In a previous column, we discussed a number to techniques to make the mixture analysis problem easier, including purification, spectral subtraction, and library searching (1). At that time, I promised to write a future column with more details on library searching and spectral subtraction. A search of my records shows that I never wrote that column, and I apologize for that. This column, therefore, focuses on library searching, and a future column will focus on spectral subtraction.

Spectral comparisons, where a known reference spectrum is compared to an unknown spectrum to assist in its identification, is an important tool in infra- red spectral interpretation (2). I am old enough that I began interpreting spectra before personal computers were invented. Back in those Jurassic days, to identify an unknown one had to compare the sample spectrum to paper copies of known spectra kept in hundreds of green three-ring binders. It could take hours of poring over these dusty tomes to find the correct library match. The company back then that published these spectra was called Sadtler Chemical (3).

Fortunately, computerized library searching now exists. In this technique, an unknown spectrum is compared to a collection of known infrared spectra kept in a digital infrared spectral library. Revealing my age once again, the first Fourier transform infrared (FT-IR) system I ever used that had computerized library searching capabilities took 15 min to complete one library search! I should point out that this system had a computer, manufactured by the FT-IR maker, with a whopping 64 kb of memory. Here in 2021, I think my toaster has more computing capability than my FT-IR computer did back then. Today, of course, library searches of thousands of spectra take place in a flash.

Regardless of the computer, a library search is performed by mathematically comparing your unknown spectrum to each of the spectra in a library. If there are 1000 spectra in the library, then 1000 comparisons will be performed. The library search software uses what is called a search algorithm to generate a number describing how similar, or different, the two spectra are. This number is called a Hit Quality Index (or HQI, for short), and it is discussed later on in the column.

Where Can I Buy Infrared Spectral Libraries?

To perform library searching, the first thing you have to do is obtain some libraries. I normally don’t call out individual companies in this column, but there are only a handful of companies selling infrared spectral libraries, so I will name several here as a convenience to my readers. I apologize if I leave anyone out.

The company with perhaps the largest collection of infrared spectral libraries is what was Sadtler, became Bio-Rad, and is now part of Wiley (3–5). According to their website, Wiley has available 264,000 spectra you can search against (5). Access to these libraries is normally done via a leasing arrangement.

Another vendor that sells infrared spectral libraries is Sigma-Aldrich (6). As far as I can tell, their entire collection consists of about 100,000 spectra. A couple of smaller companies that sell infrared spectral libraries include ST Japan (7) and Fiveash Data Management (8). A Google search will also turn up some free libraries on the internet. My experience with free libraries is that they are generally small and specific to a narrow range of samples, and sometimes the data are not correct because they have not been vetted properly. As a result, be very careful when using free libraries off the internet. The libraries you purchase may come on a compact disk, be copied to your computer’s storage device, or installed on your company’s computer network.

Ultimately, the best source of infrared spectral libraries is you. Most FT-IR software packages allow you to build your own libraries. I strongly encourage this, because only you have access to the sample types typical of your work. You can even build multiple libraries of your own of different types. For example, you could build your own polymer, gas phase, and inorganics library, and search appropriate unknowns against them. Every time you come across a new sample for which you have a good identification, add it to an appropriate library. Spectra you add to your libraries become de facto references. For this reason, please make sure your data is of high quality, you are certain of the identification, and the data matches the instrumental resolution of the other spectra in the library (9).

What About the Instrumental Resolution of Library Spectra? 

Instrumental resolution is a measure of how well a spectrometer distinguishes spectral features from each other (9). In FT-IR, instrumental resolution is measured in cm

. For example, a spectrum measured at 8 cm

 resolution can resolve features that are 8 cm-1 or further apart (9). When library searching is performed, the unknown and library spectra must be measured at the same instrumental resolution. This is why when you purchase infrared spectral libraries, they come in different resolutions. For example, the same library might be available in 8 cm

 resolutions. When purchasing a library, you must make sure that the libraries’ resolution matches that of the samples you will be measuring.

After starting up your library searching software package, you will need to choose which libraries to search against. It may be tempting to try the shotgun approach and search your sample spectrum against all the libraries you have, but this may end up being a waste of time. Think about the nature of your sample. For example, a spectrum of a polymer searched against a gas phase spectral library will never produce good results. On the other hand, a polymer spectrum searched against a large organics library might yield good results if the repeat unit of the polymer is similar to any of the molecules in the library. Judicious choice then of what libraries you search against is important.

Another thing to think about before performing a library search is to select what wavenumber region or regions to use. Typically, the default region in use by your search software will be the entire spectrum, lets say 4000 to 400 cm

However, the software will typically allow you to choose spectral regions to include or exclude from the search, and this sometimes makes sense. For example, if your sample spectrum has unwanted water vapor or CO

 peaks, you should exclude those regions from your search, since they are not part of the sample and their presence can throw off your search results. Additionally, if a sample is a mixture and you have identified a major component, excluding its peaks from a search might make the search more sensitive to the identity of minor components.

Some years ago, while I was consulting for a forensics laboratory, they had a prob- lem distinguishing between the methyl (CH

-) variants of a controlled substance using library searching. They were using the sample’s full spectrum in the search. Considering the chemical differences between the two variants, I thought it made sense to look at the C-H stretching region, which is from 3200 to 2800 cm

 as we have discussed in previous columns (2). A comparison of the spectra of the two variants in this region is seen in Figure 1.

Note that the spectra are similar but not identical. By narrowing the wavenumber region used in the search to that seen in Figure 1, a library search was then able to distinguish between these two variants and the problem was solved. This a real and powerful example of how judiciously choosing the spectral regions used in a search can improve search results.

A search algorithm is the mathematical calculation used to compare two spectra to each other and generate a HQI. A discussion of these algorithms is beyond the scope of this article, but you can find more info in the help function of your library search software or here (9).

Most library searching software packages present you with a choice of search algorithms. The differences between these algorithms is that some emphasize peak position, some emphasize peak height, and some strike a balance between the two. Algorithms with the word “derivative” in them tend to emphasize peak positions, algorithms with the words “absolute” or “absolute value” tend to emphasize peak heights, and two algorithms that strike a balance are euclidean distance and correlation (9).

From my own experience, I find that balanced algorithms, such as correlation and euclidean distance, are good for general search purposes. However, if a given search algorithm does not produce satisfactory results, it is all right to change the search algorithm and see if it improves your results. It is generally easy to switch search algorithms in library search software packages, so this is a parameter you can experiment with.

Recall that the number of spectral comparisons performed equals the number of spectra in your library. Manually sorting through all these matches would be a challenge, but fortunately the search software program organizes the hits and presents the best ones in the form of a search report. An example of a search report is seen in Figure 2. The “unknown” spectrum was of pure polystyrene, and the correlation search algorithm was used.

The table at the bottom of Figure 2 lists the identity of the eight best hits along with the HQIs and the library where the example spectrum was found. In the software used in this example, a 0 to 100 HQI scale was used, where 100 is a perfect match, and 0 is the worst possible match. Also note that the search report shows the sample spectrum and the spectra of the best matches. The best match, with an HQI of 93.18, is of pure polystyrene, which is good and expected. The second best match, with an HQI of 85.90, is of a mixture of polystyrene and another polymer. This makes sense, since impure polystyrene should be a worse match than pure polystyrene.

