Richard A. Crocombe

	Portable spectroscopic instruments have not had significant visibility within the scientific community compared with, for instance, the current generation of high-performance laboratory mass spectrometers. This is attributable to a number of reasons, including their very applied (as opposed to research) nature, resulting in few publications in the refereed literature, and their predominant use by non-scientists. The numbers of these instruments deployed today can therefore come as a surprise. For instance, as early as the 1990s, more than 60,000 ion mobility spectrometers (IMS) were deployed worldwide. Cumulative handheld X-ray fluorescence (XRF) shipments today total more than 100,000, and portable Raman shipments number in the tens of thousands. For these, and some other categories of portable spectrometers, there are more instruments in the field than there are in laboratories. 

	The rationale for portable and handheld spectrometers is simple: allowing the spectrometer to be taken to the sample, as opposed to bringing the sample to the spectrometer, thereby moving the laboratory to the point of need-the location of the sample. Portable instruments provide the ability to make informed decisions on the spot, and by rapidly delivering actionable results at the point of need, you change or transform the way the customer does business, whether a hazardous materials (hazmat) technician, a scrap metal dealer, personnel in pharmaceutical manufacturing, or the military. To generate these actionable answers, the developers of these instruments have put extensive resources into reliable identification algorithms, spectroscopic libraries or databases, and qualitative and quantitative calibrations. 

	These instruments have dramatically increased their capabilities over the past 20 years, while at the same time becoming smaller and lighter. On one hand, this is due to all the developments in consumer electronics and computing power, and, on the other, to ongoing research and development (R&D) and manufacturing experience and innovation in the companies producing the instruments. Nonetheless, some challenges remain, and, as an example, we can look at portable Raman spectroscopy, especially in its safety and security applications.

	All Raman spectroscopists, no matter what kind of samples they deal with, are aware of the potential for fluorescence inference. Excitation at 785 nm has been a good compromise between long and short wavelength excitation while still using lower cost, and lower noise, silicon-based detectors. There has been a recent trend to move to 1064 nm excitation, although a penalty is paid in terms of measurement time for comparable signal-to-noise spectra. Deep-ultraviolet (UV) excitation is intriguing, especially for standoff detection, but brings along its own issues, and obviously does not work with glass containers. Various shifted-excitation schemes are also in use, and time-gating is commercially available in a laboratory instrument. It’s not clear if there is a “best approach,” or what a universal solution, could be.

	Standoff detection is of obvious benefit for the military and other practitioners dealing with potential explosives or highly toxic materials. Truck- and backpack-mounted instruments have existed for some time, often using deep-UV excitation, which has the benefit of being able to use solar-blind detectors. But this technology is still relatively immature, and has issues related to the lifetime of key components. A handheld spectrometer operating at 830 nm excitation, capable of a one-meter standoff, has recently been announced, and that represents the first true handheld device of this kind. Sample heating is a concern for potentially explosive samples, as is the pressure required to obtain good spectra using portable Fourier-transform infrared (FT-IR) instrumentation with diamond attenuated total reflectance (ATR) sampling. Sample heating is being addressed in schemes involving motion of the laser spot over the sample.

	Real-life samples are most often mixtures, and this challenge is being addressed with a combination of library and algorithm development, aided by improvements in mobile processing power and memory, and also with the possibility of wireless communications and cloud processing. Finally, examination of a sample contained in packaging is desirable, in areas as different as pharmaceutical incoming raw material inspection, smuggling of contraband into prisons, and detection of potential explosives at airport checkpoints. Spatially-offset Raman spectroscopy can address that issue for many packaging materials (not metals, though!), and other methods have been developed.	

