- Email: [email protected]
Elemental source attribution signatures for calcium ammonium nitrate (CAN) fertilizers used in homemade explosives
Talanta 174 (2017) 131–138
Contents lists available at ScienceDirect
Talanta journal homepage: www.elsevier.com/locate/talanta
Elemental source attribution signatures for calcium ammonium nitrate (CAN) fertilizers used in homemade explosives
MARK
⁎
Carlos G. Fraga , Alexander V. Mitroshkov, Nikhil S. Mirjankar, Brian P. Dockendorff, Angie M. Melville Pacific Northwest National Laboratory, 902 Battelle Boulevard, Richland, WA 99352, USA
A R T I C L E I N F O
A BS T RAC T
Keywords: Calcium ammonium nitrate fertilizers Chemical attribution signatures Chemical forensics Inductively coupled plasma-mass spectrometry Elemental profiling Chemometrics
Calcium ammonium nitrate (CAN) is a widely available fertilizer composed of ammonium nitrate (AN) mixed with some form of calcium carbonate such as limestone or dolomite. CAN is also frequently used to make homemade explosives. The potential of using elemental profiling and chemometrics to match both pristine and reprocessed CAN fertilizers to their factories of origin for use in future forensic investigations was examined. Inductively coupled plasma-mass spectrometry (ICP-MS) was used to determine the concentrations of 64 elements in 125 samples from 11 CAN stocks from 6 different CAN factories. Using Fisher ratio and degree-ofclass-separation, the elements Na, V, Mn, Cu, Ga, Sr, Ba and U were selected for classification of the CAN samples into 5 factory groups; one group was two factories from the same fertilizer company. Partial least squares discriminant analysis (PLSDA) was used to develop a classification model which was tested on a separate set of samples. The test set included samples that were analyzed at a different time period and samples from factory stocks that were not part of the training set. For pristine CAN samples, i.e., unadulterated prills, 73% of the test samples were matched to their correct factory group with the remaining 27% undetermined using strict classification. The same PLSDA model was used to correctly match all CAN samples that were reprocessed by mixing with powdered sugar. For CAN samples that were reprocessed by mixing with aluminum or by extraction of AN with tap or bottled water, correct classification was observed for one factory group, but source matching was confounded with adulterant interference for two other factories. The elemental signatures of the water-insoluble (calcium carbonate) portions of CAN provided a greater degree of discrimination between factories than the water-soluble portions of CAN. In summary, this work illustrates the strong potential for matching unadulterated CAN fertilizer samples to their manufacturing facility using elemental profiling and chemometrics. The effectiveness of this method for source determination of reprocessed CAN is dependent on how much an adulterant alters the recovered elemental profile of CAN.
1. Introduction Calcium ammonium nitrate (CAN) is a widely available inorganic fertilizer that is frequently used by terrorists and insurgents to make homemade explosives (HMEs) for improvised explosive devices [1,2]. CAN consists of ammonium nitrate (AN), at approximately 75% (w/w), mixed with some form of calcium carbonate such as limestone or dolomite [3–5]. AN is the oxidizing component in many binary HMEs. CAN fertilizer can be reprocessed to make a HME by first mixing it with hot water to separate the soluble AN from the inert and insoluble components (e.g., calcium carbonate), followed by evaporation to
recover pure AN which is dried and crushed [6]. Alternatively, CAN fertilizer can also be crushed to a powder and used directly without extraction of the inert material [6,7]. The final reprocessing step requires mixing the extracted AN or powdered CAN with reducing agents such as aluminum powder, powdered sugar, fuel oil or a combination to produce an explosive mixture [6]. Reprocessed CAN, at one point, was used as the primary explosive in over 85% of the improvised explosive devices used against coalition forces in Afghanistan [1,2]. CAN's wide availability as a legitimate fertilizer will continue to make it and other AN based fertilizers a global HME threat for the foreseeable future. Hence efforts are underway to examine new
Abbreviations: CAN, Calcium ammonium nitrate; AN, Ammonium nitrate; HME, Homemade explosives; ICP-MS, Inductively coupled plasma-mass spectrometry; FTIR, Fouriertransform infrared spectroscopy; DCS, Degree-of-class separation; HCA, Hierarchical cluster analysis; PCA, Principal Component Analysis; PLSDA, Partial least squares discriminant analysis; DI, Deionized; RSD, Relative standard deviation ⁎ Corresponding author. E-mail address: [email protected] (C.G. Fraga). http://dx.doi.org/10.1016/j.talanta.2017.05.066 Received 25 April 2017; Received in revised form 20 May 2017; Accepted 23 May 2017 Available online 26 May 2017 0039-9140/ © 2017 Elsevier B.V. All rights reserved.
Talanta 174 (2017) 131–138
C.G. Fraga et al.
Table 1 CAN stocks and number of stock samples used in each sample set. Stock ID
Source
A12 B98 B12 B13 C11 C12 C13 D13 E13a E13b F83
Factory Factory Factory Factory Factory Factory Factory Factory Factory Factory Factory
a b
A B B B C C C D E E F
Manufacture date
Set 1 May 2013a
Set 2 Jan 2015a
Set 3 May 2015a
2012 1998 2012 2013 2011 2012 2013 2013 2013 2013 1983
3 (6b) 3 3 3 (6b) 3 – – – – – –
3 3 3 3 3 3 3 3 3 3 3
8 (6b) 4 (6b) 4 4 4 4 4 3 2 2 8 (6b)
(1st Qtr.)
(17 May) (25 Apr) (22 Mar)
time period CAN samples were prepared and analyzed by ICP-MS. number of “reprocessed” CAN samples that involved crushing CAN and mixing it with either sugar, aluminum or non-DI water.
2. Materials and methods
analytical approaches for establishing the origin of CAN and other HME precursors to support forensic investigations such as those that aim to identify and disrupt illicit HME-precursor distribution networks. CAN fertilizers, like AN fertilizers, are available as prills or granules that are typically white, grey or beige in color and 2–5 mm in diameter. CAN fertilizer is produced in various closely related formulations by mixing concentrated AN with finely ground limestone or dolomite resulting in a CAN product that is 20–28% N [8,9]. AN is produced by reacting nitric acid with ammonia. CAN is also made by reacting calcium nitrate with ammonia and carbon dioxide to yield a mixture of AN and calcium carbonate containing 20–26% N [8]. Due to its hygroscopic nature, CAN is coated with anticaking agents and additives such as gypsum, kieselguhr, and magnesium nitrate, which are applied during prilling or granulation [8]. A typical composition of CAN fertilizer consists 27% N (13.5% ammonia-N and 13.5% nitrate-N), 6–9% Ca, and 4% Mg [10], with calcium and magnesium present as carbonates (CaCO3 and MgCO3) as well as oxides (CaO and MgO) [9]. These various constituents and additives could provide unique chemical signatures specific to the manufacturing source of CAN because of bulk and trace compositional differences in CAN that are source dependent. For example, trace elements in the water and mineral additives used in making CAN are dependent on the local geology where acquired and therefore may impart trace elemental signatures into CAN that are source dependent. There are no previous published works on the source determination of CAN fertilizers. Recent papers have demonstrated the potential of linking AN fertilizers to their places of manufacture using isotopic and elemental profiling [11–13]. In particular, Brust and co-workers used linear discriminant analysis and likelihood ratios to demonstrate the discrimination of AN from one manufacturer from 16 others based on the concentrations of 32 elements and the stable isotopic ratios of oxygen and nitrogen [12]. Their work was quite convincing in part because of their large sample collection that included samples from 17 batches and two fertilizer types from one AN fertilizer manufacturer. Similar work has demonstrated the potential of sourcing commercial cyanides to their places of manufacture using isotopic, elemental and ionic profiles combined with chemometric analysis [14]. The work described here further develops the concept of sourcing commercial chemical products to their places of manufacture through the discovery and exploitation of intrinsic chemical signatures in CAN fertilizers. Specifically, the use of elemental profiling to source CAN fertilizers to a larger number of factories than previously shown for AN and cyanide is demonstrated. This includes investigating the sourcing of CAN that has been adulterated by reprocessing methods used for making HMEs. Lastly, the potential of using chemical attribution signatures from specific chemical constituents to obtain a more precise level of source discrimination is illustrated using the water-insoluble components of CAN.
2.1. CAN stocks and sample sets A total of 11 stocks of commercial CAN fertilizer (26 – 27% N) that originated from six factories were used in this study. Each stock was a bulk sample of CAN prills or granules made at a specific factory and time period. From this point forth, the term “prills” refers to either granules or true prills. Each CAN stock and corresponding source information (factory and manufacture date) were obtained either from a manufacturer, distributor, or reputable field representative. The factory sites were confirmed using a list of known CAN factories [8]. F83 was the only stock that was not matched to a specific factory, but based on its region of origin and its distinct prill size and appearance was believed to not be from any of the other factories. The identity of each stock was confirmed as CAN by FTIR analysis using a Gemini analyzer (Thermo Scientific). CAN samples from these 11 stocks were prepared and then analyzed by inductively coupled plasma-mass spectrometry (ICP-MS) in three sets. Table 1 lists the 11 CAN stocks with their corresponding source information and number of samples taken from each stock for each sample set. 2.2. Sample preparation CAN samples were prepared according to four methods: (1) CAN prills were dissolved in deionized (DI) water to produce a water-soluble portion and a water-insoluble portion, (2) water-insoluble portions were dissolved in nitric acid to produce a solution, (3) CAN prills were crushed and mixed with either aluminum powder or powdered sugar in water to produce a water-soluble portion and a water-insoluble portion, and (4) CAN prills were crushed, mixed with either tap or bottled water, and the supernatant dried to yield ammonium nitrate (AN). Methods (3) and (4) simulated CAN that is reprocessed as part of HME production. For all methods, the vessels in contact with CAN or its solutions were plastic. Details for each method are described below. For method (1), a total of 95 pristine CAN samples (those in Table 1 not denoted as “reprocessed”) were prepared by mixing 100 mg of CAN prills (typically 2–3 prills) with 10 mL of DI water (18 MΩ•cm from Nanopure, Thermo Scientific) using a 30 s vortex followed by 15 min sonication. Each sample mixture was then centrifuged for 10 min to separate the insoluble components, primarily some form of calcium carbonate. The water-soluble portion of each CAN sample was stored for ICP-MS analysis. For method (2), 20 mg of a water-insoluble portion was dissolved completely by mixing with 10 mL of 5% nitric acid (Optima, Fisher Scientific) and stored for ICP-MS analysis. This was done for a subset of Set 1 and Set 3 CAN samples. For method (3), a subgroup of 12 CAN samples from Set 1 and Set 3 132
Talanta 174 (2017) 131–138
C.G. Fraga et al.
