Food fraud and adulteration is a serious and increasing problem with significant health and economic impacts. Although robust and sophisticated laboratory methods have been developed to detect fraud, the question of how that experience can be deployed in the field remains. Ideally, handheld instruments operated by nonscientists would be used at a port, in a food distribution center or on a supermarket loading dock. Cases of food fraud — some of which cause serious illness and death — make the headlines every few months. Although some of the fraud cases have caused a significant number of hospitalizations and fatalities1, the economic costs are not widely appreciated; each year, lost sales, product recalls and loss of consumer confidence are estimated to cost legitimate retailers around $15 billion2. As a result, a number of government agencies and organizations — including the Food Protection and Defense Institute, headquartered at the University of Minnesota, and the U.S. Pharmacopeial Convention’s Food Fraud Database — now track these incidents and determine how to detect and prevent them. But how can analytical instruments help, and how do advances in photonics enable portable instruments? Analytical chemistry of food fraud The analytical science for food authenticity testing is well-established1,3,4, with laboratory methods that include liquid and gas chromatography techniques coupled with mass spectrometry, calorimetry, polymerase chain reaction analysis (for DNA fingerprinting) and nuclear magnetic resonance (NMR). These methods are highly accurate and precise, and provide the reference analytical data required for screening techniques. Optical techniques include fluorescence, UV-visible, near-infrared, mid-infrared and Raman spectroscopies. As examples, the American Association of Cereal Chemists has approved near-infrared techniques for determining protein in grain, while the consensus standards body AOAC International has official and internationally recognized methods for food analysis. In this fight, laboratory techniques can only take us so far: A sample comes to the lab, significant preparation is done to it, a long analytical procedure performed, and the data is worked up. Typically two weeks later, it is presented to nonscientists for further action. Field-based detection and screening can speed this up, enhance protection of the public and enable suspect materials to be impounded more rapidly. Field-based screening takes place not in a laboratory, but at the sample. That quick response thereby prevents those materials from entering the food chain and being dispersed. Screening is designed to give red light/green light answers — does it appear to be authentic, or is it suspicious? — and not detailed or quantitative analytical results. As optically based screening methods for food fraud are developed, it’s important to know the strengths, weaknesses, limitations and detection limits for each technique (Table 1). Near-infrared spectroscopy, usually employing diffuse reflectance sampling, is routinely used to determine the bulk composition (protein, fat, carbohydrate, moisture) of foods and grains. The error sources and variables are well understood4. Because optical spectroscopy techniques generate a single ensemble spectrum from the sample, with limited signal-to-noise, the detection limits tend to be in the 0.1 percent to 1 percent range, without enrichment or enhancement. Further, speciation of a food (for example, the types of fish or meat) requires DNA testing, and is outside the realm of optical spectroscopy. The estimated concentrations of some materials in foods are very low, well under 1 percent, and are classified as trace components. For allergens, one protein in peanuts is present at ~0.1 percent of the peanut itself, and susceptible individuals react to much less. Pesticides and fungicides can be present at the 0.0025 percent level. Preservatives can have concentrations of 0.5 percent, while colorants in common foods are present at levels of 0.06 percent. The detection of trace components in optical absorption spectroscopy, especially near-infrared and mid-infrared, relies on detecting small absorbances within a large and varying background. It is highly unlikely that trace components will be detected by an optical spectroscopy technique, especially one using a handheld or portable instrument, with shorter measurement times and lower signal-to-noise performance, compared to a laboratory instrument. Established methods for trace analyses of pesticide residues, for instance, typically require a separation technique coupled with mass spectrometric detection. This has the analytical advantage of isolating the component(s) of interest via chromatography, and detecting them using a highly sensitive technique with an inherent low background (mass spectrometry). Raman spectroscopy, as a quasi-emission method, potentially has a zero background, but is an optically inefficient technique, with only about 1 in 10∧6 input photons are Raman scattered5. In principle, both surface-enhanced Raman spectroscopy (SERS) and fluorescence would be able to detect some minor components in