the physical, chemical and isotopic characteristics of the material [4]. There is neither single methodology that could be applied to any found material, nor any single measurable parameter that would suffice in exclusive identification of the origin of the material. In this sense, the analysis of a combination of several characteristic parameters would elevate the confidence in nuclear forensic conclusions. This is the reason why, research is always carried out to look into prospective signatures of various nuclear materials (uranium and plutonium) at all stages of the nuclear fuel cycle. In particular, this paper studies a group of compounds known as uranium ore concentrates (UOCs) and other uranium compounds such as UF4 and UO2. UOCs are products of uranium mining and milling processes at the early stages of the uranium fuel cycle. Uranium is leached from its ore, purified and subsequently

http://dx.doi.org/10.1016/j.forsciint.2015.03.002 0379-0738/ß 2015 The Authors. Published by Elsevier Ireland Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-ncnd/4.0/).

62

D.M.L. Ho et al. / Forensic Science International 251 (2015) 61–68

concentrated. The concentrated uranium solution is then precipitated with a reagent, which can be ammonia/ammonium hydroxide, hydrogen peroxide, sodium hydroxide, ammonium carbonate or milk of magnesia. Products of ammonium diuranate, uranyl peroxide, sodium diuranate, ammonium uranyl carbonate and uranyl hydroxide are thus formed. Further drying and calcinations lead to the formation of UO3 and U3O8. Although the colours of these powders range from yellow to orange, brown, green and black, they are collectively and colloquially known as yellow cakes [5]. These UOCs are further subjected to conversion processes where UF4 and UO2 (termed as non-UOC in this paper) are produced [6]. Various measurable parameters associated with yellow cakes have been identified as useful nuclear forensic signatures, such as analysis of uranium (major isotopes) [7,8], thorium [9], lead [10–12], strontium [11], sulphur [13] and neodymium [14] (minor isotopes), trace elements including rare-earth elements [7,10,12,15–17], non-volatile organics [18] and anionic impurities [17,19,20]. Spectroscopic techniques such as infrared [20], near infrared reflectance [21,22] and laser-induced breakdown spectroscopy have also been used to measure yellow cakes [23]. More recently, we have demonstrated the feasibility of using Raman spectroscopy as a tool for nuclear forensics purposes in the analysis of yellow cakes [24]. In fact, nuclear forensics has been applied to real cases of found or stolen UOCs [25–27] and therefore, the analytical scenario is real for this class of material, often traded in large quantities. Although Raman spectroscopy has been applied for the analysis of UOCs [28–33], the measurement of larger number of samples and different sample composition is necessary in evaluating its robustness as a signature for nuclear forensic purposes. To add further demonstration of Raman as a suitable nuclear forensic technique, it is applied in this study to 95 UOC and non-UOC samples of varying compositions with the goal of using different multivariate analysis for evaluation. 2. Experimental 2.1. Investigated samples 2.1.1. Industrial yellow cakes There were a total of 89 industrial UOCs investigated in this study. These samples of UOCs represent a large number of the world’s uranium mines. All the samples are in powder form and were pressed into pellet using hydraulic press 3630 X-Press (SPEX Industries Inc., USA) to ease handling and to avoid possible contamination. A pressure of approximately 74 MPa was applied to the samples. 2.1.2. Laboratory synthesized yellow cakes Six additional yellow cakes were prepared in our laboratory. The compounds were synthesized under controlled conditions by precipitation from chemically pure uranium nitrate solutions with various reagents. The details can be found elsewhere [24]. The synthesized compounds are sodium diuranate, ammonium diuranate, uranyl peroxide, uranyl hydroxide and two samples of ammonium uranyl carbonate. Five samples were dried at 105 8C while one of the two ammonium uranyl carbonate samples was dried at room temperature. In total, the given/assumed composition (from the industry) or known composition (from laboratory) of all the 95 samples are distributed as ammonium diuranate – ADU (35), ammonium uranyl carbonate – AUC (4), sodium diuranate – SDU (6), U3O8 (13), uranium hydroxide – UH (20), UF4 (3), UO2 (2), UO3 (4) and uranium peroxide – UO4 (8). In addition, the assumed compositions were verified or modified based on our infrared data published recently [20].