(denoted in Table 1 as “reprocessed”) were each crushed to produce 100 mg of powdered CAN that was mixed with 10 mg of powdered sugar (C & H Pure Cane) or 10 mg of aluminum powder (Tannerite Kill Shot Target) in 10 mL of DI water. Each aqueous mixture of CAN and reducing agent (sugar or Al) was then vortexed, sonicated, and centrifuged just as the pristine CAN samples described in method (1). The water soluble portion of each CAN sample was stored for ICPMS analysis. For method (4), 200 mg of crushed CAN (individually from 18 CAN samples denoted in Table 1 as “reprocessed”) was mixed with either 10 mL of tap water (Richland, WA) or 10 mL of bottled water (Crystal Geyser Alpine Spring Water by CG Roxane, Weed, CA). Each aqueous mixture was then vortexed, sonicated, and centrifuged similar to the pristine CAN samples in method (1). Each water soluble portion was then placed into a Reacti-Therm module (Thermo Scientific) where it was heated at 40 °C under a constant stream of dry nitrogen until all the water evaporated leaving behind dried AN. Approximately 100 mg of each AN portion extracted from CAN was then mixed with 10 mL of DI water and stored for ICP-MS analysis.
(Mathworks, Natick, MA) and PLS Toolbox 7.9.2 (Eigenvector Research, Manson, WA). The data from Set 1 and Set 2 were combined for exploratory analysis to determine the optimal elements for classifying CAN samples according to source. Variable selection was performed using the Fisher-ratio method [17] and degree-of-class separation (DCS) [18]. This approach has been successfully used for obtaining chemical attribution signatures [14,19]. Clustering of sample data based on the selected variables, i.e., concentrations of elements, was assessed using hierarchical cluster analysis (HCA) and principal component analysis (PCA) [20]. Supervised classification using the selected elements was then performed by partial least squares discriminant analysis (PLSDA) [21] using samples from Set 1 and Set 2 as the training set. PLSDA uses PLS regression for directly identifying variations in data space that discriminate classes with less noise compared to linear discriminant analysis. This technique has worked well for similar data sets from our previous work [14]. The PLSDA classification was assessed on a test set consisting of samples from Set 3 and samples from Set 1 and Set 2 whose stocks were not represented in the training set. For HCA, PCA, and PLSDA, the data were preprocessed by unit-area normalization; i.e., for each sample, the concentration of each selected element was divided by the sum of all selected elemental concentrations.
2.3. ICP-MS analysis A NU Plasma high resolution multi-collector ICP-MS (NU Instruments) was used to determine the concentrations of 64 elements: Li, Be, Na, Al, K, Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Ga, Ge, As, Se, Rb, Sr, Y, Zr, Nb, Mo, Ru, Rh, Pd, Ag, Cd, In, Sn, Sb, Te, Cs, Ba, La, Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Ho, Er, Tm, Yb, Lu, Hf, Ta, W, Re, Os, Ir, Pt, Au, Hg, Tl, Pb, Bi, Th, and U. The ICP-MS instrument included 11 Faraday cups and 3 ion counters used as detectors. Aqueous samples were introduced using an ASX-112FR autosampler connected to a QuickWash accessory and an Aridus II Desolvating Nebulizer (all from CETAC). The aqueous CAN samples, i.e., their water-soluble fractions in DI water or water-insoluble fractions in 5% nitric acid, were diluted to 1/100 (v/v) with 2% nitric acid (Optima, Fisher Scientific). Representative CAN samples were first screened using Elemental Mode analysis, with only the central Faraday cup Ax, in order to adjust ICP-MS parameters. For quantitative analysis, a scan rate of 1000 ms/amu and a dwell time of 2000 ms were used. Mass scale calibration was performed using two different aqueous standards: ICPMS-IS-1 and ICP-MS-68A (both from High Purity Standards). The ICPMS-68A standard consisted of 68 elements and was diluted to produce two solutions of 8 and 10 parts-per-billion (ppb) used as external standards; the 10-ppb standard was used for quantitative analysis of the Set 1 and Set 2 samples and the 8-ppm standard for the Set 3 samples. Each respective standard and a method blank were analyzed after every three sample analyses. The standard and blank analyzed closest in time to an analyzed sample were used for quantification by one-point calibration using the respective elemental ICP-MS signal intensities (in volts) as shown by Eq. (1):
⎛ Isample − Iblank ⎞ Csample = ⎜ ⎟ × Cstandard × D ⎝ Istandard − Iblank ⎠
3. Results and discussion 3.1. Elemental signatures for the factory discrimination of CAN Robust elemental signatures that point to a specific source (e.g., CAN factory) are not affected by experimental variability and withinsource variability. In this study, experimental variability was addressed by performing sample preparations and ICP-MS analysis on the same CAN stocks at two different time periods, in 2013 (Set 1) and 2015 (Set 2), in order to capture changes in instrument performance, solvent stocks (e.g., DI water), sample-preparation personnel, and environmental conditions. In terms of within-source variability, the sample preparation and analysis of multiple stocks from the same CAN factory made at different time periods provided a means to address the
(1)
where Csample and Cstandard are the concentrations of the element in the sample and standard respectively, and Cstandard is either 8 or 10 depending on whether the 8-ppb or 10-ppb standard was used. Isample, Istandard, and Iblank are the signal intensities for sample, standard, and blank, respectively. D is the dilution factor with a value of 100. Elements whose sample intensity (Isample) was less than the corresponding signal in the blank (Iblank) were assigned a zero for concentration. This quantitative approach [15] proved to be simple and effective for the elemental profiling of each CAN aqueous sample in part because of the five orders of signal linearity inherent to Faraday-cup ICP-MS instruments [16]. Fig. 1. Dendrogram using the normalized (unit area) concentrations of V, Mn, Sr, and U for Set-1 and Set-2 CAN samples. The CAN samples are the water soluble portions of 11 CAN stocks from 6 factories (A – F). The 48 samples cluster into four groups, A, B/C, D/ F, and E, with sample B13-2-2 being the only outlier. The sample label format is: (stock ID)-(sample set)-(sample replicate).
2.4. Chemometric analysis Chemometric analysis was performed using Matlab R2014a 133
Talanta 174 (2017) 131–138
C.G. Fraga et al.
compositional variability in the manufacture of CAN from changes, for example, in feedstock and operating conditions. Variable selection and cluster analysis were then used on the ICP-MS data to select and determine those elements in CAN whose concentrations resulted in CAN stocks grouping according to their factories regardless of when analyzed or when manufactured. Variable selection was performed using Fisher ratio and DCS calculations on the pristine (not reprocessed) CAN elemental concentrations of samples from Set 1 and Set 2. The elements V, Mn, Sr, and U were found to differentiate the water-soluble portions of CAN samples into four factory groups: A, B/C, D/F, and E. Fig. 1 is a HCA dendrogram (based on Euclidean distance) that depicts the differentiation of the CAN samples into the four factory groups. As seen, CAN samples from Set 1 and Set 2 cluster with other samples from the same group, thereby demonstrating the robustness of the four selected elements towards experimental differences between the two sample sets. In addition, the clustering of stock samples in Fig. 1 with other stock samples from the same factory group (B/C and E) illustrates the robustness of the four selected elements in terms of source variability. For data sets that are difficult to classify in a single step, a step-wise classification approach can be applied [22,23]. Such a multi-level classification approach is particularly useful for discriminating classes that under an initial set of variables are separated into groups that include more than one class. Variable selection was further investigated to differentiate the B/C group and D/F group into separate factories. In the case of the D/F group, the elements Na, Cu, Ga, and Ba were found to separate the D and F CAN samples into two clusters, one for each factory (D and F). The separation of the B/C factory group into two distinct HCA groups with variable selection was not achievable. Factories B and C are different, but they belong to the same manufacturing company and are geographically near one another. It is therefore possible that they use similar reagent stocks and processes in the making of the water-soluble components (i.e., primarily AN) used in their CAN products, accounting for the similar elemental concentrations in the water-soluble fractions. Fig. 2A shows a PCA scores plot of all 64 element concentration profiles compared to a PCA scores plot of profiles for the selected eight elements Na, V, Mn, Cu, Ga, Sr, Ba, and U (Fig. 2B). It is evident from the figure that the eight-element profiles can be used for discrimination of the CAN samples into five factory groups. Although a univariate approach was utilized for feature selection, agreement between the results obtained from the same data using HCA and PCA provides a strong justification for dividing the Set 1 and Set 2 CAN samples into five factory groups: A, B/C, D, E, and F. With the future acquisition of CAN stocks from other factories, a new set of elements (currently out of 64) may be needed to differentiate CAN samples into factory groups. Presumably, the new set would include most of the eight elements selected here because of their ability to discriminate the current CAN samples into factory groups. Table 2 lists the mean concentration and relative standard deviation (RSD) for each of the eight selected elements in each CAN stock analyzed in Set 1 and Set 2. The table illustrates some of the differences in elemental concentrations that help differentiate the CAN stocks into five factory groups. For example, the mean U concentration in stock A12 from factory A is 2–35 times greater than in the stocks from factories B, C, D, and E. Typically, elemental concentrations are compared by normalizing to the sample mass (here, amount of CAN analyzed). However, this is not always straightforward for some field samples such as those that are crushed CAN mixed with a reducing agent. To address this issue, we normalized the elemental concentrations by using unit-area normalization. Given that the ICP-MS elemental signals for the 10-ppb multielement standard had a daily RSD range of 11–21%, the RSD values in Table 2 that are greater than 21% are likely caused from within-stock variability and not instrumental variability. In the end, the measured variabilities in concentration of the selected elements in each CAN
Fig. 2. PCA scores plot of 48 CAN samples from Set-1 and Set-2 using (A) 64 element concentration profile and (B) selected 8 element profile of Na, V, Mn, Cu, Ga, Sr, Ba, and U. With all 64 elements, the CAN samples do not cluster according to factory groups, however using the selected 8 elements, the CAN samples cluster into five factory groups: A, B/C, D, E, and F. The data was normalized to unit area and autoscaled prior for PCA.
stock are small relative to the differences in elemental profiles between the factories (except for B and C) as indicated by the clustering of CAN samples according to factory group (Fig. 1 and Fig. 2B). This notion was assessed through the classification of a test set of samples according to factory group, as described in the following sections. 3.2. Factory classification using the water-soluble portions of pristine CAN A single PLSDA model was created for five factory groups (A, B/C, D, E, and F) using the ICP-MS concentrations of the eight elements Na, V, Mn, Cu, Ga, Sr, Ba, and U in the water-soluble portions of the pristine (not reprocessed) Set 1 and Set 2 CAN samples (see Table 1). Only 6 out of the 11 CAN stocks (A12, B12, C11, D13, E13a, and F83) were used for training the PLSDA model. Elemental concentrations were normalized to unit area and autoscaled prior to model creation. The PLSDA model, based on 27 CAN samples, consisted of 4 latent variables that were determined through venetian blind cross-validation with 10 splits and 1 sample per split. The PLSDA model captured 94% of the elemental data variance. The created PLSDA model was tested on the elemental concentrations obtained from the water-soluble portions of 68 pristine CAN samples. The test set included all pristine samples from Set 3, and the 134
Talanta 174 (2017) 131–138
C.G. Fraga et al.