complex mixtures, if they are SERS-active (for Raman), or fluoresce as in fluorescence spectroscopy. But these do not represent universal approaches. Handheld optical spectrometers With advances in electronics and photonics, it is possible to design and build a variety of portable optical spectrometers5,6,7, enabling the laboratory to move to the field (Figure 1). Optical-based analyses often require zero sample preparation, can be performed in a “point-and-shoot” method using a portable instrument and require a data acquisition time of only a few seconds. For screening purposes, the instruments may not need the precision, accuracy or limit of detection of a laboratory instrument. Problematic samples may be referred for laboratory analysis later. Figure 1. A handheld Raman spectrometer. Courtesy of B&W Tek. It is not enough to have a portable instrument — it has to generate answers, not just spectra. These instruments are, and will be, used by nonscientists, sometimes in stressful situations. So the instruments must generate, with confidence, clear answers, often in a “green screen/red screen” format. For identification, these answers may be to 8 can indicate pass/fail, and may be able give an identification answer. Quantitative analyses typically rely on measuring the spectra of a very large number of samples, often in the thousands, correlating those spectra with the reference chemical data provided by lab methods, via data science, to build a “model,” and then deploying that model to give results. In the case of near-infrared analysis of grains, that work has been done over tens of years, and the resulting models are quantitative and widely used4,9. No single optical technique will cover every situation or every type of food; there are multiple optical techniques that can be used for analysis (Table 2). Photonics and portable optical instruments There are two broad classes of instruments: conventional designs, such as multiple spectrographs and interferometers adapted for field use, and photonic-based instruments and engines, often using technologies derived from optical telecommunications. Axsun Technologies uses broadband gain media in conjunction with a MEMS scanning Fabry-Pérot interferometer to produce a widely tunable near-infrared laser. As a predispersive instrument, this allows illumination of a large sample area and detection via a single element detector, which could be a large-area detector, in order to maximize light collection from the diffusely reflecting sample. Spectral Engines takes a different approach, with a scanning Fabry-Pérot interferometer mounted directly over a single element detector. This allows a choice of illumination sources, but attention has to be paid to the étendue of the analyzer. Grating light valves can provide a multiplex advantage, while using a single element detector, via Hadamard techniques, and this is implemented in different ways by Thermo Fisher Scientific in the microPHAZIR, and Texas Instruments in their DLP NIRscan Nano, both of which employ MOEMS-based engines. The MEMS theme is continued by Fraunhofer IPMS’s engine, which uses a MEMS scanning grating, implemented via a comb drive. A comb drive is also the actuator for the moving mirror of SiWare’s MEMS FT-NIR engine. Lastly, linear variable filter technology, originally designed for telecom applications by OCLI, is employed by Viavi. It is an advantage for a portable instrument to use a single-element detector for cost, power consumption and heat dissipation reasons, and this drives some of the design choices in these instruments, especially the use of multiplexing techniques. Optics are only a portion of the instrument: These photonic innovations need to be paired with the recent advances in consumer electronics to yield portable instruments. In recent years, we have seen dramatic improvements in lower-power processors, operating systems, user interfaces, displays, memory and communications technology, driven by the mobile phone market. In fact, the power of smartphones has encouraged researchers to adapt them for visible spectroscopy (colorimetry), especially for medical diagnostic testing in low-resource areas of the world10. Homogeneity and heterogeneity Because food is inherently heterogeneous, instruments for food analysis require a different optical arrangement from those used for homogeneous samples like incoming raw materials, bulk chemicals, plastics, liquids, solutions and gases. For homogeneous samples, a small spot size of a few millimeters or less is practical, and sampling one spot is acceptable, using either free-space optics or a fiber-in, fiber-out probe. Many identification applications use Raman spectroscopy, although near-IR and mid-IR spectroscopy can also be used. Both Raman and near-IR are “point-and-shoot” techniques, and can be used with optically transparent containers having minimal spectral contributions — for instance, clear glass and polyethylene. However, for heterogeneous samples, a large spot size of several centimeters in diameter is required, often with a sample rotator (Figure 2), and sometimes an integrating sphere will be employed. In any case, the