2.2. Instrumentation 2.2.1. Raman spectrometer Senterra bench-top model of Raman spectrometer from Bruker1 (Germany) was used for the measurements of the ore concentrates. Measurement with silicon was done daily prior to measurements to ensure that the sensitivity of the instrument does not change significantly on the different days of measurement. Calibration of wavenumbers is done automatically by the instrument with neon. Two different frequency lasers, 785 and 532 nm are available with Senterra. The former permits largely the measurements of all samples, while the latter is not favourable as the higher energy causes much fluorescence and sample degradation due to the heat induced from the laser. Laser power of 10 mW or 25 mW is mostly used on the samples. For calcined samples (darkly coloured samples), 10 mW has to be used as the use of 25 mW will lead to the instantaneous formation of oxide, thus changing the real composition of the sample. Integration times are typically 10 s and are increased for samples that yield weaker signal. The selected spectral range is 1560–90 cm 1 as no peaks are observed outside this region. 2.2.2. Infrared spectrometer Fourier Transform Infrared (FT-IR) measurements were performed using Perkin Elmer System 2000 spectrometer (from Perkin Elmer Ltd., Beaconsfield, UK) with the spectra range of 400–4000 cm 1 and 2 cm 1 resolution. The full details can be found elsewhere [20]. 2.3. Multivariate data evaluation A PLS Toolbox version 7.5.2 (Eigenvectors Research, Inc., USA) for Matlab version 8.1 (The Matworks Inc., Natick, MA, USA) software was used for multivariate data analysis; in particular principal component analysis (PCA), partial least square discriminant analysis (PLS-DA) and Fisher Discriminant Analysis (FDA) were applied on Raman spectra. 2.3.1. Principal component analysis PCA is a central part in multivariate explorative data analysis. High dimensional and collinear datasets are reduced in dimensionality to reveal latent and relevant information in data [33]. In this study, an exploratory analysis based on PCA, was applied on Raman spectra and, as internal cross validation, the random subsets cross validation method was applied using a data split of 9 and 20 as number of interactions. Before applying the algorithm, the spectral range included in the evaluation was reduced from 1560–90 cm 1 to 1150– 120 cm 1 to exclude regions where there are no peaks, arising from molecular vibrations. Pre-processing includes baseline correction, normalization and smoothing (Savitzky–Golay). The spectra obtained by Raman spectroscopy give information about the molecular composition and impurities, therefore the sample compositions were selected as labels. 2.3.2. Partial least square discriminant analysis A sample classification was performed by PLS-DA. The PLS-DA is considered a supervised class-modelling method in which prior knowledge of the classes of the samples test is required. The PLS-DA is based on Partial Least Square Regression (PLSR). The X matrix is the PCA scores while the Y matrix is constructed on columns of binary numbers where 1 represents the sample member of a class and 0 if it is not a member. The sample membership of a class is then modelled and predicted as a 0 or 1 within a threshold limit of usually 0.5. Diagnostic plots of different

D.M.L. Ho et al. / Forensic Science International 251 (2015) 61–68

63

thresholds will reveal the stability of the classifications [33]. A fundamental step to build a PLS-DA model is the determination of the correct number of latent variable. This choice is commonly performed by using cross validation where the samples are separated into the calibration/validation set; the model is built with the calibration set and validated with the other. A typical cross validation procedure involves more than one sub-validation test, each of which involves the selection of different subsets of samples for model building and model testing. There are different cross validation methods and they vary with respect to how different samples subsets are selected for these sub-validation experiments. In this work, the PLS-DA model was validated by cross-validation applying the random subsets cross validation method using a data split of 9 and 20 as number of interactions [34]. The goodness of the model was evaluated, both in calibration and cross-validation, through the examination of sensitivity, specificity, the classification error and the RMSEC/CV (root mean square error in calibration and cross-validation) parameters [35–37]. 2.3.3. Fisher Discriminant Analysis Additionally, supervised dimensional reduction has been performed on the dataset for visualization of Raman spectra using FDA. Unlike unsupervised dimensional reduction techniques such as PCA, supervised methods take advantage of the class labels (here we use the different UOC compositions) to help create an embedding of the data. Using the class labels enables these techniques to perform a transformation that not only reduces the dimensions for the data but also describes the underlying discrimination task. This is done by maximizing the distance between samples of different classes and minimizing the distance between samples of the same class. Before supervised dimensional reduction, some additional preprocessing has been carried out. Discrete Cosine Transformation (DCT) has been used to compress the 2941 variables of the original dataset. DCT is very good for the signal processing compression as the majority of the signal information can be contained in a few components of the DCT [38]. In this case, 40 features provide a good separation of the samples without over-fitting to the data. As shown later (Fig. 5), this provides a good separation of the sample groups. This has been implemented using the DCT function in the Matlab signal processing toolbox. Supervised dimensional reduction has been implemented using FDA [39,40]. For this implementation, the cosine metric was used to determine the distances between samples. 3. Results and discussions Fig. 1 shows the Raman spectra taken from the various uranium compounds. The detailed interpretation of these Raman spectra is discussed elsewhere [24]. In brief, a typical Raman spectrum using Senterra could be divided into three regions of vibrations relating to uranium (900–90 cm 1), vibrations due to anionic impurities if present (1200–900 cm 1) and non-Raman electronic transitions if present (1560–1200 cm 1). 3.1. PCA PCA was used as exploratory analysis to investigate the major trends in the whole collected spectra and to figure out the factors that gave the possibility to discriminate among the UOCs samples. Fig. 2 depicts the results from PCA of 95 Raman spectra that correspond to 95 different samples represented by each point on the plot. Three principal components (PCs) were selected, which explained 74.50% of the total variance of the sample and it is important to note also that all the samples fall within the 95%