Table 2 Mean concentration in ppb and RSD (in parentheses) for each of the eight selected elements in the DI water-soluble portionsa of the CAN Set 1 and Set 2 stock samples. Stock ID
No. of samples
Na
V
Mn
Cu
Ga
Sr
Ba
U
A12
6
550 (18%)
0.67 (24%)
14 (4.9%)
4.6 (78%)
0.50 (18%)
21 (8.1%)
2.4 (5.7%)
0.38 (22%)
B98
6
550 (34%)
0.17b (44%)
60 (33%)
26 (100%)
0.64 (39%)
13 (32%)
3.2 (52%)
0.013b (75%)
B12
6
690 (14%)
0.16 (67%)
80 (4.5%)
23 (140%)
0.78 (25%)
29 (5.5%)
4.2 (9.1%)
0.048 (42%)
B13
6
710 (36%)
0.17 (38%)
76 (25%)
33 (73%)
0.97 (11%)
47 (39%)
4.7 (14%)
0.056 (46%)
C11
6
170 (31%)
0.099c (61%)
35 (26%)
14 (79%)
0.67 (8.7%)
13 (31%)
2.8 (12%)
0.036b (18%)
C12
3
140 (22%)
0.13 (17%)
34 (5.0%)
26 (59%)
6.8 (76%)
7.3 (11%)
30 (29%)
0.029 (70%)
C13
3
280 (54%)
0.14 (30%)
27 (13%)
24 (34%)
3.4 (150%)
7.3 (14%)
12 (150%)
0.040 (6.7%)
D13
3
190 (13%)
1.8 (20%)
24 (10%)
11 (6.4%)
4.4 (15%)
150 (15%)
15 (8.9%)
0.19 (33%)
E13a
3
710 (13%)
1.7 (6.9%)
0.40 (7.7%)
4.5 (9.0%)
0.76 (13%)
190 (7.8%)
3.4 (5.1%)
0.14 (9.6%)
E13b
3
780 (6.3%)
1.8 (3.7%)
1.1 (10%)
6.2 (19%)
1.8 (13%)
300 (8.0%)
7.4 (4.8%)
0.17 (8.8%)
F83
3
590 (14%)
0.87 (22%)
20 (8.9%)
5.5 (18%)
11 (13%)
170 (7.4%)
41 (6.1%)
0.27 (18%)
a b c
The supernatant from mixing 100 mg of CAN with 10 mL of DI water. One sample had a concentration of zero and was not included. Two samples had concentrations of zero and were not included.
magnitude. In probable classification, all samples are classified so there are no unclassified samples. Several interesting observations and conclusions can be made from Table 3 regarding the ability to match the CAN samples to their sources. First, strict classification correctly matched 73% of the CAN samples to their correct class with no misclassifications; 27% were unclassified. This means that a sample that is reported as unclassified may have originated from one of the stocks that was actually modeled; therefore, if an unknown sample is listed as unclassified, it does not necessarily mean it came from a new source that was not modeled. Second, correct matching was observed for samples from factories A, B
Set 1 and Set 2 samples of 5 CAN stocks (B98, B13, C12, C13, and E13b) that were not included in training the PLSDA model. These samples were treated as unknowns. Table 3 summarizes their classification results. The term “strict” in Table 3 refers to strict classification where unknown samples are classified to a specific class if the probability is greater than a specified threshold (probability value of 0.5) for only one class. If no class has a probability greater than the threshold or if more than one class has probability exceeding the threshold, then the sample is assigned “no class” and is considered “unclassified”. In “probable” classification, unknown samples are matched to a class that has the highest probability regardless of the
Table 3 ICP-MS elemental PLSDAa classification results for test set of water-soluble pristine CAN samples. Samples
A12: 8 samples B98b: 10 samples B12: 4 samples B13b: 10 samples C11: 4 samples C12b: 7 samples C13b: 7 samples D13: 3 samples E13a: 2 samples E13bb: 5 samples F83: 8 samples 68 samples
Class
A B/C B/C B/C B/C B/C B/C D E E F 5 classes
Strict
Probable
Correct
Misclassified
Unclassified
Correct
Misclassified
7 (87.5%) 10 (100%) 3 (75%) 10 (100%) 1 (25%) 0 (0%) 4 (57.1%) 2 (66.7%) 2 (100%) 3 (60%) 8 (100%) 50 (73.5%)
0 0 0 0 0 0 0 0 0 0 0 0
1 (12.5%) 0 (0%) 1 (25%) 0 (0%) 3 (75%) 7 (100%) 3 (42.9%) 1 (33.3%) 0 (0%) 2 (40%) 0 (0%) 18 (26.5%)
8 (100%) 10 (100%) 4 (100%) 10 (100%) 3 (75%) 3 (42.9%) 6 (85.7%) 2 (66.7%) 2 (100%) 4 (80%) 8 (100%) 60 (88.2%)
0 0 0 0 1 4 1 1 0 1 0 8
(0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%)
(0%) (0%) (0%) (0%) (25%) (57.1%) (14.3%) (33.3%) (0%) (20%) (0%) (11.8%)
a PLSDA model created for 5 classes using ICP-MS elemental concentrations of Na, V, Mn, Cu, Ga, Sr, Ba and U from pristine CAN Set-1 and Set-2 water-soluble samples: A12 (6 samples), B12 (6 samples), C11 (6 samples), D13 (3 samples), E13a (3 samples), F83 (3 samples). b Training set did not include samples from these stocks.
135
Talanta 174 (2017) 131–138
C.G. Fraga et al.
Table 4 ICP-MS Elemental PLSDAa Classification Results for Reprocessed CAN Samples form Set 1 and Set 3. Samples
Class
Strict
Probable
Correct
Misclassified
Unclassified
Correct
Misclassified
A12 + sugar: 2 samples (Set 3)
A
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
B98b + sugar: 2 samples (Set 3)
B/C
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
F83 + sugar: 2 samples (set 3)
F
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
A12 + Al: 2 samples (Set 3)
A
0 (0%)
1 (50%)
1 (50%)
0 (0%)
2 (100%)
B98b + Al: 2 samples (Set 3)
B/C
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
F83 + Al: 2 samples (Set 3)
F
1 (50%)
0 (0%)
1 (50%)
2 (100%)
0 (0%)
A12 tap-water extracted AN: 5 samples (Set 1 & Set 3)
A
4 (80%)
1 (20%)
0 (0%)
4 (80%)
1 (20%)
A12 bottle water extracted AN: 3 samples (Set 1)
A
1 (33.3%)
0 (0%)
2 (66.7%)
3 (100%)
0 (0%)
B98b tap-water extracted AN: 2 samples (Set 3)
B/C
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
B13 tap-water extracted AN: 3 samples (Set 1)
B/C
3 (100%)
0 (0%)
0 (0%)
3 (100%)
0 (0%)
B13 bottle-water extracted AN: 3 samples (Set 1)
B/C
3 (100%)
0 (0%)
0 (0%)
3 (100%)
0 (0%)
F83 tap water extracted AN: 2 samples (Set 3) 30 samples
F
0 (0%)
2 (100%)
0 (0%)
0 (0%)
2 (100%)
3 classes
22 (73.3%)
4 (13.3%)
4 (13.3%)
25 (83.3%)
5 (16.7%)
a b
Same 5-classes PLSDA model used on the pristine CAN samples in Table 3 based on concentrations of Na, V, Mn, Cu, Ga, Sr, Ba and U was utilized. Training set did not include samples from these stocks.
correctly classified. This result is not surprising given that household sugar should have a zero to low metal content and therefore adulterant interference is unlikely. On the other hand, aluminum powder has metal impurities that likely caused half of the samples from stock A12 to be misclassified (half were unclassified) and half of those from stock F83 to be unclassified (half were correctly classified) by strict classification. Surprisingly, all of the B98 CAN samples mixed with aluminum powder were correctly classified. Apparently, the 8-element profile for B98 CAN was not adversely affected by the metal impurities in aluminum. The tap and bottled waters used in this study also have metal constituents. Surprisingly, all the B98 and B13 CAN samples whose AN was extracted using tap or bottled water and then dried and analyzed by ICP-MS were correctly classified. Just as for the aluminum powder, the 8-element profiles for B98 and B13 were not adversely affected by the metals in the bottled or tap waters. Although a few samples from factory B were tested, it seems so far that the 8-element profile from factory B is the most robust in terms of withstanding adulterant interference. At a lesser level, 4 out 5 of the samples from stock A12 whose AN was extracted using tap water were correctly classified while 1 out 5 were misclassified under both strict and probable classification. For the bottled water samples, 1 out 3 of the A12 samples were correctly classified with zero misclassifications under strict classification; under probable classification, all of the A12 samples were correctly classified. For the F83 stock, all two reprocessed samples involving tap water were misclassified under both strict and probable
and F with high classification success rates; ~ 88% for stock A12 and 100% for stocks B98, B13, and F83. This suggests that similar stocks from these three factories can be likely matched to their designated source. Third, probable classification correctly matched 88% of all samples to their correct class while 12% were incorrectly matched. The probable approach is best utilized when the unknown samples are suspected to belong to one of the modeled classes. 3.3. Factory classification using the water-soluble portions of reprocessed CAN In the previous section, the “unknown” pristine CAN samples represent field samples of CAN, either whole prills or crushed prills, that have not been adulterated by mixing with a reducing agent or environmental media. Another type of field sample is one in which the CAN has been reprocessed by either crushing and mixing the CAN with a reducing agent or crushing the CAN and mixing it with water to separate the AN from CAN and then drying the aqueous supernatant. Some of the Set 1 and Set 3 CAN samples were reprocessed as mentioned. These “reprocessed” samples (see Table 1) were 30 CAN samples that were either mixed with sugar or aluminum, or had their AN extracted from CAN using tap or bottled water and then prepared for ICP-MS analysis. Table 4 summarizes the classification results for the reprocessed samples using their ICP-MS elemental data and the same PLSDA model used for the pristine CAN samples. As shown in Table 4, all of the CAN samples mixed with sugar were 136
Talanta 174 (2017) 131–138
C.G. Fraga et al.