HQIs are useful indicators of spectral similarity, but they are not perfect. Random similarity in the noise and artifacts between two spectra can give search results more similar than reality. Conversely, random differences in noise and artifacts between two spectra can make them poorer matches than reality. This is why the HQI by itself should never be used identify a sample. I have seen inexperienced people look at a search report, incorrectly assume the first match in the table identifies the sample, and suffer career-ending consequences.

Search algorithms are just mathematical formulas; they do not understand spectroscopy or chemistry and make mistakes. I cannot emphasize enough the importance of always visually comparing the sample spectrum to that of the best library matches. The HQI is not a measure of the probability of having found the right answer; it is not a measure of purity. It is strictly used to organize the matches for a given search. The search will always produce a result whether there are any good matches or not, and the fact that there is a result does not mean you have identified the sample.

Ultimately, it is your job as the user of the search software, and not the computer’s (since you have a brain and eyeballs), to make the final conclusion as to the identity of a sample based on a library search. The library search narrows things down for you; it is your job to make the final decision as to whether two spectra are a match or not.

How then do we interpret the HQI? At minimum, the range of the HQI can be useful. Assuming a 0 to 100 HQI scale as discussed above, matches between 80 and 100 are excellent and may be good enough to obtain an identification. HQIs between 50 and 80 are good, but typically not good enough for a complete identification. However, you can often times obtain information about the functional groups present from matches of this quality. Lastly, a HQI less than 50 is a poor result, and is generally not useful.

You could panic, but that is not an appropriate response for a scientist, unless said scientist is being chased by a grizzly bear (more on that in a later column). Ultimately, bad library search results are not necessarily your fault—you are at the mercy of the libraries you have access to. If, by chance, your unknown looks nothing like any of the spectra in your libraries, there is not much to be done about it. This is an argument then for having access to a large number of library spectra so you can maximize the probability of finding a good match. Additionally, remember that you can experiment with wavenumber regions and search algorithms to try and improve your search results.

Infrared spectral library searching is useful in identifying unknowns and in mixture analysis. It works by mathematically comparing your sample spectrum to a collection of spectra kept in a digital library. HQI measures how similar or different two spectra are. However, the HQI only ranks the quality of the matches for a given search, and should not be used by itself to make an identification. Visual comparison of the sample and library spectra should be used in combination with the HQI achieve a sample identification. Experimenting with the search algorithm and wavenumber region may improve search results.

(4) Wiley Acquires Bio-Rad’s Informatics Spectroscopy Software and Spectral Databases | Business Wire

(5) KnowItAll IR Spectral Database Collection - Wiley Science Solutions

(6) FT-IR Spectral Libraries Nicolet/Aldrich Condensed Phase Library, edition 2 | Sigma-Aldrich

(7) Spectra Databases - katja-hm (jimdofree.com).

(8) FDM FTIR and Raman Libraries (www.fdmspectra.com).

Fundamentals of Fourier Transform Infrared Spectroscopy 2nd Edition

, PhD, is the founder and CEO of Big Sur Scientific, a maker of portable mid-infrared cannabis analyzers. He has over 30 years experience as an industrial infrared spectroscopist, has published numerous peer-reviewed papers, and has written three books on spectroscopy. As a trainer, he has helped thousands of people around the world improve their infrared analyses. In addition to writing for Spectroscopy, Dr. Smith writes a regular column for its sister publication Cannabis Science and Technology and sits on its editorial board. He earned his PhD in physical chemistry from Dartmouth College. He can be reached at: 

Articles

IR Spectral Interpretation Workshop

One of the biggest practical limitations of infrared spectroscopy is its difficulty in analyzing mixtures. Infrared library searching can help, but it must be done right.

Library Searching

Scopolamine and promethazine can be used as a substitute for heroin. It is relatively fast and simple to use surface-enhanced Raman spectroscopy (SERS) for the detection and monitoring of drug usage. Silver nanosol is an enhanced substrate that is commonly used in the SERS technique; the Raman signal of the sample is amplified by the substrate. However, for the detection of mixtures, the competitive adsorption of different components relative to the substrate will result in difficult qualitative identification of spectra. By taking advantage of this defect, the adsorbed components of a mixture were separated from silver nanoparticles by centrifugation, thus changing the mixing ratio of the two components in the mixture. As a result, the proportion of the components with strong adsorption in the mixture is relatively reduced. Thus, the components with weak adsorption can be displayed in the spectrum of the mixed solution.

Heroin is one of the most widely used narcotics in the world, and it is a key target to be monitored because of its addictive quality and the difficulty in quitting the addiction (1–6). After long-term use, heroin will cause serious damage to the nervous system of the human body, and will also adversely affect judgment, possibly causing harm to the addict or society (4,7,8). Some drug rehabilitation institutions use less harmful drug-like substances as substitutes to reduce the suffering of drug addicts and damage to their nervous system (9–14). For example, a mixture of buprenorphine, scopolamine, and promethazine, referred to as BSP, has been used because its toxicity is much lower than that of heroin (15). However, long-term use of BSP will still cause great damage to the human body (9,10). Because the price of BSP is low and it is easily accessible, it has become an alternative drug for some drug users. Therefore, it is necessary to engage in the detection and monitoring of this new type of drug abuse.

Surface-enhanced Raman spectroscopy (SERS) is a combination of Raman spectroscopy and nanotechnology, and it can enhance the Raman signal by 7–10 orders of magnitude when the tested sample is near or adsorbed on a substrate at the nanometer scale (16–19). The Raman spectrum of each sample is specific, and can be qualitatively identified as the fingerprint of the sample.

There are two popular viewpoints about the enhancement mechanism of SERS: One viewpoint is electromagnetic enhancement, and the other viewpoint is chemical enhancement (17,20). Electromagnetic enhancement generally refers to the enhancement of Raman signals by the strong electromagnetic field generated by the local surface plasmon (LSPR) excited by the rough surface of metal materials or the surface of metal nanoparticles (21). The chemical mechanism refers to the contribution of Raman scattering independent of the electromagnetic environment (such as plasma polaritons). The increase in SERS signal is usually attributed to electron transfer between the adsorbed molecule and the substrate. Current research results show that the two enhancement mechanisms generally play a role at the same time, and that signal amplification effect of electromagnetic enhancement is greater than that of chemical enhancement. Based on these enhancement phenomena, the SERS signal can be improved to some extent through substrate design or gap control (22–25).

Surface-enhanced spectroscopy as a drug detection technology has become increasingly mature (26–29). Because of its characteristics, it is more suitable for the detection of drugs in solution than many other analytical techniques. Compared with other detection technologies, the advantages of SERS lie in the simple operation of the instrument, the uncomplicated pretreatment of the sample before detection, the fast detection time, and the convenience of the portable instrument for the in situ detection of samples (30).

For SERS detection of mixtures, there are two problems associated with direct detection: One is the overlapping of peaks, and the other is competitive adsorption (31). Overlap of spectral peaks is most likely to occur when the molecular configurations of the components of the mixture are relatively similar.

There are some differences in the molecular configuration of the two components we tested. They were distinguished from the spectrum of the mixture as described in our previous studies (15). From the spectral observation, there is no chemical reaction of the two components after mixing; that is, no new substance is formed. The current study focuses on eliminating the interference of competitive adsorption in practical detection.

A BWS415-785H portable Raman spectrometer (B&W Tek, Inc.) was used to obtain the spectral data. The maximum working power of the Raman spectrometer was 0.3 W. The laser wavelength was 785 nm. After spectrum acquisition, the spectrum is smoothed and the baseline was corrected by the instrument software.