Fig. 1. Raman spectra (baseline corrected) of different industrial UOCs and nonUOCs measured with 785 nm.

confidence limit. The 2D and 3D score plots are shown in Fig. 2A and B, respectively. They reveal the relationship among the samples; in general, compounds having the same composition could be grouped together, as indicated by the full circles in Fig. 2B, but there are also exceptions. These are small clusters of U3O8, UF4, AUC, UO4 and possibly SDU. For the samples found close to each other, it meant that their respective Raman spectra are similar. On the other hand, two major clusters of ADU and UH (dotted circles in Fig. 2B) can also be observed although there are some overlaps between these two clusters. With the aid of PCA for visualization, one can observe the usefulness and shortfalls of using Raman spectroscopy for measuring different uranium compounds. The loading plot as seen in Fig. 2C shows the wavenumbers that explain the position of the samples in the score plot. Although there are a few points which are located farther away from their local clusters, their results can be readily explained by examining the spectra. In a recent paper, we have noted the absence of nitrogen– hydrogen (N–H) and oxygen–hydrogen (O–H) bands using Raman spectroscopy at the particular laser wavelength of 785 nm [41]. These N–H and O–H bands are known to be Raman active [42] and their observation would likely result in a larger differentiation between the ADU and UH clusters. In the case of ADU, both N–H and O–H bands should be observed and for UH, only O–H band(s) should be observed. Absorption bands related to these vibrations were observed for infrared spectroscopy [20]. As iterated from Fig. 2B, ADU and UH formed overlapping clusters and this suggests that some ADU or UH are not distinguishable from each other. Fig. 3 illustrates the infrared and Raman data for three selected pairs of ADU and UH for further discussion. Fig. 3A and B belong to samples El Dorado (ADU) and South Alligator (UH), Fig. 3C and D belong to that of Delft (ADU) and North Span (UH) and Fig. 3E and F are laboratory synthesized ADU and UH. It should be noted that the N–H band in El Dorado (blue line of Fig. 3A) is much more prominent than its O–H band unlike Delft which is completely opposite (blue line of Fig. 3C). In a qualitative aspect, one should expect the ratio of N–H/O–H bands to be high for a pure ADU (blue line of Fig. 3E). On the other hand, the Raman spectra of El Dorado/South Alligator (Fig. 3B) and Delft/North Span (Fig. 3D) are highly similar to each other unlike their infrared spectra. In contrast, the

64

D.M.L. Ho et al. / Forensic Science International 251 (2015) 61–68

Fig. 2. 2D score plot (A), 3D score plot (B) and loading plot (C) from PCA of 95 Raman spectra.