operator may need to sample several areas. The typical application is quantitative, and the database building to support that is complex and requires reference analyses of all the samples. In addition, samples may often be “wet” and/or “dark,” dramatically reducing the signal-to-noise of the measurement. Figure 2. A near-infrared spectrometer with sample cup and rotator. Courtesy of ©2013-2017 PerkinElmer Inc. All rights reserved. Printed with permission. Near-IR solid samples are typically examined in reflectance, and the analyst needs to understand what is actually being sampled by the beam. For instance, is the composition of the surface the same as the bulk? Are you analyzing just the surface, or what depth is being probed? In the near-infrared, these answers will be wavelength dependent: The absorption coefficients drop (and therefore penetration depth increases) dramatically from the longwave NIR (~1.7 to 2.5 μm), through the midwave NIR (~1.0 to 1.7 μm), to the shortwave NIR (~visible to 1.0 μm). A graphic illustration of the sampling problem in foods is provided by a blueberry muffin (Figure 3). If the optical instrument has a small sampling spot size, then inside the muffin, if you sample where a blueberry is, the measured fat content will be low, because blueberries don’t have fat. However, if you sample where there is pastry, the measured fat content will be higher. But if the butter was not mixed very well, there will be localized high fat areas. Figure 3. Cross sections of two store-bought blueberry muffins, showing their variability and internal heterogeneity. Courtesy of Eric Crocombe. Food fraud detection using portable optical instruments Ultimately, near-IR is a secondary technique. It relies on primary analysis of the samples using an established method, such as the Kjeldahl method for organic nitrogen, and combining that data in a chemometric method with the spectra of those samples (Figure 4). Because natural products vary significantly, and because near-IR spectra are not very specific, a large number of calibration samples are required, along with a validation step to ensure that the correlations the method arrives at are valid and not coincidental. In the case of the long-established grain analysis, this can be many thousands of calibration samples, collected over time and updated yearly. Figure 4. Flow chart for near-infrared analysis of food — both raw materials and food products. Courtesy of Richard Crocombe. In detecting food fraud, the user cannot rely on the product labels for the primary analysis of the calibration samples. Calibration samples must be authentic. Commercial samples are routinely identified as fraudulent so if one or more samples are not what they are labeled, but are input in the model as genuine, the model will no longer be able to detect those fraudulent samples. They will be reported as genuine, and the result will be a “garbage in, garbage out” model. “Garbage in” represents a systematic error, one which is not averaged out by adding more data. The same reasoning applies to other types of errors, for instance in the recording of the spectra. The spectra need to be acquired under carefully controlled conditions, so that signatures of containers, surfaces, fluorescent lights, clothing, hands, etc., are not introduced into the database. Random errors can be reduced by including a large number of samples, but systematic errors will remain. Heterogeneity and optical depth penetration need to be understood so that it’s clear where the spectrum originates: an average of the whole sample, the surface, the skin, seeds, wax coating, etc. Figure 5. Extra virgin olive oils — note their ‘greenish’ hue. Courtesy of Eric Crocombe. To perform screening for food fraud and adulteration using portable optical spectrometers, there must be a commitment to analyzing a very large number of samples by standard methods, recording their “metadata” (reported origin, age, appearance, etc.), then collecting their optical spectra with a full understanding of sampling and characteristics of each spectroscopic technique, and using all of that data to build validated chemometric methods. This is not something that can be done in a “crowd-sourced” fashion — it has to be performed using good laboratory practice, and familiarity with “big data” analytics. An organization in Massachusettes is taking that approach — TeakOrigin is a data analytics company that works as a truth detector for food by bringing together molecular spectroscopy, analytical chemistry and deep machine learning. Olive oil and vegetable oils To give an idea of the challenges involved, the near-infrared spectra of several olive oils (Figure 5) and vegetable oils were recorded on portable instruments. The spectra of olive oils and canola oil appear very similar, with subtle but real peak intensity differences, and a significant distinction around 5520 cm-1 (~1.9 μm) (Figure 6). Figure 6. Near-infrared transmission spectra of olive oils (upper traces) and canola oil (lowest trace). The spectra are very similar, but a significant difference can be seen around 5250 cm−1 (~1.9 μm), with a band present in the olive