The morphology of nanoparticles was characterized by a high-resolution field emission scanning electron microscope (Hitachi Company). The dynamic light scattering test was measured by a PALS/90P particle size analyzer (Brookhaven Instruments Co., Ltd.). A Cary 4000 UV-vis spectrophotometer was also used (Agilent Technology Co., Ltd.).

Potassium iodide (KI), silver nitrate (AgNO

) were purchased from Sinophram Chemical Reagent Co., Ltd. Scopolamine hydrobromide powder (1G) and promethazine hydrochloride powder (5G) were purchased from Dalian Meilun Biotechnology Co., Ltd.

The solutions of scopolamine and promethazine were obtained by diluting scopolamine hydrobromide and promethazine hydrochloride, respectively, with deionized water, are used to obtain solutions with different concentrations. Each solution is mixed with the substrate at the same time.

The preparation of the silver nano-substrate is as follows: Soluble silver salt nitrate (AgNO

) was used as an oxidant, and ascorbic acid was used as the reducing agent to prevent the aggregation of silver nanoparticles. Sodium citrate was used as the stable solution. Their concentrations in the initial solution were 0.168 mg/mL, 0.105 mg/mL, and 0.873 mg/mL. First, the water bath condition was prepared at 30 

C. Then, the three concentrations (soluble silver salt nitrate, ascorbic acid, and sodium citrate) were placed in the bath proportional to each other, then the solutions (sols) were stirred continuously for 15 min with a magnetic mixture and both sol and nanosol are used. Then, the temperature of the water bath was continually raised to the boiling state and stayed at that temperature level for 2 h until the color turned grayish-green. Then, the sol was refrigerated after cooling. Before using it, the sol needed to be centrifuged and purified to remove the smaller particles and improve the monodispersibility of the particles. Ultrasonic oscillation dispersion was performed before the test was used.

Figure 1 shows the SEM characterization of the silver sol substrate. The main process used is described in the literature (32). The silver nanoparticles are basically in a spherical state, and there is no mixing of other shapes of the substrate. Figure 2a shows the UV absorption spectrum, and the scanning range is from 300 nm to 800 nm. The absorption resonance peak is approximately 439 nm, which is consistent with the absorption peak of the silver sol. Figure 2b shows the statistical weight distribution of the particle size of the silver sol measured by dynamic light scattering (DLS) technology. Among them, 60 nm particles account for the largest proportion in the system, with its statistical weight being the total number of particles set to 100% and the statistical weight of other particle sizes compared with its value as the ordinate value of other particle sizes. It can be seen that several main particle sizes are distributed at 50–80 nm.

 M) scopolamine were used to compare the mixture, as shown in Figures 3a and 3c. It should be noted that because the signal of scopolamine is weak, the intensity of scopolamine is exaggerated by about three times. When the molar concentration ratio of the two mixtures was 1:1 (1.5 x 10

 M), a weak Raman signal of scopolamine was detected at 1002 cm

. As shown in Figure 3b, the characteristic peaks in the mixture mainly belong to promethazine. However, in the previous study, when the concentration of the two components was the same, the spectrum was mainly the spectrum of scopolamine (15). The difference is that in our current study, the mixture was in contact with the silver nanoparticles at the same time; in the previous study, scopolamine was first added to the silver nanosol (nanosolution), and then the promethazine was subsequently added. From the trend of these two spectra, we can deduce that there is a competitive adsorption phenomenon when they are mixed simultaneously. Promethazine will mask the characteristic peak of scopolamine at the same concentration, which is not conducive to identifying the spectrum of the mixture. What is more troublesome is that in the actual mixing ratio, the concentration of scopolamine is much lower than that of promethazine. To enhance the signal of scopolamine, we added potassium iodide into the mixture system to increase the strength of scopolamine.

To improve the signal of scopolamine in the mixed solution in Figure 3, we added potassium iodide (1 M). As shown in Figure 4, the signal of scopolamine at 1002 cm

 increases with the addition of electrolytes, but it greatly decreases compared with the detection of scopolamine at the same concentration alone. Moreover, the concentration ratio of the silver sol we used is further increased, and the signal should be stronger in theory than what is shown in Figure 4. The characteristic peaks of promethazine in the spectrum of the mixed solution are basically all displayed. Promethazine of the same concentration was added to the electrolytes for independent detection and is compared with the mixture spectrum, as shown in Figure 5. The position and intensity of the peak of the mixture spectrum are almost the same as that of promethazine alone. Table I compares the intensity of the mixture spectrum with that of promethazine when it is separately detected in several main characteristic peaks. The results show that the signal of promethazine was dominant even though both of them were enhanced. This further shows that the enhanced signal of the two is mainly caused by the molecules adsorbing on the silver nanoparticles. The addition of electrolytes did not change the adsorption ratio of the two but further reduced the distance between nanoparticles and enhanced their signal.

In the mixture with the same concentration, the signal of scopolamine is weak, which is not conducive to spectral recognition. The more difficult problem is that the concentration of promethazine with strong adsorption is much higher than that of scopolamine in the actual ratio. It can predict the difficulty of actual detection. To verify this assumption, we mixed them according to the actual ratio, in which the mass concentration of scopolamine is 10 ppm (3.3 x 10

 M), and the doping of promethazine is 1670 ppm (5.9 x 10

 M). Through SERS detection, as shown in Figure 6a, scopolamine in the actual ratio cannot be identified in the Raman spectrum. Only the characteristic peaks of promethazine can be identified, which is consistent with the expected results.

According to the experimental verification, the nanoparticles in the silver sol are relatively easy to separate from the solution, as shown in Figure 8. For example, most of the nanoparticles will be located at the bottom of the tube after high-speed centrifugation. Figure 8a shows the original sol solution, and the supernatant after centrifugation is shown in Figure 8b. We used this method in combination with the “defects” of competitive adsorption to separate the components of the mixture in solution. After the mixed solution was added to the silver sol and then centrifuged, the silver sol traveled to the bottom of the centrifuge tube. Some samples adsorbed on the silver nanoparticles were separated from the mixed solution.

The adsorption of promethazine is stronger than that of scopolamine. Thus, promethazine should be the main component separated. The proportion of mixture in the remaining supernatant changed: The proportion of promethazine decreased, and the relative concentration of scopolamine increased. After repeating this several times, the characteristic peak of scopolamine can be seen in the mixed solution.

First, we centrifuged the mixed solution at a high speed for 20 min and removed the supernatant for detection. The results as shown in Figure 6b indicate that there was no signal of scopolamine. The signal intensity of promethazine greatly changed, indicating that the concentration of promethazine changed accordingly. However, the signal intensity of promethazine in the supernatant increased because the relationship between concentration and signal intensity in the higher concentration range is not necessarily consistent. This is mainly because of the intensity saturation effect in SERS detection.

After the first centrifugation, the supernatant was again mixed with concentrated silver sol and prepared for the second centrifugation. After the second centrifugation, the supernatant was removed for testing, and the results are shown in Figure 6c. The 1034 cm

 signal intensity for promethazine changed, indicating that the concentration also changed, but the main peak at 1002 cm

 for scopolamine still did not appear. On this basis, we continued to centrifuge the supernatant mixture of the previous step, and the supernatant after centrifugation was detected, as shown in Figure 6d. The signal of scopolamine was finally detected, and the signal strength of promethazine was much weaker. At the concentration of the mixture according to the actual use ratio, it was finally detected that both existed in the spectrum.