	Finally, the shrinking size and increased performance of these Raman instruments lends them to “hyphenation” or “combination” with other portable spectrometers. A combined Raman–FT-IR instrument has been on the market for several years now, but combinations like Raman–NIR, Raman–XRF, and Raman with laser-induced breakdown spectroscopy (LIBS) are also possible. Indeed, the next year or two should see the launch of both U.S. and European Mars missions, with Raman and LIBS capabilities, a fantastic example of portable and standoff spectroscopy. We shouldn’t forget that a gas chromatography–mass spectrometry (GC–MS) instrument was on a Viking mission to Mars in the late 1970s, and that FT-IR instruments were on the two Voyager missions, launched in late 1970s, continuing on their planetary tours throughout the 1980s. Hyperspectral instruments have also been shrinking in size, especially those based on silicon detectors. Handheld and drone-mounted versions are now available, and we will can anticipate increasing applications of those instruments in the future.

	So we can expect to see developments in all of these areas, and also significant diminution of the size of these instruments. Today, a Raman instrument about 1” x 1” x 1” in size is possible. This is a far cry from the scanning triple monochromator, argon ion excitation, photomultiplier detection, and chart paper output instrument I used for my PhD studies!

	All of the capabilities described above come from instruments developed by the major analytical instrument companies, or by smaller companies and institutions under government contracts. Recent years have seen the start of a new market segment, enabled by the very low cost of ambient temperature, silicon-based, detectors, especially the cameras in smartphones. Small, portable, low-cost spectrometers operating in the visible region have been available for many years, used, for instance, in colorimetric analysis of water, and color matching (such as paint). One research thread has been to investigate whether existing clinical analyses and assays can be adapted to smartphones, enabling this work to be done at the point-of-care, potentially in remote areas with very limited infrastructures. This research leverages existing “chemistries” (that is, established reagents and procedures), microfluidics, and both the spectroscopic potential of smartphone cameras, and their inherent processing power and communications abilities. This was the topic of a recent Pittsburgh Conference Symposium, organized by the Society for Applied Spectroscopy (SAS). Also featured in that symposium was a detailed description of the capabilities of smartphone cameras, and what is required to obtain good spectroscopic data from them. This leads to the possibility of scientific organizations equipping members of the public with accessories and software for their phones, and then coordinating “citizen science” studies involving thousands of people. Work is going on in this area in The Netherlands, following on from a project a few years ago measuring atmospheric aerosols, in which over 3000 members of the public participated.

	However, there are also new companies marketing devices in the visible-to-~1000 nm region directly to the public, and making wide claims about their analytical capabilities. Although there is obviously interest in the general public to be able to analyze materials, including food and medicines, there are numerous significant issues here. The first is whether there is appropriate information in the spectral region covered. You can obviously match colors, and you can certainly examine classes of compounds like lycopenes and carotenoids in fruits and vegetables. But general food analysis has traditionally used the near infrared, especially the 1000– 1700 nm region, where overtone and combination bands are much stronger. (This author has even seen a claim to assay gold using a low cost visible spectroscopy device, something that is straightforward using handheld or portable XRF, but impossible in the visible region of the spectrum!) Second, foods, in particular, are highly heterogeneous, and laboratory instruments typically employ large sample areas, often combined with sample spinners or integrating spheres for their measurements. Third, the interrogation of foods using a handheld instrument, especially by an untrained operator, can be highly irreproducible, giving erroneous results. Fourth, years of experience in the near-infrared community have shown the need for careful and validated calibrations, to eliminate the possibility of accidental correlations when limited sample sets are used. Fifth, some of these new vendors seem to be relying on “crowd-sourced” data. Sixth, there seems to be a lack of understanding about the possible detection limits of optical spectroscopy in the condensed phase, with claims to detect trace amounts, such as allergens, pathogens, pesticide and herbicide residues, and so on. With all the issues noted above, and the lack (most likely) of reference analyses, the potential for “garbage in, garbage out” analyses is extremely high. This list of caveats can go on and on, but suffice it to say this is an area of concern, with the potential to mislead the general public, with even fatal consequences. 

	To sum up, portable and handheld spectroscopy is a rapidly growing area, with instruments spanning the electromagnetic spectrum from the X-ray region to the mid-infrared, and ranging in techniques from GC–MS to nuclear magnetic resonance (NMR) relaxometry. Instruments are becoming smaller, lighter, and more capable, but we still have to understand the basic analytical spectroscopy concepts of sampling, heterogeneity, detection limits, signal-to-noise, and so on.	 