Fig. 3. Dendrograms using concentrations of V, Mn, Sr, and U for (A) 14 samples that are the water soluble portions of 5 CAN stocks and (B) their corresponding samples from the 5% nitric-acid dissolutions of the water insoluble (IS) portions. The water soluble samples cluster into two groups, A and B/C. The nitric-dissolved samples cluster into three tight groups, A, B, and C, corresponding to three CAN factories. The water insoluble portions of CAN have elemental signatures that clearly differentiate B and C stocks.
stocks that were not part of the training set. The same set of 8 elements was used to correctly match 100% of the CAN samples that were reprocessed by crushing and mixing the CAN with powdered sugar. In the case of CAN samples that were either mixed with aluminum powder or had their AN extracted with tap or bottled water, classification results were confounded due to adulterant interference depending on the factory from which the CAN originated. Remarkably, for one factory, adulterant effect on the elemental signature of its reprocessed samples was not significant, as all were correctly matched to its factory group. In summary, this work illustrates that sourcing pristine CAN, i.e., unadulterated CAN prills or powder, can potentially be done using elemental profiling and chemometrics. In terms of reprocessed CAN, the potential to match a sample to its correct factory depends on how much a factory's unique elemental profile is affected by adulterants. In both cases, a factory's elemental signature for CAN needs to be sufficiently stable or well modeled. These results support the continued development of chemical forensics capabilities for the source determination of various commercial chemical threats such as CAN. Future work will likely investigate other potential chemical attribution signatures such as those from the water-insoluble CAN components that are, at least for elemental profiles, more precise at differentiating the tested CAN stocks according to factory. Additionally, more CAN stocks that are representative of those factories studied here, plus others, have to be analyzed to reliably determine the full potential and limits of elemental profiling and chemometrics for sourcing CAN.
classification. This spread in classification performance illustrates how the elemental signatures of different CAN sources are affected differently by adulterant interference. Future work may look at other types of chemical attribution signatures that are not affected by adulterant components in reprocessed CAN samples. Based on the results described, our current hypothesis is that parts of the elemental profiles for CAN are characteristic to the facility where either the starting materials or the CAN was manufactured. For instance, the water from the nitric acid used in manufacturing the AN in CAN is likely local to where the nitric acid was made or diluted and therefore may have characteristic trace elements coming from the local water that are determined by the local geology. Any work to test our hypothesis will need nitric acid stocks used in the commercial manufacture of ammonium nitrate. 3.4. Factory discrimination using the water-insoluble portions of CAN Approximately 25% of CAN is insoluble in water and consists primarily of calcium carbonate and some additives. We therefore briefly investigated whether the water-insoluble components of CAN could distinguish CAN according to factory. To address this question, the DI water-insoluble portions of 14 CAN samples from Set 1 and Set 3 were dissolved in 5% nitric acid and analyzed by ICP-MS. Fig. 3A and B depict the HCA dendograms for the water-soluble portions and the corresponding water-insoluble portions (dissolved in 5% nitric), respectively, for 14 CAN samples from stocks A12, B98, B12, B13, and C12. The elements V, Mn, Sr, and U were used to produce the two dendograms and clearly illustrate that the water-insoluble portion of CAN differentiates the B/C factory group into separate factories, a distinction that was not possible using the water-soluble portions of CAN. The water-insoluble components of CAN may provide a more precise elemental signature to distinguish current and future CAN stocks from different manufacturing facilities.
Acknowledgements Funding for this project was provided by the Technical Support Working Group of the Combatting Terrorism Technical Support Office (CTTSO) (Contract N4175615MP50216). CTTSO support does not represent an endorsement of the contents or conclusions of this paper. We thank the Terrorist Explosive Device Analytical Center for providing some of the CAN stocks.
4. Conclusions
References
Strong evidence has been provided for the potential use of elemental profiling and chemometric analysis for the matching of pristine and reprocessed CAN fertilizers to their places of manufacture. Specifically, PLSDA using the ICP-MS determined concentrations of Na, V, Mn, Cu, Ga, Sr, Ba and U was used to match 73% of a test set of pristine CAN samples from 11 stocks to their correct source (i.e., one of 5 factory groups). The test set also included samples from factory
[1] M.D. Barbero, Remarks by Michael D. Barbero, Director JIEDDO, Homemade Explosives Threat, In: Proceedings of the Homemade Explosives Conference, Crystal City, Virginia, 2011. [2] M.D. Barbero, Remarks by Michael D. Barbero, Director JIEDDO, In: Proceedings of the International Fertilizer Association Annual Conference, Doha, Qatar, 2012. [3] G. Maxwell, Synthetic Nitrogen Products: a Practical Guide to the Products and Processes, Springer, 2004. [4] U.N.I.D. Organization, I.F.D. Center, Fertilizer Manual, Springer, Netherlands,
137
Talanta 174 (2017) 131–138
C.G. Fraga et al.
chemometrics, Anal. Chem. 88 (2016) 1827–1834. [15] A. Mitroshkov, J.V. Ryan, T. May-Lin, J.J. Neeway, Comprehensive isotopic and elemental analysis of a multi-oxide glass by multicollector ICP-MS in isotope substitution studies, J. Chromatogr. Sep. Tech. 7 (2016). [16] A.A. Scheidemann, R.B. Darling, F.J. Schumacher, A. Isakharov, Faraday cup detector array with electronic multiplexing for multichannel mass spectrometry, J. Vac. Sci. Technol. A 20 (2002) 597–604. [17] K.J. Johnson, R.E. Synovec, Pattern recognition of jet fuels: comprehensive GCXGC with ANOVA-based feature selection and principal component analysis, Chemom. Intell. Lab. Syst. 60 (2002) 225–237. [18] K.M. Pierce, J.L. Hope, K.J. Johnson, B.W. Wright, R.E. Synovec, Classification of gasoline data obtained by gas chromatography using a piecewise alignment algorithm combined with feature selection and principal component analysis, J. Chromatogr. A 1096 (2005) 101–110. [19] C.G. Fraga, K. Bronk, B.P. Dockendorff, A. Heredia-Langner, Organic chemical attribution signatures for the sourcing of a mustard agent and its starting materials, Anal. Chem. 88 (2016) 5406–5413. [20] K.R. Beebe, R.J. Pell, M.B. Seasholtz, Chemometrics: a Practical Guide, John Wiley & Sons, Inc, New York, 1998. [21] D. Ballabio, V. Consonni, Classification tools in chemistry. Part 1: linear models, PLS-DA, Anal. Methods 5 (2013) 3790–3798. [22] L.S. Ramos, Characterization of mycobacteria Species by HPLC and pattern recognition, J. Chromatogr. Sci. 32 (1994) 219–227. [23] B.K. Lavine, A. Fasasi, N. Mirjankar, M. Sandercock, Development of search prefilters for infrared library searching of clear coat paint smears, Talanta 119 (2014) 331–340.