corresponding Raman spectra of the laboratory synthesized ADU and UH are quite different. These results suggest that sample like Delft (for example) probably does not possess a single UOC composition but rather and possibly, a mixture of UOCs. This postulation or assumption is realistic for two reasons. Firstly, the compositions of the UOCs were provided by external parties to the best of their knowledge. Secondly, the composition of UOC is probable a less critical factor compared to its purity. Therefore, the various reagents used in the milling process can in fact precipitate uranium solution (even in minor amount) before the precipitation step itself. It is often the case that uranium solution has to be brought to optimum pH before precipitation with NH3, NaOH, lime or MgO before its subsequent precipitation [6]. The use of lime in the first step is also used as a mean to precipitate impurities [43]. These steps could have produced varying compositions of UOCs. Although laboratory synthesized UH (red line of Fig. 3E) did not appear to be very pure (due to the presence of starting

material such as MgO and nitrate), it was precipitated without the pre-adjustment of pH and the same was applied to the precipitation of ADU. To our point of view, the latter was more important than the former to ensure that only one single composition exists. In the case of some ADU samples found in the UH clusters and some UH found in the ADU samples, their infrared spectra were compared to verify its composition. 3.2. PLS-DA After a preliminary step of feature extraction to compress the relevant information using PCA, PLS-DA model was developed to predict a class membership for the samples; nine classes were then modelled. The model was evaluated using sensitivity, specificity, both in calibration and cross-validation, the classification error and the root mean square error in calibration and cross validation

D.M.L. Ho et al. / Forensic Science International 251 (2015) 61–68

65

Fig. 3. IR and Raman of El Dorado and South Alligator samples (A and B); IR and Raman of Delft and North Span samples (C and D); IR and Raman of laboratory synthesized ADU and UO2(OH)2 samples (E and F). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

D.M.L. Ho et al. / Forensic Science International 251 (2015) 61–68

66

Table 1 Sensitivity, specificity, classification error, R2 and RMSE of PLS-DA model. Modelled

ADU

AUC

SDU

U3 O8

UF4

UH

UO2

UO3

UO4

Sensitivity (Cal) Specificity (Cal) Sensitivity (CV) Specificity (CV) Class. Err. (Cal) Class. Err. (CV) R2 (Cal) R2 (CV) RMSEC RMSECV

0.83 0.90 0.77 0.85 0.14 0.19 0.6 0.4 0.30 0.31

1.00 1.00 1.00 0.99 0.00 0.01 0.9 0.8 0.06 0.09

1.00 0.97 0.83 0.97 0.02 0.11 0.7 0.6 0.12 0.15

1.00 0.99 1.00 0.99 0.01 0.01 0.8 0.8 0.13 0.15

1.00 1.00 0.60 0.99 0.00 0.02 0.7 0.2 0.10 0.15

0.85 0.92 0.75 0.87 0.12 0.19 0.6 0.3 0.25 0.35

1.00 0.99 0.63 0.98 0.01 0.19 0.7 0.1 0.07 0.13

1.00 1.00 1.00 1.00 0.00 0.00 1.0 0.9 0.03 0.05

1.00 1.00 0.88 0.97 0.00 0.07 0.9 0.9 0.07 0.09

[44], as presented in Table 1. The value of R2, also known as coefficient of determination (a measure of the quality of the model) was also reported. The best classification model was obtained using 14 latent variables. The data clearly shows that the better classified samples were AUC, U3O8, UO3 and UO4 that were characterized by the highest values, close to 1, of sensitivity and specificity both in calibration and cross validation and the lower values of classification errors and RMSEC/ RMSECV. Also, the R2 values indicated that the predictive ability of the model, for these samples, can be considered robust. With regards to ADU and UH, they were the samples that showed the highest values of the classification errors and the RMSEC/ RMSECV. Also in this case, these samples appear overlapped in the PLS-DA score plot reported in Fig. 4 and showed in fact the same behaviour obtained by PCA. In the PLS-DA score plot, the clusters are better distinguished than in the PCA clusters reported in Fig. 2A. These results confirm that the classification model could be considered, as preliminary study, as a good model and the combination of Raman spectra and multivariate data analysis (PCA and PLS-DA), showed a great potential in nuclear forensics investigation.

3.3. Supervised visualization of Raman spectra using FDA Fig. 5 shows the results from using FDA for visualization of the spectra as described in Section 2.3.3. As with the PCA results shown in Fig. 2, this is a dimensional reduction technique. However, FDA is a supervised dimensional reduction method and this means that the embedding shown in Fig. 5 has grouped samples belonging to the same sample group and separated samples from different group. Supervised visualization would not be possible during forensic analysis of an unknown sample, as the class of the sample would not be known in advance. However, Fig. 4 is useful for understanding the relationship between the different sample groups and how they relate to one another. As shown by the PLS-DA results in Section 3.2, the classification results are not 100% accurate. The FDA results help to understand which of the classes are more likely to be mis-classified with one another. A number of the sample groups have been well separated from the other composition. Notably, the UO3, UF4, UO4, ADU, U3O8 and UO2 materials have been well distinguished within the FDA embedding. This suggests that based on the Raman spectra done in this study, these particular compositions can successfully be identified and

Fig. 4. PLS-DA score plot (LV1–LV2) of Raman Spectra.