	This short article does not contain any references; for those the reader is directed to this author’s recent review article: Richard A. Crocombe, “Portable Spectroscopy”, 

 is the principal of Crocombe Spectroscopic Consulting, and the 2020 President of the Society for Applied Spectroscopy. Direct correspondence to 

Articles

Portable spectroscopic instruments have not had significant visibility within the scientific community compared with, for instance, the current generation of high-performance laboratory mass spectrometers.

The Future of Portable Spectroscopy

Here in Part IV, we will cover new spectrometer fabrication techniques, new light sources, and wavelength-selective devices. Researchers are combining new wavelength-selective devices with focal plane arrays to produce tunable arrays, and photonics technology is enabling new classes of multiplex spectrometers, based upon Hadamard transform techniques. So we are looking at "the art of the possible" (4): what new spectrometers can be constructed, and how they can be constructed, using all the technologies developed during the 1990s. In particular, many new NIR technology developments employ Fabry–Pérot tunable filters, the theory and properties of which were described in Part II of this series (1). We will concentrate on spectrometers made possible by micro electro mechanical systems (MEMS), micro-opto-electro-mechanical systems (MOEMS), and LIGA technologies. Some other interesting developments are omitted, but should be noted: slab waveguide spectrometers (5–9); handheld tunable diode laser absorption spectrometers (TDLAS) (10); a biomedical, in-vivo, scanning spectrometer (11); linear variable filters (12); fixed-filter instruments based upon spectroscopic principal components (13,14); developments in NIR array detector technology (15,16); a macroscopic tunable Fabry-Pérot filter coupled to a focal plane (17); and Fourier-transform (FT)-IR spectrometers using photoelastic modulators (18) and liquid crystals (19) to provide optical path differences, just to name a few. 

LIGA (20–23) is a German acronym for X-ray lithography (Lithographie), electroplating (Galvanoformung), and molding (Abformung), and this technology was described in Part II (1). A number of spectrometers have been fabricated using the LIGA process. They fall into two classes: first, those in which the metal LIGA parts are used directly in the spectrometer, and second, those in which the metal LIGA parts are used as the master mold for a plastic-body spectrometer. 

A LIGA NIR FT spectrometer has been described (24–26), in which the whole optical system fits on a chip with dimensions of 11.5 mm × 9.4 mm. The structural components of the spectrometer were fabricated in an iron-nickel (FeNi) alloy, as were the plunger and microcoil cores of the drive mechanism and the fiber-optic mounts. The mirrors were also FeNi structures, made by LIGA. The beamsplitter, detector, and lenses were commercially available optical components. A mirror travel of about 300 μm was planned, but in practice, only 54 μm was achieved, with a resolution of about 25 nm.  

LIGA has been used in the manufacture of visible and NIR spectrographs, by production of a nickel master molding tool, and then hot-embossing polymethylmethacrylate (PMMA) (27,28). These are monolithic, hollow, "slab-waveguide" spectrometers (Figure 1). The UV–vis and NIR spectrometers from microParts share the same design principle and production process (29). These designs are based upon a monolithic spectrometer chip, which contains all the optical components — entrance slit, self-focusing echelette diffraction grating, and deflection mirror — all on a single injection-molded part. A reflective coating (aluminum in the NIR and gold in the UV–vis) is deposited on the chip. The UV–vis spectrometers use a 256-pixel detector and cover the range of 350–1050 nm at <10-nm resolution, while the NIR version is equipped with an InGaAs array detector and covers 1000–1700 nm at <16-nm resolution. The NIR spectrometer module is 61 mm × 42 mm × 16 mm in size.  