In the mixed solution, according to the actual ratio, we selected 1002 cm

 so we can clearly see the change of the spectrum of scopolamine in the process, which is shown in Figure 6. Figure 7 displays the results in a column chart. Although the first three times have weak intensity, they are below the detection limit so they can be regarded as the absence of scopolamine. The fourth time is the success of the actual detection of scopolamine.

We used simple experimental conditions to identify the spectrum of a solution mixture and utilized the disadvantageous SERS detection conditions to create a means of separating mixtures. This solves the problem that the strong adsorption component in the mixture can inhibit the weak adsorption component in the Raman spectrum. It should be noted that in the process of continuous centrifugation, the mixed solution is diluted to a certain extent. Although the relative concentration ratio of scopolamine is increasing, the concentration of scopolamine itself is decreasing, and therefore, the number of centrifugations needs to be properly controlled. The amount of silver sol used for centrifugation also requires optimization. Compared with other methods of separating mixtures, the procedure described herein is simple, and it is not only a feasible new method for the SERS detection of solution mixtures, but also for a rapid screening of new alternative drugs.

This work was supported by the National Natural Science Foundation of China( Grant No. 31871873), Inner Mongolia University for the Nationalities (Grant No. NMDYB18006), and the Inner Mongolia Autonomous Region Natural Science Foundation of China (Grant No. 2018LH08055).

(2) J. Valentin, T. Mateo, M. Mariana, S. Virginie, S. Anne-Marie, L. Denis, S. Anais, L. Virginie, B. Fabrice, A. Serge, and R. Gerald, 

(5) K. Julia, W. Sophia, R. Francois, S. Norbert, B. Matthias, and W. Elisa, 

(6) B. Valentina, C. Ciro, M. Angelo, G.I.L. Francesco, P. Matteo, and M. Icro, 

(7) G. Yong-Long, Z. Yang, C. Jiang-Peng, W. Sheng-Bing, C. Xing-Hui, Z. Yan-Chun, et al., 

(8) L. Yixin, X. Baijuan, L. Rongrong, Y. Dan, W. Yanlin and L. Wenmei, 

(9) Z.Denghui, Z. Dengke, D. Huiqiong, Z. Xuehui, Q Jinping, S. Hongxian, X. Xiaojun, W. Xuyi and H. W, Chin. 

(10) Z. Xuhui, W. Xuyi, L. Jun and H. Wei, 

(11) S. Wenwen, W. Qing, Z. Jianbin, P. Wenkai, Z. Jiawen, Y. Weiting, H. Qianyu, C. Deniz, and Z. Wenhua, 

(13) C.P. Fang, S.C. Wang, H.H. Tsou, R.H. Chung, Y.T. Hsu, S.C. Liu, H.W. Kuo, T.H. Liu, A.C.H. Chen, and Y.L. Liu, 

14) W.W. Shen, Q. Wang, J.B. Zhang, W.K. Ping, J.W. Zhang, W.T. Ye, Q.Y. Hu, D. Cerci, and W.H. Zhou, 

(15) L. Bao, S. Han, X.Y. Sha, H. Zhao, Y.P. Liu, D.Y. Lin, and W. Hasi, 

(16) J. Langer, D.J. de Aberasturi, J. Aizpurua, R. A. Alvarez-Puebla, B. Auguie, J.J. Baumberg, G.C. Bazan, S.E.J. Bell, A. Boisen, et al., 

(17) R. Pilot, R. Signorini, C. Durante, L. Orian, M. Bhamidipati, and L. Fabris, 

(19) J. Faulds, R.E. Littleford, D. Graham, G. Dent, and W.E. Smith, 

(20) C. Caro, P. Quaresma, E. Pereira, et al., 

(21) J. Langer, D.J. de Aberasturi, J. Aizpurua, et al., 

(23) C. Carlos, J.S. Maria, F. Victorino, C. Alejandro, Z. Paula and G. Francisco, 

(24) K. Kneipp, Y. Wang, H. Kneipp, et al., 

(25) A.C. Sara, H. Szushen, R.G. Benito, et al., 

(26) R.A. Sulk, R.C. Corcoran and K.T. Carron, 

(27) K.J. Si, P.Z. Guo, Q.Q. Shi. and W.L. Cheng, 

(28) X. Lin, W.L.J. Hasi, X.T. Lou, S. Han, D.Y. Lin and Z.W. Lu, 

(29) S.Z. Weng, M.Q. Qiu, R.L. Dong, F. Wang, L.S. Huang, D.Y. Zhang and J.L. Zhao, 

Spectroc. Acta Pt. A-Molec, Biomolec. Spectr.

(30) H. Zhao, W. Hasi, L. Bao, Y.P. Liu, S. Han and D.Y. Lin, 

(32) Y.Q. Qin, X.H. Ji, J. Jing, H. Liu and H.L. Wu, 

 is with the National Laboratory of Science and Technology at the Harbin Institute of Technology in Harbin, China. 

 is with the National Laboratory of Science and Technology at the Harbin Institute of Technology in Harbin, China, and the College of Physics and Electronic Information at the Inner Mongolia Universities for Nationalities. 

 is with the Affiliated Hospital at Inner Mongolia University for the Nationalities. Direct correspondence to Wuliji Hasi at 

Peer-Reviewed Research

SERS can be used for the detection and monitoring of drugs as pure compounds and mixtures. A demonstration of a sample preparation method used to detect components with weak substrate adsorption in the spectrum of a mixed solution is shown.

Analysis of a Mixture Solution Using Silver Nanoparticles Based on Surface-Enhanced Raman Spectroscopy (SERS)

Data mining and fusion of Raman and mid-infrared (mid-IR) spectra was studied to improve the identification ability for geographical origins of rice. Relative standard deviation (RSD) analysis can predict whether there are outlier Raman spectra. Hierarchical clustering analysis (HCA) can find out the potential outlier data, and then RSD analysis can finally determine the outlier data. The recognition accuracy of the model built by eliminating the outlier data was higher than that of the model using all the data. The identification accuracy of the data fusion model was 97.8%, 4.5% higher than that of the Raman and mid-IR models. The model was further applied to identify the geographical origins of 10 japonica rice varieties, with an accuracy of 96.7%. A combination of data mining and data fusion can enhance the discrimination ability for the geographical origin of rice using a combination of Raman and mid-IR spectroscopy.

Rice is one of the most important cereal crops. It is rich in protein, fat, carbohydrates, and other nutrients (1,2). Its quality is influenced by many factors, such as genetics, growing conditions, and processing (3). With living standards improving, geographical indication products are becoming more popular among consumers because of their better quality. Meanwhile, the phenomenon of confusion or adulteration appears, bringing unfair economic benefits to the fraudsters and destroying the credibility of consumers to producers (4). Therefore, it is necessary to develop a reliable tool to identify the geographical origin of rice for consumers, honest producers, retailers, and governments.