oils but missing in canola oil. Courtesy of TeakOrigin. Food fraud is a continuing problem with significant consequences for human health and economic damage. Handheld optical instruments, enabled by advances in consumer electronics and photonics, have the potential to screen for fraud, but the underlying spectroscopy and analytical science need to be fully understood for those instruments to have valid methods. In addition, the sensitivity (detection limit) of a portable optical instrument should not be overstated — these will be screening methods only, and trace detection will likely require a laboratory technique. Meet the authors Richard Crocombe has a Ph.D. in chemistry and spectroscopy, worked with major manufacturers of spectroscopic instrumentation for over 30 years and recently set up his own consulting company; email: racrocombe@gmail.com. Ellen Miseo has a Ph.D. in physical chemistry with a concentration in spectroscopy and has worked in both spectroscopic instrumentation and food analysis. Currently she is the chief scientist at a startup devoted to using spectroscopy to understand high-value foods; email: ellen.miseo@gmail.com. Acknowledgments The authors want to thank Brent Overcash (TeakOrigin), Rob Packer (PerkinElmer) and Peter Larkin (Cytec Solvay Group) for helpful discussions; B&WTek for running the Raman spectra of vegetable oil samples; Marla Harris for baking the muffins; Eric Crocombe for the olive oil and muffin photographs; and B&WTek and PerkinElmer for permission to reproduce instrument pictures. References 1. R. Evershed and N. Temple (2016). Sorting the Beef from the Bull: The Science of Food Fraud Forensics. 1st Ed. London: Bloomsbury Sigma. 2. S. Yager (2015). The wheels of crime are greased with olive oil. The Atlantic. https://www.theatlantic.com/magazine/archive/ 2015/07/high-cost-food-fraud/395327/. 3. G. Downey (Ed.) (2016). Advances in Food Authenticity Testing. 1st Ed. Cambridge, England: Woodhead Publishing (Elsevier). 4. P. Williams and K. Norris (2001). Near-Infrared Technology in the Agricultural and Food Industries. St. Paul, Minn.: AACC Publications. 5. M.D. Hargreaves (2014). Handheld Raman spectrometers and their applications. In Encyclopedia of Analytical Chemistry. Hoboken, N.J.: John Wiley & Sons. 6. J. Antila et al. (2014). MEMS- and MOEMS-based near-infrared spectrometers. In Encyclopedia of Analytical Chemistry. Hoboken, N.J.: John Wiley & Sons. 7. C.A. Teixeira dos Santos et al. (2015). Applications of portable near-infrared spectrometers. In Encyclopedia of Analytical Chemistry. Hoboken, N.J.: John Wiley & Sons. 8. C. Gardner and R.L. Green (2014). Identification and confirmation algorithms for handheld spectrometers. In Encyclopedia of Analytical Chemistry. Hoboken, N.J.: John Wiley & Sons. 9. C.A. Roberts et al. (2004). Near-Infrared Spectroscopy in Agriculture. Madison, Wis.: American Society of Agronomy, Crop Science Society of America, Soil Science Society of America. 10. J. Canning (2016). Smartphone spectrometers and other instrumentation. SPIE Newsroom. DOI: 10.1117/2.1201512.006220. Additional Web Exclusive content below. References 1. P.A. Smith (Jan. 19, 2015). The New York Times, https://www.nytimes.com/2015/01/20/science/a-lab-is-trying-to-keep-china-from-dodging-us-tariffs-on-honey.html?_r=0. 2. For a crisis timeline, see http://news.bbc.co.uk/2/hi/7720404.stm and https://en.wikipedia.org/wiki/2008_Chinese_milk_scandal. 3. See http://tna.europarchive.org/20111030113958/ and http:/www.food.gov.uk/safereating/chemsafe/sudani/. Portable Optical Spectrometers in Focus In many cases, the funding and initial applications for portable spectroscopic instruments were in the safety and security sector. The first portable Fourier transform infrared (FTIR) spectrometer was produced by SensIR in about 2000, and many of these were sold following the events of 9/11. With the developments in telecom optics, it was relatively straightforward to build a small Raman instrument with a 785-nm laser, spectrograph and CCD. But the first highly engineered instrument of this type — designed for emergency response — was from Ahura Scientific in around 2005. The first FTIR handheld was marketed by Ahura Scientific (now part of Thermo Fisher Scientific) in 2010. Portable near-IR instruments have had a different arc. Some were developed for niche applications in fuel and petroleum, or for sugar in fruit, but a major impetus was for recycling. In about 2004, Axsun Technologies developed the Anavo, followed by Polychromix (now also part of Thermo Fisher Scientific) in 2005, initially targeted at carpet fiber recycling. Overall, the strategies for these companies was to develop the platform and the first application, preferably with some outside funding, then move into adjacent markets (for example from first responders to the military), and then into commercial applications such as raw material identification in pharmaceutical manufacturing. Moving into detection of food fraud and adulteration was a natural step. For portable Raman instruments1, concerns about sample fluorescence drove designers toward longer wavelength excitation, with the combination