A second spectrometer, also fabricated by LIGA precision injection molding, operates in the visible and short wavelength NIR (30). With a 256-pixel detector, the NIR version covers 740–1050 nm at <8-nm resolution. The size of this "spectrometer engine" is 41 mm × 21 mm × 26.5 mm. The use of LIGA to produce a master mold, and subsequent hot-embossing or precision injection molding, leads to a lower-cost manufacturing technique as compared with conventional assembly processes. The microParts UV–vis spectrometer described earlier can sell for less than $400 in very large quantities (29). LIGA spectrometers of this type have found a number of diverse, dedicated applications, principally in the visible region; for instance, in noninvasive measurement of newborn jaundice (31), measurement of blood parameters (32), and color grading of diamonds (33). 

As we have seen (2), light-emitting diodes (LEDs) can be used as sources in NIR spectrometers, but a variety of enhancements can improve their utility. Adding an optical waveguide region, consisting of a ridge structure on the LED's upper surface, will guide the emitted light by total internal reflection, and allows efficient coupling of that light into small optical fibers. A super luminescent light-emitting diode (SLED) is a ridge-waveguide light-emitting diode, pumped at high current levels so that stimulated emission occurs, enhancing emission into the waveguide (34). A SLED is not a laser, and precautions are taken to prevent lasing; typically, the rear of the device is not electrically pumped, and light suffers both absorption and diffraction in this "lossy" region, preventing the formation of a resonant cavity. Figure 2 shows a SLED on a miniature optical bench, with its emission coupled into a miniature GaP lens, mounted on a metal LIGA alignment structure. 

SLED light sources are available commercially from a number of vendors (35) in the traditional telecommunications region of the spectrum (O-, C-, and L-bands), and also can be fabricated for use at longer wavelengths. A typical SLED has an emission width of 60–100 nm; therefore, they fall into an intermediate area in which they are not broadband enough for an individual SLED to be a general purpose source, nor are they narrow enough to be a useful spectroscopic sensor on their own. An additional wavelength-selective device is required, and Axsun Technologies' (36) first analytical spectroscopic instruments were "tunable-SLEDs" (T-SLEDs), coupling a SLED to a MEMS Fabry-Pérot tunable filter, and based partially on their earlier telecommunications optical channel monitor (OCM) products (37–40).  

The optics for the T-SLED are on a 14-mm long miniature optical bench, itself contained in a 14-pin, chip-sized butterfly package (Figure 3). This level of integration is made possible by a number of MEMS components and techniques: SLED sources, miniature lenses made by a mass-transport process, a MEMS Fabry–Pérot tunable filter, broadband high-reflectivity coatings, LIGA-fabricated optical mounts, and a robotic active-alignment process (37). One key step is reducing the problem of miniature optical assembly to one that has been solved already: placement and laser soldering of miniature electronic components onto printed circuit boards using pick-and-place machines. 

The MEMS tunable Fabry–Pérot filter (Figure 4) consists of two mirrors, facing each other and separated by a few micrometers. Changing the mirror separation by applying voltage to the MEMS structure tunes the transmitted wavelength (1). Because of the extremely small size and low mass of the movable mirror, the mechanical resonant frequency of the filter is over 100 kHz, and it can be scanned physically over its entire range in less than 50 μs (36,37). These optical filters address the issues raised in a 2004 paper (41) that commented that the resolution (expressed as 

 = Δλ/λ) of a miniature Fabry–Pérot filter would be limited to ~30; these spectroscopy designs commonly achieve ~1500, and telecommunications designs achieve considerably higher than that. 

The T-SLED spectrometer concepts were developed further to make a broadband tunable laser. A laser typically consists of a cavity formed by two mirrors, containing a gain medium and a wavelength-selective element. Construction of this laser was achieved by taking the optical components described for the T-SLED spectrometer, but substituting a semiconductor optical amplifier (SOA) for the SLED, using the tunable Fabry–Pérot filter in the cavity as the wavelength-selective element. An SOA is pumped along its full length, in contrast to a SLED, and light is emitted from both ends. With a typical SOA, lasing is achieved over a ~200 nm range, with a resolution of ~1 nm and a power of ~1 mW. They offer three ranges (1350–1550 nm, 1550–1800 nm, and 1800–1970 nm), and these can be combined in a single spectrometer to cover 1350–1970 nm.  