Sensory identification (5), biological identification (6,7), chemical identification (stable isotope [8,9]), trace elements (10,11), volatile components (12–14), and other technologies have been widely used to identify the geographical origin of rice. Among them, cheap and fast spectral technologies, combined with pattern recognition methods, are becoming more popular, and are generally used in the analysis of multicomponent samples. The literature reports the application of Raman spectroscopy in identifying the geographical origin of rice. For example, Hwang and others reported the identification of Korean and imported rice with the recognition accuracy of 98% (15). Kim and co-authors identified rice that came from China and Korea with the recognition accuracy of 96% (16). Feng and others identified rice that originated from different provinces in China with the highest recognition accuracy of 100% (17,18). Most of the rice samples used in these studies are shelled rice grains, so only the information of components outside the rice grains are collected. It is well known that the weight percentages of carbohydrates, protein, fat, and water in rice grains are about 70–80%, 7–8%, 1–2%, and 11–12%, respectively, and the distribution of the components in rice grains is inhomogeneous (19). Besides, the rice processing technology, such as polishing, will cause differences on the surface of rice. Therefore, the Raman spectra obtained are insufficient to reflect the information of the whole composition of rice. To analyze the rice grains accurately, they need to be pulverized and sieved to obtain the rice flour sample as evenly as possible (20). In addition, the existing literature focused on the identification of geographical origins and varieties of rice and other inhomogeneous products mainly obtain the spectral information based on the average of two or three measurements (21). The possibility of outlier data caused by large fluctuations is not taken into account, and the accuracy of the data by taking the average is unknown. To make the obtained spectra reflect the sample information more comprehensively, it is necessary to collect enough spectra and eliminate the outliers.

As the growing environment changes, so does the content or structure of components in rice, thus the position and intensity of the absorption peaks of infrared spectra are different, and these differences can be used to reflect the regional characteristics of rice products (22). Xia and colleagues used near-infrared spectroscopy (NIR) and Fisher’s linear discriminant to discriminate Xiangshui rice with the accuracy of 100% (23). In the discrimination of non-Xiangshui rice originated from nine provinces, 92.3% of the modeled samples were classified correctly, and 90.9% of the validated samples were correctly discriminated. Compared with the NIR spectrum, the absorption peak of the mid-IR spectrum is usually sharper, with good resolution and high peak value. Nowadays, the updated mid-IR spectrometer in the market does not require sample preparation, greatly saving testing time, and is an ideal fast detection technique. However, the application of mid-IR spectroscopy in identifying the geographical origin of rice is rare.

The above methods for identification of the geographical origin of rice are mostly based on a single technique that can not reflect the composition information of rice sufficiently. For example, artificial falsification of specific chemical indicators will cause serious interference to chemical identification methods, so the reliability of identification results is poor. Moreover, the existing identification methods are commonly aimed at rice samples grown in different countries or different provinces in the same country. The geographical areas are far apart from each other, making identification relatively easy. There have been many reports on the application of data fusion technology in food quality detection (24,25). Sun and others established a model for identifying official rhubarb using NIR spectroscopy and mid-IR spectroscopy, using data fusion strategies to improve the classification model and allow the correct classification of all the samples (26). Hohmann and associates can effectively distinguish a total of 205 tomato samples of nine varieties from seven farms with an accuracy rate of 95 to 100% by combining proton nuclear magnetic resonance, mid-IR spectroscopy, and isotope ratio mass spectrometry (27). According to the literature, data fusion of Raman and infrared spectra for geographical identification of rice is rarely reported.

Therefore, in this work, rice products grown in four geographical regions (Xijiang, Panjin, Jiansanjiang, and Wuchang) in northeast China were taken as representative examples to develop an excellent identification model. Rice samples were refined, crushed, and screened to obtain relatively uniform rice flour samples. Raman and mid-IR spectra were obtained. Outlier Raman spectra were eliminated. A model was established by support vector machine based on data fusion of Raman and infrared spectra aimed at improving the identification accuracy and reliability for geographical origins of rice. Finally, the promotion and application value of the established method and model were studied in 10 kinds of rice.

For this experiment, 10 types of geographical indication rice products were collected from different areas of China, as seen in Figure S1. Fangzheng rice, Jiansanjiang rice, Wuchang rice, and Xiangshui rice were all grown in Heilongjiang province. Yanbian rice and Xijiang rice were both grown in Jilin province. Xinghua rice and Heheng rice were grown in Jiangsu province. Panjin rice was grown in Liaoning province. Yutai rice was grown in Shandong province. All of the rice samples were cultivated in 2017 and 2018, and collected in different planting areas to ensure they are representative of the samples. Xijiang, Panjin, Jiansanjiang, Wuchang, Xiangshui, Yanbian, Fangzheng, Heheng, Xinghua, and Yutai rice were collected from 18, 30, 21, 21, 30, 15, 33, 21, 21, and 33 positions, respectively. Two samples were collected in each position, each sample took about 2 kg of rice, and all the samples were japonica rice with specific varieties.

Rice milling process mainly used paddy huller (NA12345) and rice mill (NA-JCB) which were both produced by Kemai Instrument Co., Ltd. A vertical grinder (15B, Baling Electric Appliance Co., Ltd.) was used to crush rice samples, and the diameter of stainless steel sieve in it was 0.6 mm. Raman spectra were recorded using a portable laser Raman spectrometer (Prott-ezRaman-d3, Enwave Optronics), samples were simply transferred on a quartz glass with 2 mm thick. Mid-IR spectra were recorded using an attenuated total reflectance (ATR) Fourier transform infrared spectrometer (Nicolet IS-10, Thermo Fisher Scientific).

Rice samples were processed according to the requirements of the first grade japonica rice in GB/T 1354-2018. Then, 20 g of each grain sample was slowly added to the grinder within half a minute and crushed for 2 min to ensure complete crushing. The rice flour obtained was screened by 100 and 140 mesh sieves successively. Rice flour with the size of 100–140 mesh was stored in a freezer and balanced to room temperature in a dryer before spectral analysis.

The parameters for Raman spectrum acquisition were as follows. The excitation wavelength of laser was 785 nm, the laser power was 450 mW, the temperature of the charge-coupled detector (CCD) detector was -85 °C, the spectrometer operated from 250 to 2339 cm

, and the number of scans was three with each having an accumulation time of 4 s. Five replicated spectra were acquired at different positions for each sample. The total test time of a sample was about 2 min, including sample loading, testing (five times), spectrum saving, and desktop cleaning.

The parameters for the mid-IR spectrum acquisition were as follows: The spectrometer operated from 525 to 4000 cm

, and each sample was scanned 32 times. An appropriate amount of rice flour was covered on the testing window and was compacted for test subsequently. All the samples were analyzed in triplicate. The total test time of a sample was about 2.5 min.

To remove interfering and irrelevant information from the original spectral data, pretreatment should be done before data analysis. In this work, Raman spectra and mid-IR spectra were denoised by a wden wavelet function (28), which eliminated the effect of scattering by multiplicative scatter correction (29), and normalized by the mapminmax function (30) successively. As Raman measurement was relatively interfered by the environment and instrument, relative standard deviation (RSD) analysis and hierarchical clustering analysis (HCA) (31,32) were comprehensively applied to eliminate outliers. Then, the average Raman spectrum and mid-IR spectrum for each sample were taken as the representative spectra. Finally, a support vector machine (SVM) (33) was used to establish the model, with the radial basis function (RBF) (34) being used. In addition, Gamma and C parameters were optimized by a grid search technique within the region of 10

. The data of one sample of each sampling point was used as the training set, and the data of the another sample was used as the prediction set. All of the data preprocessing and model construction were based on Matlab 2019b.