1998. [5] V. Smil, Enriching the Earth: Fritz Haber, Carl Bosch, and the Transformation of World Food Production, MIT Press, 2004. [6] JIEDDO, JIEDDO HME Profile: Calcium Ammonium Nitrate (CAN-26), in: JIEDDO (Ed.) JIDA Information Resources, Training Aids, 〈https://www.jieddo. mil/content/docs/JIEDDO_HME_Tri-fold_v3.pdf〉. [7] B. Zygmunt, D. Buczkowski, Agriculture grade ammonium nitrate as the basic ingredient of massive explosive charges, Propellants Explos. Pyrotech. 37 (2012) 685–690. [8] J. Glauser, Chemical Economics Handbook: Ammonium Nitrate, IHS Chemical 2013. [9] E.F.M. Association, Production of ammonium nitrate and calcium ammonium nitrate, European Fertilizer Manufacturers’ Association 2000. [10] Product Specifications for CAN-27, Achema,〈http://www.achema.com/uploads/ achema%20products%20en/Fertilisers/Nitrogen%20fertilisers/Calcium %20ammonium%20nitrat%20N27/2015.05.29%20CAN%20EN.pdf〉. [11] S.J. Benson, C.J. Lennard, P. Maynard, D.M. Hill, A.S. Andrew, C. Roux, Forensic analysis of explosives using isotope ratio mass spectrometry (IRMS) — Discrimination of ammonium nitrate sources, Sci. Justice 49 (2009) 73–80. [12] H. Brust, M. Koeberg, A. van der Heijden, W. Wiarda, I. Mügler, M. Schrader, G. Vivo-Truyols, P. Schoenmakers, A. van Asten, Isotopic and elemental profiling of ammonium nitrate in forensic explosives investigations, Forensic Sci. Int. 248 (2015) 101–112. [13] D. Widory, J.-J. Minet, M. Barbe-Leborgne, Sourcing explosives: a multi-isotope apprach, Sci. Justice 49 (2009) 62–72. [14] N.S. Mirjankar, C.G. Fraga, A.J. Carman, J.J. Moran, Source attribution of cyanides using anionic impurity profiling, stable isotope ratios, trace elemental analysis and
138
Contents lists available at ScienceDirect
Talanta journal homepage: www.elsevier.com/locate/talanta
Elemental source attribution signatures for calcium ammonium nitrate (CAN) fertilizers used in homemade explosives
MARK
⁎
Carlos G. Fraga , Alexander V. Mitroshkov, Nikhil S. Mirjankar, Brian P. Dockendorff, Angie M. Melville Pacific Northwest National Laboratory, 902 Battelle Boulevard, Richland, WA 99352, USA
A R T I C L E I N F O
A BS T RAC T
Keywords: Calcium ammonium nitrate fertilizers Chemical attribution signatures Chemical forensics Inductively coupled plasma-mass spectrometry Elemental profiling Chemometrics
Calcium ammonium nitrate (CAN) is a widely available fertilizer composed of ammonium nitrate (AN) mixed with some form of calcium carbonate such as limestone or dolomite. CAN is also frequently used to make homemade explosives. The potential of using elemental profiling and chemometrics to match both pristine and reprocessed CAN fertilizers to their factories of origin for use in future forensic investigations was examined. Inductively coupled plasma-mass spectrometry (ICP-MS) was used to determine the concentrations of 64 elements in 125 samples from 11 CAN stocks from 6 different CAN factories. Using Fisher ratio and degree-ofclass-separation, the elements Na, V, Mn, Cu, Ga, Sr, Ba and U were selected for classification of the CAN samples into 5 factory groups; one group was two factories from the same fertilizer company. Partial least squares discriminant analysis (PLSDA) was used to develop a classification model which was tested on a separate set of samples. The test set included samples that were analyzed at a different time period and samples from factory stocks that were not part of the training set. For pristine CAN samples, i.e., unadulterated prills, 73% of the test samples were matched to their correct factory group with the remaining 27% undetermined using strict classification. The same PLSDA model was used to correctly match all CAN samples that were reprocessed by mixing with powdered sugar. For CAN samples that were reprocessed by mixing with aluminum or by extraction of AN with tap or bottled water, correct classification was observed for one factory group, but source matching was confounded with adulterant interference for two other factories. The elemental signatures of the water-insoluble (calcium carbonate) portions of CAN provided a greater degree of discrimination between factories than the water-soluble portions of CAN. In summary, this work illustrates the strong potential for matching unadulterated CAN fertilizer samples to their manufacturing facility using elemental profiling and chemometrics. The effectiveness of this method for source determination of reprocessed CAN is dependent on how much an adulterant alters the recovered elemental profile of CAN.
1. Introduction Calcium ammonium nitrate (CAN) is a widely available inorganic fertilizer that is frequently used by terrorists and insurgents to make homemade explosives (HMEs) for improvised explosive devices [1,2]. CAN consists of ammonium nitrate (AN), at approximately 75% (w/w), mixed with some form of calcium carbonate such as limestone or dolomite [3–5]. AN is the oxidizing component in many binary HMEs. CAN fertilizer can be reprocessed to make a HME by first mixing it with hot water to separate the soluble AN from the inert and insoluble components (e.g., calcium carbonate), followed by evaporation to
recover pure AN which is dried and crushed [6]. Alternatively, CAN fertilizer can also be crushed to a powder and used directly without extraction of the inert material [6,7]. The final reprocessing step requires mixing the extracted AN or powdered CAN with reducing agents such as aluminum powder, powdered sugar, fuel oil or a combination to produce an explosive mixture [6]. Reprocessed CAN, at one point, was used as the primary explosive in over 85% of the improvised explosive devices used against coalition forces in Afghanistan [1,2]. CAN's wide availability as a legitimate fertilizer will continue to make it and other AN based fertilizers a global HME threat for the foreseeable future. Hence efforts are underway to examine new
Abbreviations: CAN, Calcium ammonium nitrate; AN, Ammonium nitrate; HME, Homemade explosives; ICP-MS, Inductively coupled plasma-mass spectrometry; FTIR, Fouriertransform infrared spectroscopy; DCS, Degree-of-class separation; HCA, Hierarchical cluster analysis; PCA, Principal Component Analysis; PLSDA, Partial least squares discriminant analysis; DI, Deionized; RSD, Relative standard deviation ⁎ Corresponding author. E-mail address: [email protected] (C.G. Fraga). http://dx.doi.org/10.1016/j.talanta.2017.05.066 Received 25 April 2017; Received in revised form 20 May 2017; Accepted 23 May 2017 Available online 26 May 2017 0039-9140/ © 2017 Elsevier B.V. All rights reserved.
Talanta 174 (2017) 131–138
C.G. Fraga et al.
Table 1 CAN stocks and number of stock samples used in each sample set. Stock ID
Source
A12 B98 B12 B13 C11 C12 C13 D13 E13a E13b F83
Factory Factory Factory Factory Factory Factory Factory Factory Factory Factory Factory
a b
A B B B C C C D E E F
Manufacture date
Set 1 May 2013a
Set 2 Jan 2015a
Set 3 May 2015a
2012 1998 2012 2013 2011 2012 2013 2013 2013 2013 1983
3 (6b) 3 3 3 (6b) 3 – – – – – –
3 3 3 3 3 3 3 3 3 3 3
8 (6b) 4 (6b) 4 4 4 4 4 3 2 2 8 (6b)
(1st Qtr.)
(17 May) (25 Apr) (22 Mar)
time period CAN samples were prepared and analyzed by ICP-MS. number of “reprocessed” CAN samples that involved crushing CAN and mixing it with either sugar, aluminum or non-DI water.
2. Materials and methods
analytical approaches for establishing the origin of CAN and other HME precursors to support forensic investigations such as those that aim to identify and disrupt illicit HME-precursor distribution networks. CAN fertilizers, like AN fertilizers, are available as prills or granules that are typically white, grey or beige in color and 2–5 mm in diameter. CAN fertilizer is produced in various closely related formulations by mixing concentrated AN with finely ground limestone or dolomite resulting in a CAN product that is 20–28% N [8,9]. AN is produced by reacting nitric acid with ammonia. CAN is also made by reacting calcium nitrate with ammonia and carbon dioxide to yield a mixture of AN and calcium carbonate containing 20–26% N [8]. Due to its hygroscopic nature, CAN is coated with anticaking agents and additives such as gypsum, kieselguhr, and magnesium nitrate, which are applied during prilling or granulation [8]. A typical composition of CAN fertilizer consists 27% N (13.5% ammonia-N and 13.5% nitrate-N), 6–9% Ca, and 4% Mg [10], with calcium and magnesium present as carbonates (CaCO3 and MgCO3) as well as oxides (CaO and MgO) [9]. These various constituents and additives could provide unique chemical signatures specific to the manufacturing source of CAN because of bulk and trace compositional differences in CAN that are source dependent. For example, trace elements in the water and mineral additives used in making CAN are dependent on the local geology where acquired and therefore may impart trace elemental signatures into CAN that are source dependent. There are no previous published works on the source determination of CAN fertilizers. Recent papers have demonstrated the potential of linking AN fertilizers to their places of manufacture using isotopic and elemental profiling [11–13]. In particular, Brust and co-workers used linear discriminant analysis and likelihood ratios to demonstrate the discrimination of AN from one manufacturer from 16 others based on the concentrations of 32 elements and the stable isotopic ratios of oxygen and nitrogen [12]. Their work was quite convincing in part because of their large sample collection that included samples from 17 batches and two fertilizer types from one AN fertilizer manufacturer. Similar work has demonstrated the potential of sourcing commercial cyanides to their places of manufacture using isotopic, elemental and ionic profiles combined with chemometric analysis [14]. The work described here further develops the concept of sourcing commercial chemical products to their places of manufacture through the discovery and exploitation of intrinsic chemical signatures in CAN fertilizers. Specifically, the use of elemental profiling to source CAN fertilizers to a larger number of factories than previously shown for AN and cyanide is demonstrated. This includes investigating the sourcing of CAN that has been adulterated by reprocessing methods used for making HMEs. Lastly, the potential of using chemical attribution signatures from specific chemical constituents to obtain a more precise level of source discrimination is illustrated using the water-insoluble components of CAN.
2.1. CAN stocks and sample sets A total of 11 stocks of commercial CAN fertilizer (26 – 27% N) that originated from six factories were used in this study. Each stock was a bulk sample of CAN prills or granules made at a specific factory and time period. From this point forth, the term “prills” refers to either granules or true prills. Each CAN stock and corresponding source information (factory and manufacture date) were obtained either from a manufacturer, distributor, or reputable field representative. The factory sites were confirmed using a list of known CAN factories [8]. F83 was the only stock that was not matched to a specific factory, but based on its region of origin and its distinct prill size and appearance was believed to not be from any of the other factories. The identity of each stock was confirmed as CAN by FTIR analysis using a Gemini analyzer (Thermo Scientific). CAN samples from these 11 stocks were prepared and then analyzed by inductively coupled plasma-mass spectrometry (ICP-MS) in three sets. Table 1 lists the 11 CAN stocks with their corresponding source information and number of samples taken from each stock for each sample set. 2.2. Sample preparation CAN samples were prepared according to four methods: (1) CAN prills were dissolved in deionized (DI) water to produce a water-soluble portion and a water-insoluble portion, (2) water-insoluble portions were dissolved in nitric acid to produce a solution, (3) CAN prills were crushed and mixed with either aluminum powder or powdered sugar in water to produce a water-soluble portion and a water-insoluble portion, and (4) CAN prills were crushed, mixed with either tap or bottled water, and the supernatant dried to yield ammonium nitrate (AN). Methods (3) and (4) simulated CAN that is reprocessed as part of HME production. For all methods, the vessels in contact with CAN or its solutions were plastic. Details for each method are described below. For method (1), a total of 95 pristine CAN samples (those in Table 1 not denoted as “reprocessed”) were prepared by mixing 100 mg of CAN prills (typically 2–3 prills) with 10 mL of DI water (18 MΩ•cm from Nanopure, Thermo Scientific) using a 30 s vortex followed by 15 min sonication. Each sample mixture was then centrifuged for 10 min to separate the insoluble components, primarily some form of calcium carbonate. The water-soluble portion of each CAN sample was stored for ICP-MS analysis. For method (2), 20 mg of a water-insoluble portion was dissolved completely by mixing with 10 mL of 5% nitric acid (Optima, Fisher Scientific) and stored for ICP-MS analysis. This was done for a subset of Set 1 and Set 3 CAN samples. For method (3), a subgroup of 12 CAN samples from Set 1 and Set 3 132
Talanta 174 (2017) 131–138
C.G. Fraga et al.