of 785-nm excitation and silicon-based detectors being the most frequent choice. Some 1064-nm excitation instruments became available, using InGaAs detectors, but their performance was reduced relative to 785-nm excitation due to both the υ4 Raman cross-section dependence, and the diminished performance of InGaAs detectors, relative to silicon. Portable mid-IR instruments almost uniformly employ interferometers, and single element DLaTGS pyroelectric detectors. Portable near-infrared instruments2,3 use a variety of technologies, often multiplex techniques, stimulated by the high cost and cooling requirements of InGaAs array detectors. Understanding of optical throughput, or étendue (area and solid angle product) is critical, especially if the instrument has to illuminate and/or collect light from large areas. Some of the available instruments and engines are described below (Table 3). This does not include spectrographs from optical equipment manufacturers like Ocean Optics, Avantes, StellarNet, etc., nor does it include laboratory instruments from manufacturers like FOSS, Metrohm, Perten, etc. Courtesy of Richard Crocombe. References 1. M.D. Hargreaves (2014). Handheld Raman spectrometers and their applications. In Encyclopedia of Analytical Chemistry, John Wiley. 2. J. Antila et al. (2104). MEMS- and MOEMS-based near-infrared spectrometers. In Encyclopedia of Analytical Chemistry, John Wiley. 3. C.A. Teixeira dos Santos et al. (2015). Applications of portable near-infrared spectrometers. In Encyclopedia of Analytical Chemistry, John Wiley. Sampling Challenge: The Case of the Blueberry Muffin A graphic illustration of the challenges in sampling food lies with the blueberry muffin. On the surface of the muffin are sugar crystals, about 1 to 2 mm on the edge, and blueberries about 1 cm in diameter (Figure 1). Some parts of the surface are light brown, and others dark brown, as a result of chemical changes induced by heat. (For chemists, this is a Maillard reaction, between an amino acid and a reducing sugar1.) Figure 1. A tray of store-bought blueberry muffins, showing their surface heterogeneity. Courtesy of Eric Crocombe. A cross section of two store-bought muffins reveals their internal heterogeneity, and sample-to-sample variability (Figure 2). The muffin at the top appears to have higher blueberry content. If the optical instrument has a small sampling spot size, then inside the muffin, if you sample where a blueberry is, the measured fat content will be low, because blueberries don’t have fat. However, if you sample where there is pastry, the measured fat content will be higher. But if the butter was not mixed very well, there will be localized high fat areas. Figure 2. Cross sections of two store-bought blueberry muffins, showing their internal heterogeneity. Courtesy of Eric Crocombe. Nutritional labels (Figure 3) convey information for the “average muffin.” But how do we associate the label information with a spectrum or spectra from the muffin, and hope that it makes sense within a chemometric procedure? And if, unbeknown to the sampler, this was a fraudulent product, then using any spectra from the muffins, combined with the fraudulent information on the label, would contaminate the database we are building. Figure 3. Nutritional label for the store-bought muffins. Courtesy of Eric Crocombe. It’s no easier if you are baking your own muffins and you are confident about the ingredients (Figure 4). The muffin size will be different, and even some of the ingredients. A common recipe calls for two large eggs; if you use medium eggs, the fat content will be lower. The recipe calls for milk, but is it whole milk or skim milk? In this homemade example, frozen blueberries were used and allowed to thaw. The effect of this is to generate juice, which has colored a large portion of these muffins (Figure 5.). Figure 4. Ingredients for homemade muffins. Courtesy of Eric Crocombe. Figure 5. The baked homemade muffins. Note the completely different appearance, as compared with store-bought muffins. Courtesy of Eric Crocombe. References 1. See, for instance, the “Science of Cooking” website: http://www.scienceofcooking.com/maillard_reaction.htm. Raman Spectroscopy of Olive Oil and Vegetable Oils Raman spectroscopy is emerging as a vital way to test for the differences between good olive oil and run-of-the-mill vegetable oils. Using portable instruments, the spectra of several olive oils and vegetable oils were recorded, and the results are shown below. As might be expected from the color, olive oil fluoresced with 785-nm excitation, although the bands are clear above the fluorescence background. The 1064-nm excitation dramatically reduced the fluorescence, although on the particular portable instrument used the Raman shift range was not as large, and the resolution was lower, making detection of small spectral differences more difficult. Raman spectra of an extra virgin olive oil sample, with 785-nm (black) and 1064-nm (blue) excitation. Note the sloping baseline for 785-nm excitation, due to fluorescence. The 1064-nm excitation spectrum was run at lower resolution. Courtesy of TeakOrigin/B&WTek.