Taking Xijiang, Panjin, Jiansanjiang, and Wuchang rice as examples that originated from three neighboring provinces in northeast China, the spectra of sample 1 in position 1 were pretreated and shown in Figure 1. The peaks of Raman spectra mainly located in the region of 251–1500 cm

, especially located at 492, 878, 952, 1094, 1126, 1139, 1273, 1351, 1390, and 1474 cm

. Through RSD analysis, it was found that the bands with high stability of mid-IR spectra located at 578–4000 cm

, and the peaks mainly occurred at 859, 928, 997, 1080, 1150, 1640, 2930, and 3290 cm

. Therefore, the data of these regions were adopted for subsequent analysis. As seen in Figure 1, the spectra of the four kinds of rice from different producing areas were highly similar which were difficult to be recognized by naked eyes.

Geographical Identification of Rice by Raman Spectroscopy

, taking Xijiang rice as an example, RSD values of the five Raman spectra of each sample are shown in Table I. Because of the fluctuation of instrument and environment, inhomogeneity of rice flour, and some man-made factors, the RSD values generally fluctuated below 14%. Among them, the RSD value of some rice samples were so large that there might be outlier data.

The HCA of five Raman spectra of each sample was performed to find out the possible outlier spectrum intuitively. Euclidean distance (35), standard Euclidean distance, city block distance, and the congruence coefficient (cosine) (36) were used as distance measures to quantify the similarity between spectra. Linkage methods, such as average linkage (37), single linkage, and complete linkage, were applied to construct the clustering tree (38). Taking sample 1 of Xijiang rice in position 2 for example, the values of cophenetic correlation coefficient for different methods were calculated and shown in Table SI. The best created clustering tree using seuclidean distance measure and average linkage method was shown in Figure 2a. Spectra 1 and 2 had high similarity and were clustered together while spectra 3–5 clustered together. The distinction between spectra 1 and 2 and spectra 3–5 was very obvious. However, it was still difficult to judge the outlier data.

The HCA was further carried out on the ten spectra of Xijiang rice in position 2, and the results were shown in Figure 2b. Spectra 1 and 2 were significantly different from the other eight spectra, indicating that they were the potential outliers. After eliminating spectra 1 and 2, RSD value of the remaining three Raman spectra of sample 1 in position 2 was 13.9%, which was within the fluctuation range (below 14%), confirming that spectra 1 and 2 were the outlier data. Spectra 3–8 in Figure 2b were mixed together indicating that the difference between the two rice samples in the same position was small. The above research results showed that RSD analysis can predict whether outliers exist, HCA analysis can find out the potential outliers, and the outliers can be finally identified and verified by RSD analysis. Combination of RSD analysis and HCA provides an accurate and reliable method to eliminate the outliers.

The same analysis of Xijiang rice in the remaining 17 positions was processed, and the outlier data found are shown in Table I. It was found that outlier data exist in all the rice samples which RSD values were over 14%, and the RSD values after eliminating the outlier data were all less than 14%, indicating that the existence of outlier data can be preliminarily judged by the fluctuation range of the RSD value. The 15 outlier spectra in Table I should be eliminated for subsequent modeling. Furthermore, after analyzing the Raman spectra of Panjin, Jiansanjiang, and Wuchang rice, it was found that there were seven, four, and one outlier spectra, respectively (Tables SI–SIV).

Average of Raman spectra after eliminating outlier data were taken as the representative spectra for each sample, values of the Raman absorbance were taken as independent variables, and the geographical origin was used as the dependent variable. Thus, matrix X (180 × 1250) and matrix Y (180 × 1) were formed. The number of samples was 180, and 1250 was the number of independent variables, along with one dependent variable. The recognition accuracy of the model based on matrix X was 93.3%, and the total recognition time was 54.9 s. Another prediction model was established by using the data without eliminating the outlier spectra, the recognition rate was 91.1%, and the recognition time was 55.1 s.

Taking into account the relevant variables will influence the efficiency of the model, and the principal component analysis (PCA) can extract the main information of the data, so PCA of matrix X was carried out before modeling, and the results were shown in Table II. The accumulated contribution rate of the first 179 principal components (PCs) was 100.0%, demonstrating that the 179 PCs can represent the overall information of matrix X. The recognition accuracy of the model using the 179 PCs was consistent with the result of matrix X, but its value was not the largest in Table II, indicating that there are some useless information in the first 179 PCs. When the numbers of PCs were 13, 15, 16, 18–21, and 24, the recognition accuracy was the highest (95.6%), and the total recognition time was 2.4–3.0 s, which were both better than those of the model built with matrix X. The misidentified samples were shown in Figure 3a. When matrix X was used, there were six wrongly classified samples. Three samples of Xijiang rice were classified as Panjin rice, three samples of Panjin rice were classified as Jiansanjiang rice. When the numbers of PCs were 13, 15, 16, 18–21, and 24, four samples of Panjin rice were wrongly classified as Jiansanjiang rice. It could be seen that the results varied greatly using different data, and the recognition accuracy of the model built by eliminating the outlier data was always higher than that of the model using all the data (Table II).

Geographical Identification by Mid-IR Spectroscopy

Mid-IR spectra of 180 samples of four kinds of rice were averaged as the representative spectra for each sample, values of the transmittance of mid-infrared spectra were taken as independent variables, and the geographical origin was used as dependent variable. Thus, matrix Z (180 × 7100) and matrix Y (180 × 1) were formed, with 7100 being the number of independent variables. The recognition accuracy of the model based on matrix Z was 93.3%, and the total recognition time was 307.4 s. PCA was then carried out on matrix Z, the results of the identification model using different PCs were shown in Table III. With the increase of the number of PCs, total recognition time increased when the number of PCs was between 13 and 15 and identification accuracy was 94.4%, which was the largest in Table III and much better than the results using matrix Z. When the identification accuracy was 94.4%, five samples were wrongly identified as can be seen in Figure 3a. Distribution of rice samples in the space formed by the first 3 PCs was shown as Figure 3b. Panjin rice was close to Xijiang and Jiansanjiang rice, some samples of Xijiang and Jiansanjiang rice were crossed together, and a few samples of Jiansanjiang and Wuchang rice overlapped. These samples were easy to be misjudged during identification, which explained why the recognition accuracy was only 71.1% when the number of PCs was three.

Transitions between vibrational and rotational levels of chemical bonds or functional groups absorb infrared light, so the infrared spectrum can reflect the information of functional groups or chemical bonds contained in molecules. Besides, infrared absorption takes place only when there is a change of dipole moment. Raman spectrometer analyses the scattered light with different frequency from the incident light to obtain the information of molecular structure. Different from infrared spectrum, both polar molecules and nonpolar molecules can produce Raman spectrum. Therefore, if the two spectra is integrated, the information representing the composition of rice can be more comprehensive. According to the above experimental results, the misclassified samples corresponding to the best identification accuracy of Raman and mid-IR spectroscopy were entirely different, so data fusion of the two spectra can play a complementary role.

Low-level fusion of data was conducted first. Matrix X (180 × 1250) and matrix Z (180 × 7100) were fused into matrix M (180 × 8350) according to the method shown in Figure S2 (taking sample 1 of Xijiang rice in position 1 as an example). It was found that the recognition accuracy of the model using matrix M was 97.8% and the recognition time was 387.3 s. Data fusion by this method had achieved ideal recognition accuracy, but the recognition took a long time.