(denoted in Table 1 as “reprocessed”) were each crushed to produce 100 mg of powdered CAN that was mixed with 10 mg of powdered sugar (C & H Pure Cane) or 10 mg of aluminum powder (Tannerite Kill Shot Target) in 10 mL of DI water. Each aqueous mixture of CAN and reducing agent (sugar or Al) was then vortexed, sonicated, and centrifuged just as the pristine CAN samples described in method (1). The water soluble portion of each CAN sample was stored for ICPMS analysis. For method (4), 200 mg of crushed CAN (individually from 18 CAN samples denoted in Table 1 as “reprocessed”) was mixed with either 10 mL of tap water (Richland, WA) or 10 mL of bottled water (Crystal Geyser Alpine Spring Water by CG Roxane, Weed, CA). Each aqueous mixture was then vortexed, sonicated, and centrifuged similar to the pristine CAN samples in method (1). Each water soluble portion was then placed into a Reacti-Therm module (Thermo Scientific) where it was heated at 40 °C under a constant stream of dry nitrogen until all the water evaporated leaving behind dried AN. Approximately 100 mg of each AN portion extracted from CAN was then mixed with 10 mL of DI water and stored for ICP-MS analysis.
(Mathworks, Natick, MA) and PLS Toolbox 7.9.2 (Eigenvector Research, Manson, WA). The data from Set 1 and Set 2 were combined for exploratory analysis to determine the optimal elements for classifying CAN samples according to source. Variable selection was performed using the Fisher-ratio method [17] and degree-of-class separation (DCS) [18]. This approach has been successfully used for obtaining chemical attribution signatures [14,19]. Clustering of sample data based on the selected variables, i.e., concentrations of elements, was assessed using hierarchical cluster analysis (HCA) and principal component analysis (PCA) [20]. Supervised classification using the selected elements was then performed by partial least squares discriminant analysis (PLSDA) [21] using samples from Set 1 and Set 2 as the training set. PLSDA uses PLS regression for directly identifying variations in data space that discriminate classes with less noise compared to linear discriminant analysis. This technique has worked well for similar data sets from our previous work [14]. The PLSDA classification was assessed on a test set consisting of samples from Set 3 and samples from Set 1 and Set 2 whose stocks were not represented in the training set. For HCA, PCA, and PLSDA, the data were preprocessed by unit-area normalization; i.e., for each sample, the concentration of each selected element was divided by the sum of all selected elemental concentrations.
2.3. ICP-MS analysis A NU Plasma high resolution multi-collector ICP-MS (NU Instruments) was used to determine the concentrations of 64 elements: Li, Be, Na, Al, K, Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Cu, Zn, Ga, Ge, As, Se, Rb, Sr, Y, Zr, Nb, Mo, Ru, Rh, Pd, Ag, Cd, In, Sn, Sb, Te, Cs, Ba, La, Ce, Pr, Nd, Sm, Eu, Gd, Tb, Dy, Ho, Er, Tm, Yb, Lu, Hf, Ta, W, Re, Os, Ir, Pt, Au, Hg, Tl, Pb, Bi, Th, and U. The ICP-MS instrument included 11 Faraday cups and 3 ion counters used as detectors. Aqueous samples were introduced using an ASX-112FR autosampler connected to a QuickWash accessory and an Aridus II Desolvating Nebulizer (all from CETAC). The aqueous CAN samples, i.e., their water-soluble fractions in DI water or water-insoluble fractions in 5% nitric acid, were diluted to 1/100 (v/v) with 2% nitric acid (Optima, Fisher Scientific). Representative CAN samples were first screened using Elemental Mode analysis, with only the central Faraday cup Ax, in order to adjust ICP-MS parameters. For quantitative analysis, a scan rate of 1000 ms/amu and a dwell time of 2000 ms were used. Mass scale calibration was performed using two different aqueous standards: ICPMS-IS-1 and ICP-MS-68A (both from High Purity Standards). The ICPMS-68A standard consisted of 68 elements and was diluted to produce two solutions of 8 and 10 parts-per-billion (ppb) used as external standards; the 10-ppb standard was used for quantitative analysis of the Set 1 and Set 2 samples and the 8-ppm standard for the Set 3 samples. Each respective standard and a method blank were analyzed after every three sample analyses. The standard and blank analyzed closest in time to an analyzed sample were used for quantification by one-point calibration using the respective elemental ICP-MS signal intensities (in volts) as shown by Eq. (1):
⎛ Isample − Iblank ⎞ Csample = ⎜ ⎟ × Cstandard × D ⎝ Istandard − Iblank ⎠
3. Results and discussion 3.1. Elemental signatures for the factory discrimination of CAN Robust elemental signatures that point to a specific source (e.g., CAN factory) are not affected by experimental variability and withinsource variability. In this study, experimental variability was addressed by performing sample preparations and ICP-MS analysis on the same CAN stocks at two different time periods, in 2013 (Set 1) and 2015 (Set 2), in order to capture changes in instrument performance, solvent stocks (e.g., DI water), sample-preparation personnel, and environmental conditions. In terms of within-source variability, the sample preparation and analysis of multiple stocks from the same CAN factory made at different time periods provided a means to address the
(1)
where Csample and Cstandard are the concentrations of the element in the sample and standard respectively, and Cstandard is either 8 or 10 depending on whether the 8-ppb or 10-ppb standard was used. Isample, Istandard, and Iblank are the signal intensities for sample, standard, and blank, respectively. D is the dilution factor with a value of 100. Elements whose sample intensity (Isample) was less than the corresponding signal in the blank (Iblank) were assigned a zero for concentration. This quantitative approach [15] proved to be simple and effective for the elemental profiling of each CAN aqueous sample in part because of the five orders of signal linearity inherent to Faraday-cup ICP-MS instruments [16]. Fig. 1. Dendrogram using the normalized (unit area) concentrations of V, Mn, Sr, and U for Set-1 and Set-2 CAN samples. The CAN samples are the water soluble portions of 11 CAN stocks from 6 factories (A – F). The 48 samples cluster into four groups, A, B/C, D/ F, and E, with sample B13-2-2 being the only outlier. The sample label format is: (stock ID)-(sample set)-(sample replicate).
2.4. Chemometric analysis Chemometric analysis was performed using Matlab R2014a 133
Talanta 174 (2017) 131–138
C.G. Fraga et al.
compositional variability in the manufacture of CAN from changes, for example, in feedstock and operating conditions. Variable selection and cluster analysis were then used on the ICP-MS data to select and determine those elements in CAN whose concentrations resulted in CAN stocks grouping according to their factories regardless of when analyzed or when manufactured. Variable selection was performed using Fisher ratio and DCS calculations on the pristine (not reprocessed) CAN elemental concentrations of samples from Set 1 and Set 2. The elements V, Mn, Sr, and U were found to differentiate the water-soluble portions of CAN samples into four factory groups: A, B/C, D/F, and E. Fig. 1 is a HCA dendrogram (based on Euclidean distance) that depicts the differentiation of the CAN samples into the four factory groups. As seen, CAN samples from Set 1 and Set 2 cluster with other samples from the same group, thereby demonstrating the robustness of the four selected elements towards experimental differences between the two sample sets. In addition, the clustering of stock samples in Fig. 1 with other stock samples from the same factory group (B/C and E) illustrates the robustness of the four selected elements in terms of source variability. For data sets that are difficult to classify in a single step, a step-wise classification approach can be applied [22,23]. Such a multi-level classification approach is particularly useful for discriminating classes that under an initial set of variables are separated into groups that include more than one class. Variable selection was further investigated to differentiate the B/C group and D/F group into separate factories. In the case of the D/F group, the elements Na, Cu, Ga, and Ba were found to separate the D and F CAN samples into two clusters, one for each factory (D and F). The separation of the B/C factory group into two distinct HCA groups with variable selection was not achievable. Factories B and C are different, but they belong to the same manufacturing company and are geographically near one another. It is therefore possible that they use similar reagent stocks and processes in the making of the water-soluble components (i.e., primarily AN) used in their CAN products, accounting for the similar elemental concentrations in the water-soluble fractions. Fig. 2A shows a PCA scores plot of all 64 element concentration profiles compared to a PCA scores plot of profiles for the selected eight elements Na, V, Mn, Cu, Ga, Sr, Ba, and U (Fig. 2B). It is evident from the figure that the eight-element profiles can be used for discrimination of the CAN samples into five factory groups. Although a univariate approach was utilized for feature selection, agreement between the results obtained from the same data using HCA and PCA provides a strong justification for dividing the Set 1 and Set 2 CAN samples into five factory groups: A, B/C, D, E, and F. With the future acquisition of CAN stocks from other factories, a new set of elements (currently out of 64) may be needed to differentiate CAN samples into factory groups. Presumably, the new set would include most of the eight elements selected here because of their ability to discriminate the current CAN samples into factory groups. Table 2 lists the mean concentration and relative standard deviation (RSD) for each of the eight selected elements in each CAN stock analyzed in Set 1 and Set 2. The table illustrates some of the differences in elemental concentrations that help differentiate the CAN stocks into five factory groups. For example, the mean U concentration in stock A12 from factory A is 2–35 times greater than in the stocks from factories B, C, D, and E. Typically, elemental concentrations are compared by normalizing to the sample mass (here, amount of CAN analyzed). However, this is not always straightforward for some field samples such as those that are crushed CAN mixed with a reducing agent. To address this issue, we normalized the elemental concentrations by using unit-area normalization. Given that the ICP-MS elemental signals for the 10-ppb multielement standard had a daily RSD range of 11–21%, the RSD values in Table 2 that are greater than 21% are likely caused from within-stock variability and not instrumental variability. In the end, the measured variabilities in concentration of the selected elements in each CAN
Fig. 2. PCA scores plot of 48 CAN samples from Set-1 and Set-2 using (A) 64 element concentration profile and (B) selected 8 element profile of Na, V, Mn, Cu, Ga, Sr, Ba, and U. With all 64 elements, the CAN samples do not cluster according to factory groups, however using the selected 8 elements, the CAN samples cluster into five factory groups: A, B/C, D, E, and F. The data was normalized to unit area and autoscaled prior for PCA.
stock are small relative to the differences in elemental profiles between the factories (except for B and C) as indicated by the clustering of CAN samples according to factory group (Fig. 1 and Fig. 2B). This notion was assessed through the classification of a test set of samples according to factory group, as described in the following sections. 3.2. Factory classification using the water-soluble portions of pristine CAN A single PLSDA model was created for five factory groups (A, B/C, D, E, and F) using the ICP-MS concentrations of the eight elements Na, V, Mn, Cu, Ga, Sr, Ba, and U in the water-soluble portions of the pristine (not reprocessed) Set 1 and Set 2 CAN samples (see Table 1). Only 6 out of the 11 CAN stocks (A12, B12, C11, D13, E13a, and F83) were used for training the PLSDA model. Elemental concentrations were normalized to unit area and autoscaled prior to model creation. The PLSDA model, based on 27 CAN samples, consisted of 4 latent variables that were determined through venetian blind cross-validation with 10 splits and 1 sample per split. The PLSDA model captured 94% of the elemental data variance. The created PLSDA model was tested on the elemental concentrations obtained from the water-soluble portions of 68 pristine CAN samples. The test set included all pristine samples from Set 3, and the 134
Talanta 174 (2017) 131–138
C.G. Fraga et al.