D.M.L. Ho et al. / Forensic Science International 251 (2015) 61–68

67

ADUs, Raman spectroscopy is useful as a tool for compound identification. As a rapid and non-destructive technique, Raman spectroscopy certainly complements other techniques in the multidisciplinary field of nuclear forensics. Acknowledgements The authors wish to acknowledge the UK Support Programme to the IAEA, Nexia Solutions (Springfield, UK) and IAEA for providing samples of yellow cakes. The support provided by the Australian Safeguards and Non-proliferation Office (ASNO) for the acquisition of samples from Australia is highly appreciated. We will also like to thank Institute for Nuclear Waste Disposal, Karlsruhe, Germany for allowing us to use their Raman spectrometer. Finally, the authors will like to thank Adrian Nicholl, Patric Lindqvist-Reis and Dieter Schild for their kind assistance and support.

References Fig. 5. Supervised visualization of the spectra using Fisher Linear Discriminant Analysis for dimensional reduction.

characterized. However, it should be noted that there is a cluster of samples from ADU, U3O8 and UO2 materials that are closely coupled. This suggests that materials from these UOC compositions are similarly composed and would be more difficult to discern using the results from the Raman spectral analysis. There is still a good degree of separation between these sample types, though less pronounced as with the UO3, UF4 and UO4 sample types. Fig. 5 also shows a strong group of a few compositions of samples. It would appear that there are a number of UOC materials that have very similar spectral compositions and therefore, they would be difficult to distinguish from one another. The cluster towards the bottom of the plot includes samples from the ADU, UH, SDU and AUC sample groups, these are very tightly grouped into a cluster and therefore based on this visual analysis of the samples it would be almost impossible to distinguish based on the Raman spectra of the samples. These result are consistent with the results that were shown from the PCA of the samples where the ADU and UH samples formed overlapping clusters and were difficult to differentiate. However, unlike the PCA results this cluster also includes samples from the SDU and AUC groups. It is possible that some of the spectral information may have been lost during DCT pre-processing. 4. Conclusions Two visualization methods and a classification method have been used for the evaluation of the Raman spectra obtained from the measurements of uranium compounds, for nuclear forensic purposes. PCA provides a view on the distribution of the samples. FDA provides visualization where the samples have not only been represented in a reduced dimensional space, but samples belonging to the same underlying compound types. The supervised visualization results reinforce the assertions made from the unsupervised PCA visualizations. There are a number of sample types that exhibit different properties from one another, suggesting that it would be possible to identify the UOC class that an unidentified sample belongs to. However, there are some notable exceptions to this and this is likely due to the loss of spectral information during DCT-preprocessing. PLS-DA shows rather promising results in the event of having to identify the different compounds. With exceptions to some uranyl hydroxides and