To overcome the low recognition efficiency, mid-level data fusion was used. According to the results in Table II and Table III, 13, 15, 16, 18– 21, and 24 PCs corresponding to the highest identification accuracy of Raman spectroscopy were selected respectively, and new data matrices were constructed by combined with 13 and 15 PCs corresponding to the highest identification accuracy of mid-IR spectroscopy. The newly established classification models had the highest recognition accuracy of 97.8% (Table SV), which was 4.5% higher than that of Raman and infrared spectroscopy. One sample of Xijiang rice was wrongly identified as Wuchang rice while one sample of Wuchang rice was wrongly identified as Jiansanjiang rice.

Promotion and Application of Data Mining and Data Fusion Technology 

The promotion and application of data mining and data fusion technology were further studied. Xiangshui rice, Yanbian rice, Fangzheng rice, Heheng rice, Xinghua rice, Yutai rice, and the above studied four kinds of rice were identified together. The matrix of Raman spectra data was 486 × 1250, the identification accuracy was 91.8%, and the identification time was 430.9 s. The matrix of mid-IR spectra data was 486 × 7100, the identification accuracy was 91.8%, and the identification time was 2497.7 s, the model was time-consuming and its identification accuracy was not ideal. To improve the recognition accuracy and efficiency of the model, data mining (Tables SVI–SXI) and mid-level data fusion method was adopted. When the number of Raman PCs was 40 and the number of infrared PCs was 35, the identification accuracy of the ten kinds of rice was 96.7%, and the identification time was 40.1 s. The results were relatively ideal and consistent with expectations.

Taking Xijiang rice, Panjin rice, Jiansanjiang rice, and Wuchang rice as examples, this paper explored the influence of data fusion of Raman spectra and mid-IR spectra on the geographical identification model. First of all, rice samples were processed, crushed, and screened, and rice flour sized between 100–140 mesh were obtained. Second, Raman spectra and infrared spectra of each sample were collected at 5 measurement positions and in triplicate respectively. The collected spectra were pretreated by denoising, multivariate scattering correction and normalization successively. RSD analysis and HCA were applied to detect of outliers from Raman data of each sample. Then, the Raman spectra and mid-IR spectra of each sample were averaged respectively. Finally, the geographical identification models were established by support vector machine using Raman data, mid-IR data and the fusion of the two data. The results showed that RSD analysis can predict whether there are outliers, HCA analysis can find out the potential outlier data, and then RSD analysis can finally determine the outlier data, the comprehensive application of RSD analysis and HCA provides an accurate and reliable method to eliminate the outliers. In addition, data fusion realized complementarity of Raman and mid-IR spectra information, and the accuracy of the established model was 97.8%, which was 4.5% higher than that of Raman data and mid-IR data. The data fusion method explored in this paper provided more accurate and reliable data for identifying geographical origins of rice quickly. Besides, the identification accuracy of 10 kinds of rice with close geographical location was 96.7%, which was higher than 90.0%, demonstrating that the data pretreated method and the data fusion model proposed were effective and worthy popularization.

The authors have no conflicts of interest.

This work was supported by the Natural Science Foundation of Jiangsu Province [grant number BK20180816]; the Natural Science Research of Jiangsu Higher Education Institutions of China [grant number 17KJD550001]; and the National Natural Science Foundation of China [grant number 61602217].

(1) Y.B. Monakhova, D.N. Rutledge, A. Roßmann, et al., 

(2) S.Y. Park, J.K. Kim, J.S. Jang, et al., 

(3) K.L. Bett-Garber, R.J. Bryant, C.C. Grimm, et al., 

(4) H.B. Zhu, Y.Z. Wang, H. Liang, et al., 

(5) G.A. Camelo-Méndez, B.H. Camacho-Díaz, A.A. Villar-Martínez, et al., 

(7) M.M. Voorhuijzen, J.P. van Dijk, T.W. Prins, et al., 

(8) K. Ariyama, M. Shinozaki, and A. Kawasaki, 

(9) R.A. Lagad, S.K. Singh, and V.K. Rai, 

(10) G. Li, L. Nunes, Y.J. Wang, et al., 

(11) J. Promchan, D. Günther, A. Siripinyanond, et al., 

(17) X. Feng, Q. Zhang, P. Cong, et al., 

(19) S. Chaiwanichsiri, D. Thumrongchote, T. Suzuki, et al., 

(20) M. Sha, D.D. Gui, Z.Y. Zhang, et al., 

(21) D. Ballabio, E. Robotti, F. Grisoni, et al., 

(23) L.Y. Xia, S.G. Shen, Z.H. Liu, et al., 

(24) E. Borràs, J. Ferré, R. Boqué, et al., 

(26) W.J. Sun, X. Zhang, Z.Y. Zhang, et al., 

(27) M. Hohmann, Y. Monakhova, S. Erich, et al., 

(28) M. Sha, D. Zhang, Z.Y. Zhang, et al., 

(30) H.A.H. Naji, Q.J. Xue, K. Zheng, et al., 

(31) C.M. Farrelly, S.J. Schwartz, A.L. Amodeo, et al., 

(33) S.R. Amendolia, G. Cossu, M.L. Ganadu, et al., 

(36) L.Z. Hong, Q.L. Bao, L.T. Yue, et al., 

The application of data mining combined with data fusion of Raman and mid- infrared spectra was studied to improve discrimination ability for modeling the geographical origins of rice.

Enhanced Raman and Mid-Infrared Spectroscopic Discrimination of Geographical Origin of Rice by Data Mining and Data Fusion

Pet food is a multibillion dollar per year business that affects businesses and pet owners on a daily basis. Along with increased growth, there has been increased controversy. The melamine pet food scare of 2007 affected millions of people and their pets, along with the pet food industry. The pet food scare highlighted the potential for contaminants and controversial ingredients that could be contained in pet food. In addition to organic chemical contaminants and additives, there is also the possibility of toxic elemental contamination from protein sources, fillers, and manufacturing processes. Therefore, the search for “healthy” pet food goes beyond the choice of a name brand food or nutritious ingredients on a label.

Ten years ago, SPEX CertiPrep, first introduced its popular study on “Heavy Metals in Pet Food,” which was featured in 

 (1,2). The purpose of this 2009 study was to examine pet foods from a variety of sources to determine if there were potentially toxic elements present in the foods. Many of the pet foods sampled showed significant concentrations of various toxic metals. In many cases, the concentrations exceeded the extrapolated human limit values calculated to pet-size dosages.

In the intervening 10 years, the Food Safety Modernization Act (FSMA) was enacted, and pet food became one of the targets of scrutiny. This new 2019 study was produced as a joint effort between SPEX companies and CEM Corporation, and revisits pet food brands first analyzed in 2009, while looking at new brands that have emerged after the 2007 pet food crisis and later enactment of the FSMA, to determine whether heavy metal contamination has significantly changed in the decade following the first study. Using updated cryogenic and microwave technologies, samples were tested by inductively coupled plasma (ICP) and inductively coupled plasma–mass spectrometry (ICP–MS) to determine heavy metal content, and these results were compared to our previous 2009 study.

Recently, I presented a summary of our findings in an interview for 

 (3). The purpose of this article is to expand on those findings.

Some of the first commercial dog food has been credited to a businessman in the 1860s who developed a dog biscuit after he witnessed dogs being fed leftover hardtack biscuits in the ports of England. His idea later came to the United States in the late 1890s, but the canned pet food we know today did not exist extensively until after World War I (WWI). Ken-L-Ration brand (1922) was the first commercial pet food sold because of an excess of horse meat after WWI. In the early part of the 20th century, the U.S. Food and Drug Administration (FDA) did not exist, and only human food and medicine was under consideration of the 1906 Pure Food and Drug Act.