Table 2 Mean concentration in ppb and RSD (in parentheses) for each of the eight selected elements in the DI water-soluble portionsa of the CAN Set 1 and Set 2 stock samples. Stock ID
No. of samples
Na
V
Mn
Cu
Ga
Sr
Ba
U
A12
6
550 (18%)
0.67 (24%)
14 (4.9%)
4.6 (78%)
0.50 (18%)
21 (8.1%)
2.4 (5.7%)
0.38 (22%)
B98
6
550 (34%)
0.17b (44%)
60 (33%)
26 (100%)
0.64 (39%)
13 (32%)
3.2 (52%)
0.013b (75%)
B12
6
690 (14%)
0.16 (67%)
80 (4.5%)
23 (140%)
0.78 (25%)
29 (5.5%)
4.2 (9.1%)
0.048 (42%)
B13
6
710 (36%)
0.17 (38%)
76 (25%)
33 (73%)
0.97 (11%)
47 (39%)
4.7 (14%)
0.056 (46%)
C11
6
170 (31%)
0.099c (61%)
35 (26%)
14 (79%)
0.67 (8.7%)
13 (31%)
2.8 (12%)
0.036b (18%)
C12
3
140 (22%)
0.13 (17%)
34 (5.0%)
26 (59%)
6.8 (76%)
7.3 (11%)
30 (29%)
0.029 (70%)
C13
3
280 (54%)
0.14 (30%)
27 (13%)
24 (34%)
3.4 (150%)
7.3 (14%)
12 (150%)
0.040 (6.7%)
D13
3
190 (13%)
1.8 (20%)
24 (10%)
11 (6.4%)
4.4 (15%)
150 (15%)
15 (8.9%)
0.19 (33%)
E13a
3
710 (13%)
1.7 (6.9%)
0.40 (7.7%)
4.5 (9.0%)
0.76 (13%)
190 (7.8%)
3.4 (5.1%)
0.14 (9.6%)
E13b
3
780 (6.3%)
1.8 (3.7%)
1.1 (10%)
6.2 (19%)
1.8 (13%)
300 (8.0%)
7.4 (4.8%)
0.17 (8.8%)
F83
3
590 (14%)
0.87 (22%)
20 (8.9%)
5.5 (18%)
11 (13%)
170 (7.4%)
41 (6.1%)
0.27 (18%)
a b c
The supernatant from mixing 100 mg of CAN with 10 mL of DI water. One sample had a concentration of zero and was not included. Two samples had concentrations of zero and were not included.
magnitude. In probable classification, all samples are classified so there are no unclassified samples. Several interesting observations and conclusions can be made from Table 3 regarding the ability to match the CAN samples to their sources. First, strict classification correctly matched 73% of the CAN samples to their correct class with no misclassifications; 27% were unclassified. This means that a sample that is reported as unclassified may have originated from one of the stocks that was actually modeled; therefore, if an unknown sample is listed as unclassified, it does not necessarily mean it came from a new source that was not modeled. Second, correct matching was observed for samples from factories A, B
Set 1 and Set 2 samples of 5 CAN stocks (B98, B13, C12, C13, and E13b) that were not included in training the PLSDA model. These samples were treated as unknowns. Table 3 summarizes their classification results. The term “strict” in Table 3 refers to strict classification where unknown samples are classified to a specific class if the probability is greater than a specified threshold (probability value of 0.5) for only one class. If no class has a probability greater than the threshold or if more than one class has probability exceeding the threshold, then the sample is assigned “no class” and is considered “unclassified”. In “probable” classification, unknown samples are matched to a class that has the highest probability regardless of the
Table 3 ICP-MS elemental PLSDAa classification results for test set of water-soluble pristine CAN samples. Samples
A12: 8 samples B98b: 10 samples B12: 4 samples B13b: 10 samples C11: 4 samples C12b: 7 samples C13b: 7 samples D13: 3 samples E13a: 2 samples E13bb: 5 samples F83: 8 samples 68 samples
Class
A B/C B/C B/C B/C B/C B/C D E E F 5 classes
Strict
Probable
Correct
Misclassified
Unclassified
Correct
Misclassified
7 (87.5%) 10 (100%) 3 (75%) 10 (100%) 1 (25%) 0 (0%) 4 (57.1%) 2 (66.7%) 2 (100%) 3 (60%) 8 (100%) 50 (73.5%)
0 0 0 0 0 0 0 0 0 0 0 0
1 (12.5%) 0 (0%) 1 (25%) 0 (0%) 3 (75%) 7 (100%) 3 (42.9%) 1 (33.3%) 0 (0%) 2 (40%) 0 (0%) 18 (26.5%)
8 (100%) 10 (100%) 4 (100%) 10 (100%) 3 (75%) 3 (42.9%) 6 (85.7%) 2 (66.7%) 2 (100%) 4 (80%) 8 (100%) 60 (88.2%)
0 0 0 0 1 4 1 1 0 1 0 8
(0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%) (0%)
(0%) (0%) (0%) (0%) (25%) (57.1%) (14.3%) (33.3%) (0%) (20%) (0%) (11.8%)
a PLSDA model created for 5 classes using ICP-MS elemental concentrations of Na, V, Mn, Cu, Ga, Sr, Ba and U from pristine CAN Set-1 and Set-2 water-soluble samples: A12 (6 samples), B12 (6 samples), C11 (6 samples), D13 (3 samples), E13a (3 samples), F83 (3 samples). b Training set did not include samples from these stocks.
135
Talanta 174 (2017) 131–138
C.G. Fraga et al.
Table 4 ICP-MS Elemental PLSDAa Classification Results for Reprocessed CAN Samples form Set 1 and Set 3. Samples
Class
Strict
Probable
Correct
Misclassified
Unclassified
Correct
Misclassified
A12 + sugar: 2 samples (Set 3)
A
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
B98b + sugar: 2 samples (Set 3)
B/C
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
F83 + sugar: 2 samples (set 3)
F
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
A12 + Al: 2 samples (Set 3)
A
0 (0%)
1 (50%)
1 (50%)
0 (0%)
2 (100%)
B98b + Al: 2 samples (Set 3)
B/C
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
F83 + Al: 2 samples (Set 3)
F
1 (50%)
0 (0%)
1 (50%)
2 (100%)
0 (0%)
A12 tap-water extracted AN: 5 samples (Set 1 & Set 3)
A
4 (80%)
1 (20%)
0 (0%)
4 (80%)
1 (20%)
A12 bottle water extracted AN: 3 samples (Set 1)
A
1 (33.3%)
0 (0%)
2 (66.7%)
3 (100%)
0 (0%)
B98b tap-water extracted AN: 2 samples (Set 3)
B/C
2 (100%)
0 (0%)
0 (0%)
2 (100%)
0 (0%)
B13 tap-water extracted AN: 3 samples (Set 1)
B/C
3 (100%)
0 (0%)
0 (0%)
3 (100%)
0 (0%)
B13 bottle-water extracted AN: 3 samples (Set 1)
B/C
3 (100%)
0 (0%)
0 (0%)
3 (100%)
0 (0%)
F83 tap water extracted AN: 2 samples (Set 3) 30 samples
F
0 (0%)
2 (100%)
0 (0%)
0 (0%)
2 (100%)
3 classes
22 (73.3%)
4 (13.3%)
4 (13.3%)
25 (83.3%)
5 (16.7%)
a b
Same 5-classes PLSDA model used on the pristine CAN samples in Table 3 based on concentrations of Na, V, Mn, Cu, Ga, Sr, Ba and U was utilized. Training set did not include samples from these stocks.
correctly classified. This result is not surprising given that household sugar should have a zero to low metal content and therefore adulterant interference is unlikely. On the other hand, aluminum powder has metal impurities that likely caused half of the samples from stock A12 to be misclassified (half were unclassified) and half of those from stock F83 to be unclassified (half were correctly classified) by strict classification. Surprisingly, all of the B98 CAN samples mixed with aluminum powder were correctly classified. Apparently, the 8-element profile for B98 CAN was not adversely affected by the metal impurities in aluminum. The tap and bottled waters used in this study also have metal constituents. Surprisingly, all the B98 and B13 CAN samples whose AN was extracted using tap or bottled water and then dried and analyzed by ICP-MS were correctly classified. Just as for the aluminum powder, the 8-element profiles for B98 and B13 were not adversely affected by the metals in the bottled or tap waters. Although a few samples from factory B were tested, it seems so far that the 8-element profile from factory B is the most robust in terms of withstanding adulterant interference. At a lesser level, 4 out 5 of the samples from stock A12 whose AN was extracted using tap water were correctly classified while 1 out 5 were misclassified under both strict and probable classification. For the bottled water samples, 1 out 3 of the A12 samples were correctly classified with zero misclassifications under strict classification; under probable classification, all of the A12 samples were correctly classified. For the F83 stock, all two reprocessed samples involving tap water were misclassified under both strict and probable
and F with high classification success rates; ~ 88% for stock A12 and 100% for stocks B98, B13, and F83. This suggests that similar stocks from these three factories can be likely matched to their designated source. Third, probable classification correctly matched 88% of all samples to their correct class while 12% were incorrectly matched. The probable approach is best utilized when the unknown samples are suspected to belong to one of the modeled classes. 3.3. Factory classification using the water-soluble portions of reprocessed CAN In the previous section, the “unknown” pristine CAN samples represent field samples of CAN, either whole prills or crushed prills, that have not been adulterated by mixing with a reducing agent or environmental media. Another type of field sample is one in which the CAN has been reprocessed by either crushing and mixing the CAN with a reducing agent or crushing the CAN and mixing it with water to separate the AN from CAN and then drying the aqueous supernatant. Some of the Set 1 and Set 3 CAN samples were reprocessed as mentioned. These “reprocessed” samples (see Table 1) were 30 CAN samples that were either mixed with sugar or aluminum, or had their AN extracted from CAN using tap or bottled water and then prepared for ICP-MS analysis. Table 4 summarizes the classification results for the reprocessed samples using their ICP-MS elemental data and the same PLSDA model used for the pristine CAN samples. As shown in Table 4, all of the CAN samples mixed with sugar were 136
Talanta 174 (2017) 131–138
C.G. Fraga et al.