[1] K. Mayer, M. Wallenius, T. Fangha¨nel, Nuclear forensic science – from cradle to maturity, J. Alloys Compd. 444–445 (2007) 50–56. [2] M. Wallenius, K. Mayer, I. Ray, Nuclear forensic investigations: two case studies, Forensic Sci. Int. 156 (2006) 55–62. [3] M. Wallenius, K. Lu¨tzenkirchen, K. Mayer, I. Ray, L.A. de las Heras, M. Betti, et al., Nuclear forensic investigations with a focus on plutonium, J. Alloys Compd. 444–445 (2007) 57–62. [4] M.J. Kristo, S.J. Tumey, The state of nuclear forensics, Nucl. Instrum. Methods Phys. Res., Sect. B: Beam Interact. Mater. Atoms 294 (2013) 656–661. [5] D. Connelly, Uranium processing, Int. Min. (January) (2008) 58–61. [6] J.G.H.D. Preez, A review of the industrial processes involving uranium – from the ore to the reactor, Radiat. Prot. Dosim. 26 (1989) 7–13. [7] E. Keegan, S. Richter, I. Kelly, H. Wong, P. Gadd, H. Kuehn, et al., The provenance of Australian uranium ore concentrates by elemental and isotopic analysis, Appl. Geochem. 23 (2008) 765–777. [8] G.A. Brennecka, L.E. Borg, I.D. Hutcheon, M.A. Sharp, A.D. Anbar, Natural variations in uranium isotope ratios of uranium ore concentrates: understanding the 238 U/235U fractionation mechanism, Earth Planet. Sci. Lett. 291 (2010) 228–233. [9] Z. Varga, M. Wallenius, K. Mayer, E. Hrnecek, Alternative method for the production date determination of impure uranium ore concentrate samples, J. Radioanal. Nucl. Chem. 290 (2011) 485–492. [10] J. Sˇvedkauskaite-LeGore, K. Mayer, S. Millet, A. Nicholl, G. Rasmussen, D. Baltrunas, Investigation of the isotopic composition of lead and of trace elements concentrations in natural uranium materials as a signature in nuclear forensics, Radiochim. Acta 95 (2007) 601–605. [11] Z. Varga, M. Wallenius, K. Mayer, E. Keegan, S. Millet, Application of lead and strontium isotope ratio measurements for the origin assessment of uranium ore concentrates, Anal. Chem. 81 (2009) 8327–8334. [12] J. Sˇvedkauskaite˙-LeGore, G. Rasmussen, S. Abousahl, P. Belle, Investigation of the sample characteristics needed for the determination of the origin of uraniumbearing materials, J. Radioanal. Nucl. Chem. 278 (2008) 201–209. [13] S.-H. Han, Z. Varga, J. Krajko, M. Wallenius, K. Song, K. Mayer, Measurement of the sulphur isotope ratio (34S/32S) in uranium ore concentrates (yellow cakes) for origin assessment, J. Anal. At. Spectrom. 28 (2013) 1919–1925. [14] J. Krajko´, Z. Varga, E. Yalcintas, M. Wallenius, K. Mayer, Application of neodymium isotope ratio measurements for the origin assessment of uranium ore concentrates, Talanta 129 (2014) 499–504. [15] Z. Varga, R. Katona, Z. Stefa´nka, M. Wallenius, K. Mayer, A. Nicholl, Determination of rare-earth elements in uranium-bearing materials by inductively coupled plasma mass spectrometry, Talanta 80 (2010) 1744–1749. [16] Z. Varga, M. Wallenius, K. Mayer, Origin assessment of uranium ore concentrates based on their rare-earth elemental impurity pattern, Radiochim. Acta: Int. J. Chem. Aspects Nucl. Sci. Technol. 98 (2010) 771–778. [17] E. Keegan, M. Wallenius, K. Mayer, Z. Varga, G. Rasmussen, Attribution of uranium ore concentrates using elemental and anionic data, Appl. Geochem. 27 (2012) 1600–1609. [18] A.K. Kennedy, D.A. Bostick, C.R. Hexel, R.R. Smith, J.M. Giaquinto, Non-volatile organic analysis of uranium ore concentrates, J. Radioanal. Nucl. Chem. 296 (2013) 817–821. [19] V. Badaut, M. Wallenius, K. Mayer, Anion analysis in uranium ore concentrates by ion chromatography, J. Radioanal. Nucl. Chem. 280 (2009) 57–61. ¨ ztu¨rk, M. Meppen, K. Mayer, M. Wallenius, C. Apostolidis, Charac[20] Z. Varga, B. O terization and classification of uranium ore concentrates (yellow cakes) using infrared spectrometry, Radiochim. Acta 99 (2011) 1–7. [21] J.W. Plaue, G.L. Klunder, I.D. Hutcheon, K.R. Czerwinski, Near infrared reflectance spectroscopy as a process signature in uranium oxides, J. Radioanal. Nucl. Chem. 296 (2013) 551–555. [22] G.L. Klunder, J.W. Plaue, P.E. Spackman, P.M. Grant, R.E. Lindvall, I.D. Hutcheon, Application of visible/near-infrared reflectance spectroscopy to uranium ore

68

[23]

[24]

[25] [26] [27]

[28]

[29]

[30]

[31] [32]