The first controls of pet food came under the 1938 Federal Food, Drug, and Cosmetic Act, which only required that pet food be safe to eat, produced under sanitary conditions, free from harmful substances, and labelled truthfully as to its quantity, identity, manufacturing process, ingredients (listed by weight), and nutritional information. The next significant advance of pet food also came after a major war when the consumer canned and packaged food companies that grew out of war rations, then transitioned to a post war economy. These companies capitalized on the ability to use human food by-products to create pet food.

Extruded dry kibble began to appear in the mid-1950s, and the pet food industry flourished and became mainstream, ultimately growing into a multibillion dollar a year global industry. The 1960s saw increases in specialized pet foods and an increase in brands and product lines for established companies. The first big public scare in the pet food industry started with the 2007 discovery of melamine and cyanuric acid in a wide range of pet food varieties and brands. Early in 2007, there were hundreds of pet deaths and illnesses because of kidney failure around the world, which was later traced back to illegal protein adulteration. Chinese raw material manufacturers had added the melamine to the pet food to artificially increase the protein content. Over 5000 product recalls were announced across multiple brands, manufacturers, and suppliers, linked to almost 10,000 pet deaths and illnesses.

For years after the melamine pet food disaster, pet owners continued to be concerned about their pet’s food. The initial disaster happened because of terrible oversight of supply chains and lack of testing for harmful ingredients. The pet food industry claimed that, as a result of the melamine scare, there was more active policing in their industry and products, but questions still remained whether pet food was safe or not. In 2009, SPEX CertiPrep made a decision to look at the elemental content of pet food. SPEX CertiPrep was a close-knit family company whose employees and owners were dedicated pet owners. The study solicited pet food from all the employees, and purchased food from local pet stores, groceries, dollar stores, and supercenters. We decided to investigate all price ranges, from the cheapest dollar store food to the boutique gourmet brands sold in specialty markets. The original study was conducted in the winter of 2009–2010, covering almost sixty samples. The samples included approximately equal amounts of cat and dog foods, as well as equal amounts of wet canned food and dry kibble.

Once the foods had been grounded, digested, and tested, it was found that there were substantially high amounts (up to ppm levels) of heavy metals and wear metals (from the processing equipment) in pet food, especially in dry kibbles. The amount of food that owners fed their pets increased the exposure of pets to higher levels of potentially toxic metals. Despite the fact that pet food had been under the control of the FDA since 1938, there was no guidance in 2009 regarding the limits for the most toxic elements to dogs and cats. Instead, the SPEX CertiPrep study used human exposure guidelines issued by the EPA and scaled them to a standardized size (50 lbs for a dog and 10 lbs for a cat).

In the decade since the first heavy metals study, the FSMA was created. Pet food fell under the direction of the FDA within the FSMA, and it was our hope that public awareness of the previous pet food regulatory shortfalls would be corrected and lower metal content would now be seen in pet food.

The basic techniques, materials, and methods for both the 2009 and 2019 studies were similar. Differences in the studies included the use of updated equipment and instrumentation, which will be discussed within the “Methods” section. For details on the previous 2009 experimental conditions, please refer to the originally published articles published in 

In both the 2009 and 2019 studies, concentrated optima-grade nitric acid was obtained from Fisher Scientific. Deionized water was obtained from an in-house laboratory water filtration and processing system, which produces deionized water (ASTM I grade).

The SPEX CertiPrep standards (SPEX CertiPrep Group) used in both studies included:


CLMS-2N: Claritas Multielement Solution Standard

(10 mg/L: Ag, Al, As, Ba, Be, Bi, Ca, Cd, Co, Cr, Cs, Cu, Fe, Ga, In, K, Li, Mg, Mn, Na, Ni, Pb, Rb, Se, Sr, Tl, U, V, and Zn)

CL-ICV-1: Claritas Initial Calibration Verification 1

(1000 mg/L: Fe, K, Ca, Na, Mg, Sr; 10 mg/L: Ag, Al, As, Ba, Be, Cd, Co, Cr, Cu, Mn, Mo, Ni, Pb, Sb, Se, Tl, V, Zn, Th, and U)

CLU2-2Y: Claritas Uranium Single Ele- ment Standard (1000 mg/L)

CLHG2-1AY: Claritas Mercury Single Element Standard (10 mg/L)

Standards were diluted to μg/kg concentrations appropriate for ICP– MS analysis in 5% nitric acid in deionized water solution to match the acid matrix of samples. Standards of 0.1, 0.4, 1.0, 4.0, and 10 μg/kg, were made up in 5% HNO

, and matrix matched for ICP–MS calibration.

Optima-grade nitric acid used in digestion was added to the deionized water blanks prior to analysis in a comparable concentration as to the acid concentration found in the pet food samples. In addition to water blanks, optima-grade nitric acid was processed as a sample in the microwave vessel to determine contamination from the polytetrafluoroethylene (PFTE) vessels or any memory effects of the previous samples. After each pet food sample digestion run, 10 mL of nitric acid were digested discarded from the microwave vessels to clean vessels. A second 10 mL digestion of nitric acid was then repeated, and the digest was diluted in the same method as the pet food samples and treated as a vessel blank.

Sample Collection, Preparation, and Digestion

For this investigation, a total of 61 different cat and dog foods were bought from local stores or donated by employees of the SPEX companies and CEM Corporation. The samples consisted of dry food varieties. Of the 60 dry foods, 37 of them were dog food, and 23 of them were cat food samples. The range of quality of the pet foods was from “discount” to “gourmet brands.” Pet food prices ranged from the “bargain” store foods priced at 0.02¢ per ounce to gourmet or specialty foods purchased from pet suppliers priced above 0.42¢ per ounce. There are 11 samples that were similar or identical to brands or varieties studied in the original 2009 study.

Dry cat and dog food samples (10–15 g) were ground to a uniform powder in a cryogenic mill (SPEX SamplePrep 6875D Large Dual Freezer/Mill with 6885 mid-size Poly-Vial). Operating conditions are shown in Table I.

One half gram of the powdered or blended sample was digested with 10 ml of nitric acid in a CEM Mars-6 microwave oven (CEM Corporation) using a preprogrammed pet food method. After digestion, samples were diluted to 50 mL volume with deionized water. Prior to analysis, samples were diluted to a final concentration of 1000x using deionized water.

Primary sample analysis was conducted using a PerkinElmer inductively coupled plasma–optical emission spectrometer (ICP–OES) (PerkinElmer Inc.). The ICP–OES was used to determine the macro-elemental composition of the foods to select standards which would allow for correct matrix interferences. Elements measured by ICP–OES included aluminum, calcium, copper, iron, magnesium, manganese, phosphorus, potassium, selenium, silicon, sodium, sulfur, and zinc.

Analysis of heavy metals in pet food using ICP–OES was conducted to determine if there were potentially toxic elements present. Many of the samples showed significant concentrations of various toxic metals.

March 2021

Library Searching

Analysis of a Mixture Solution Using Silver Nanoparticles Based on Surface-Enhanced Raman Spectroscopy (SERS)

Enhanced Raman and Mid-Infrared Spectroscopic Discrimination of Geographical Origin of Rice by Data Mining and Data Fusion

Heavy Metals in Pet Food: Changes in Heavy Metal Contamination in Pet Food Over the Past Decade

The 2021 Employment and Salary Survey: Covid-19 and the New Virtual World

Combined Raman and Photoluminescence Imaging of Two-Dimensional WS2