Fig. 3. Dendrograms using concentrations of V, Mn, Sr, and U for (A) 14 samples that are the water soluble portions of 5 CAN stocks and (B) their corresponding samples from the 5% nitric-acid dissolutions of the water insoluble (IS) portions. The water soluble samples cluster into two groups, A and B/C. The nitric-dissolved samples cluster into three tight groups, A, B, and C, corresponding to three CAN factories. The water insoluble portions of CAN have elemental signatures that clearly differentiate B and C stocks.
stocks that were not part of the training set. The same set of 8 elements was used to correctly match 100% of the CAN samples that were reprocessed by crushing and mixing the CAN with powdered sugar. In the case of CAN samples that were either mixed with aluminum powder or had their AN extracted with tap or bottled water, classification results were confounded due to adulterant interference depending on the factory from which the CAN originated. Remarkably, for one factory, adulterant effect on the elemental signature of its reprocessed samples was not significant, as all were correctly matched to its factory group. In summary, this work illustrates that sourcing pristine CAN, i.e., unadulterated CAN prills or powder, can potentially be done using elemental profiling and chemometrics. In terms of reprocessed CAN, the potential to match a sample to its correct factory depends on how much a factory's unique elemental profile is affected by adulterants. In both cases, a factory's elemental signature for CAN needs to be sufficiently stable or well modeled. These results support the continued development of chemical forensics capabilities for the source determination of various commercial chemical threats such as CAN. Future work will likely investigate other potential chemical attribution signatures such as those from the water-insoluble CAN components that are, at least for elemental profiles, more precise at differentiating the tested CAN stocks according to factory. Additionally, more CAN stocks that are representative of those factories studied here, plus others, have to be analyzed to reliably determine the full potential and limits of elemental profiling and chemometrics for sourcing CAN.
classification. This spread in classification performance illustrates how the elemental signatures of different CAN sources are affected differently by adulterant interference. Future work may look at other types of chemical attribution signatures that are not affected by adulterant components in reprocessed CAN samples. Based on the results described, our current hypothesis is that parts of the elemental profiles for CAN are characteristic to the facility where either the starting materials or the CAN was manufactured. For instance, the water from the nitric acid used in manufacturing the AN in CAN is likely local to where the nitric acid was made or diluted and therefore may have characteristic trace elements coming from the local water that are determined by the local geology. Any work to test our hypothesis will need nitric acid stocks used in the commercial manufacture of ammonium nitrate. 3.4. Factory discrimination using the water-insoluble portions of CAN Approximately 25% of CAN is insoluble in water and consists primarily of calcium carbonate and some additives. We therefore briefly investigated whether the water-insoluble components of CAN could distinguish CAN according to factory. To address this question, the DI water-insoluble portions of 14 CAN samples from Set 1 and Set 3 were dissolved in 5% nitric acid and analyzed by ICP-MS. Fig. 3A and B depict the HCA dendograms for the water-soluble portions and the corresponding water-insoluble portions (dissolved in 5% nitric), respectively, for 14 CAN samples from stocks A12, B98, B12, B13, and C12. The elements V, Mn, Sr, and U were used to produce the two dendograms and clearly illustrate that the water-insoluble portion of CAN differentiates the B/C factory group into separate factories, a distinction that was not possible using the water-soluble portions of CAN. The water-insoluble components of CAN may provide a more precise elemental signature to distinguish current and future CAN stocks from different manufacturing facilities.
Acknowledgements Funding for this project was provided by the Technical Support Working Group of the Combatting Terrorism Technical Support Office (CTTSO) (Contract N4175615MP50216). CTTSO support does not represent an endorsement of the contents or conclusions of this paper. We thank the Terrorist Explosive Device Analytical Center for providing some of the CAN stocks.
4. Conclusions
References
Strong evidence has been provided for the potential use of elemental profiling and chemometric analysis for the matching of pristine and reprocessed CAN fertilizers to their places of manufacture. Specifically, PLSDA using the ICP-MS determined concentrations of Na, V, Mn, Cu, Ga, Sr, Ba and U was used to match 73% of a test set of pristine CAN samples from 11 stocks to their correct source (i.e., one of 5 factory groups). The test set also included samples from factory
[1] M.D. Barbero, Remarks by Michael D. Barbero, Director JIEDDO, Homemade Explosives Threat, In: Proceedings of the Homemade Explosives Conference, Crystal City, Virginia, 2011. [2] M.D. Barbero, Remarks by Michael D. Barbero, Director JIEDDO, In: Proceedings of the International Fertilizer Association Annual Conference, Doha, Qatar, 2012. [3] G. Maxwell, Synthetic Nitrogen Products: a Practical Guide to the Products and Processes, Springer, 2004. [4] U.N.I.D. Organization, I.F.D. Center, Fertilizer Manual, Springer, Netherlands,
137
Talanta 174 (2017) 131–138
C.G. Fraga et al.
chemometrics, Anal. Chem. 88 (2016) 1827–1834. [15] A. Mitroshkov, J.V. Ryan, T. May-Lin, J.J. Neeway, Comprehensive isotopic and elemental analysis of a multi-oxide glass by multicollector ICP-MS in isotope substitution studies, J. Chromatogr. Sep. Tech. 7 (2016). [16] A.A. Scheidemann, R.B. Darling, F.J. Schumacher, A. Isakharov, Faraday cup detector array with electronic multiplexing for multichannel mass spectrometry, J. Vac. Sci. Technol. A 20 (2002) 597–604. [17] K.J. Johnson, R.E. Synovec, Pattern recognition of jet fuels: comprehensive GCXGC with ANOVA-based feature selection and principal component analysis, Chemom. Intell. Lab. Syst. 60 (2002) 225–237. [18] K.M. Pierce, J.L. Hope, K.J. Johnson, B.W. Wright, R.E. Synovec, Classification of gasoline data obtained by gas chromatography using a piecewise alignment algorithm combined with feature selection and principal component analysis, J. Chromatogr. A 1096 (2005) 101–110. [19] C.G. Fraga, K. Bronk, B.P. Dockendorff, A. Heredia-Langner, Organic chemical attribution signatures for the sourcing of a mustard agent and its starting materials, Anal. Chem. 88 (2016) 5406–5413. [20] K.R. Beebe, R.J. Pell, M.B. Seasholtz, Chemometrics: a Practical Guide, John Wiley & Sons, Inc, New York, 1998. [21] D. Ballabio, V. Consonni, Classification tools in chemistry. Part 1: linear models, PLS-DA, Anal. Methods 5 (2013) 3790–3798. [22] L.S. Ramos, Characterization of mycobacteria Species by HPLC and pattern recognition, J. Chromatogr. Sci. 32 (1994) 219–227. [23] B.K. Lavine, A. Fasasi, N. Mirjankar, M. Sandercock, Development of search prefilters for infrared library searching of clear coat paint smears, Talanta 119 (2014) 331–340.
1998. [5] V. Smil, Enriching the Earth: Fritz Haber, Carl Bosch, and the Transformation of World Food Production, MIT Press, 2004. [6] JIEDDO, JIEDDO HME Profile: Calcium Ammonium Nitrate (CAN-26), in: JIEDDO (Ed.) JIDA Information Resources, Training Aids, 〈https://www.jieddo. mil/content/docs/JIEDDO_HME_Tri-fold_v3.pdf〉. [7] B. Zygmunt, D. Buczkowski, Agriculture grade ammonium nitrate as the basic ingredient of massive explosive charges, Propellants Explos. Pyrotech. 37 (2012) 685–690. [8] J. Glauser, Chemical Economics Handbook: Ammonium Nitrate, IHS Chemical 2013. [9] E.F.M. Association, Production of ammonium nitrate and calcium ammonium nitrate, European Fertilizer Manufacturers’ Association 2000. [10] Product Specifications for CAN-27, Achema,〈http://www.achema.com/uploads/ achema%20products%20en/Fertilisers/Nitrogen%20fertilisers/Calcium %20ammonium%20nitrat%20N27/2015.05.29%20CAN%20EN.pdf〉. [11] S.J. Benson, C.J. Lennard, P. Maynard, D.M. Hill, A.S. Andrew, C. Roux, Forensic analysis of explosives using isotope ratio mass spectrometry (IRMS) — Discrimination of ammonium nitrate sources, Sci. Justice 49 (2009) 73–80. [12] H. Brust, M. Koeberg, A. van der Heijden, W. Wiarda, I. Mügler, M. Schrader, G. Vivo-Truyols, P. Schoenmakers, A. van Asten, Isotopic and elemental profiling of ammonium nitrate in forensic explosives investigations, Forensic Sci. Int. 248 (2015) 101–112. [13] D. Widory, J.-J. Minet, M. Barbe-Leborgne, Sourcing explosives: a multi-isotope apprach, Sci. Justice 49 (2009) 62–72. [14] N.S. Mirjankar, C.G. Fraga, A.J. Carman, J.J. Moran, Source attribution of cyanides using anionic impurity profiling, stable isotope ratios, trace elemental analysis and
138