D.M.L. Ho et al. / Forensic Science International 251 (2015) 61–68 concentrates for nuclear forensic analysis and attribution, Appl. Spectrosc. 67 (2013) 1049–1056. J.-B. Sirven, A. Pailloux, Y. M’Baye, N. Coulon, T. Alpettaz, S. Gosse, Towards the determination of the geographical origin of yellow cake samples by laser-induced breakdown spectroscopy and chemometrics, J. Anal. At. Spectrom. 24 (2009) 451–459. D. Ho Mer Lin, D. Manara, Z. Varga, A. Berlizov, T. Fangha¨nel, K. Mayer, Applicability of Raman spectroscopy as a tool in nuclear forensics for analysis of uranium ore concentrates, Radiochim. Acta 101 (2013) 779–784. Z. Varga, M. Wallenius, K. Mayer, M. Meppen, Analysis of uranium ore concentrates for origin assessment, Proc. Radiochim. Acta 1 (2011) 1–4. P.A. Budinger, T.L. Drenski, A.W. Varnes, J.R. Mooney, The case of the great yellow cake caper, Anal. Chem. 52 (1980) 942A–948A. E. Keegan, M.J. Kristo, M. Colella, M. Robel, R. Williams, R. Lindvall, et al., Nuclear forensic analysis of an unknown uranium ore concentrate sample seized in a criminal investigation in Australia, Forensic Sci. Int. 240 (2014) 111–121. F. Pointurier, O. Marie, Identification of the chemical forms of uranium compounds in micrometer-size particles by means of micro-Raman spectrometry and scanning electron microscope, Spectrochim. Acta B: At. Spectrosc. 65 (2010) 797–804. M. Amme, B. Renker, B. Schmid, M.P. Feth, H. Bertagnolli, W. Do¨belin, Raman microspectrometric identification of corrosion products formed on UO2 nuclear fuel during leaching experiments, J. Nucl. Mater. 306 (2002) 202–212. M. Mellini, F. Riccobono, Chemical and mineralogical transformations caused by weathering in anti-tank DU penetrators (‘‘the silver bullets’’) discharged during the Kosovo war, Chemosphere 60 (2005) 1246–1252. E.A. Stefaniak, et al., Recognition of uranium oxides in soil particulate matter by means of m-Raman spectrometry, J. Nucl. Mater. 381 (2008) 278–283. I.S. Butler, G.C. Allen, N.A. Tuan, Micro-Raman spectrum of triuranium octoxide, U3O8, Appl. Spectrosc. 42 (1988) 901–902.

[33] M.L. Palacios, S.H. Taylor, Characterization of uranium oxides using in-situ microRaman spectroscopy, Appl. Spectrosc. 54 (2000) 1372–1378. [34] B.M. Wise, N.B. Gallagher, R. Bro, J.M. Shaver, W. Windig, R.S. Koch, PLS_Toolbox Version 4.0 for Use with MATLAB, Eigenvector Research, Inc., WA, USA, 2006. [35] M.A.F. De La Ossa, C. Garcı´a-Ruiz, J.M. Amigo, Near infrared spectral imaging for the analysis of dynamite residues on human handprints, Talanta 130 (2014) 315–321. [36] M. Bassbasi, M. De Luca, G. Ioele, A. Oussama, G. Ragno, Prediction of the geographical origin of butters by partial least square discriminant analysis (PLS-DA) applied to infrared spectroscopy (FTIR) data, J. Food Compos. Anal. 33 (2014) 210–215. [37] L. Fongaro, K. Kvaal, Surface texture characterization of an Italian pasta by means of univariate and multivariate feature extraction from their texture images, Food Res. Int. 51 (2013) 693–705. [38] N. Ahmed, Discrete cosine transform, IEEE Trans. Comput. 23 (1974) 90–93. [39] M. Welling. Fisher Linear Discriminant Analysis. University of Toronto, Toronto. [40] M. Tingting, J. Goulermas, J. Tsujii, S. Ananiadou, Proximity-based frameworks for generating embeddings from multi-output data, IEEE Trans. Pattern Anal. Mach. Intell. 34 (2012) 2216–2232. [41] D. Ho Mer Lin, D. Manara, P. Lindqvist-Reis, T. Fangha¨nel, K. Mayer, The use of different dispersive Raman spectrometers for the analysis of uranium compounds, Vib. Spectrosc. 73 (2014) 102–110. [42] I.A. Degen, G.A. Newman, Raman spectra of inorganic ions, Spectrochim. Acta 49A (1993) 859–887. [43] Uranium Extraction Technology, International Atomic Energy Agency, 1993. [44] L. Lenhardt, R. Bro, I. Zekovic´, T. Dramic´anin, M.D. Dramic´anin, Fluorescence spectroscopy coupled with PARAFAC and PLS DA for characterization and classification of honey, Food Chem. 175 (2015) 284–291.