Axsun has delivered spectrometers for a number of process pharmaceutical applications, including active pharmaceutical ingredient drying (42) and blend monitoring (43), while both Axsun and Polychromix have developed portable dedicated instruments for carpet and plastic recycling (44,45). Process applications are demanding (46,47), and miniature instruments (48) have some potential advantages in that environment. In particular, the resonant frequency of a MEMS device can be very high, in the tens of kilohertz, implying that they will be unaffected by the usual low-frequency vibrations (tens to hundreds of hertz) encountered in a process environment. In the case of pharmaceutical blend monitoring (49), a battery-powered spectrometer is mounted on the lid of the rotating (10–15 rpm) intermediate bulk container, viewing its contents through a flush-mounted sapphire window. It communicates via wireless, and obtains a 2-nm-resolution spectrum, with noise less than 0.04 mAU, in about 400 ms on every rotation as the container passes through its inverted position.  

Recycling of carpet fibers (50), plastics, and polymers, while in existence for some time, has seen significant growth driven by a combination of increasing regulation, escalating disposal costs, rising feedstock prices, and the rapid expansion of economies in China and India. Typical carpet fibers are polyethylene terephthalate (PET), polypropylene (PP), Nylon 6 (N6, polycaprolactam), Nylon 66 (N66, formed from hexamethylenediamine and adipic acid), and wool; these can be recycled if the facility can identify and separate these materials rapidly. In recent years, many outlets have been developed for these recycled resins; for instance, polypropylene is used in leachfield drainage chambers (51) and "plastic lumber" products such as fence posts, decking, and railroad ties (52), while Nylon 6 can be depolymerized back to caprolactam (50). Carpet recycling, therefore, has multiple benefits: less material in landfills, lower consumption of petroleum feedstocks, and economically viable production of industrial and consumer products (53). Identification of these materials typically is performed in a sorting facility by unskilled personnel, and frequently in a hot, humid, dusty industrial environment. This market has been addressed by both Axsun and Polychromix, with dedicated, battery-powered, MEMS-based, portable spectrometers (44,45). These are "material identification" instruments — their primary output is not the display of a spectrum, but the name of the material, via an audio output, as well as visual — and are the first widespread applications of analytical MEMS-based NIR spectrometers, with about 200 instruments deployed in 2007. 

MEMS optical scanning technology enables rotation of miniature gratings (Figure 5) using a comb drive, and thus a miniature scanning spectrometer (Figure 6). This concept has been realized (54,55), and a series of spectrometer "engines" has been described (56). The 2 mm × 2 mm grating, with 500 lines per mm, is fabricated by deposition of aluminum on the silicon substrate of a scanning mirror, followed by plasma etch through a photo mask, and the complete scanning grating assembly is placed in a standard 14-pin butterfly package. The scanning grating is resonantly driven; that is, it oscillates at its natural resonant frequency, and this also means that its motion approximates that of a simple harmonic oscillator, with a sinusoidal velocity profile. The spectrometer covers the range 1.2–1.9 μm with a resolution of about 10 nm, and a single-element InGaAs detector is used. The size of the complete engine (with electronics and so forth) is 100 mm × 80 mm × 75 mm.

A conventional NIR fixed-grating spectrograph (2), equipped with a linear InGaAs array detector, enjoys a multichannel advantage, collecting every resolution element in the spectrum simultaneously, and 512-element arrays are now common. However, this MEMS-grating spectrometer scans sequentially, ceding that multichannel, and thus signal-to-noise ratio (S/N), improvement. Its potential advantage is in lower-cost fabrication, essentially replacing a relatively expensive InGaAs array detector with a potentially low-cost MEMS scanning grating in applications in which the reduced S/N (for a given measurement time) is acceptable.

Two-dimensional NIR array detectors, based upon InGaAs, indium antimonide (InSb), and mercury–cadmium–telluride (MCT or HgCdTe), have been available for many years and, in analytical applications, are employed widely in spectroscopic imaging (57). Recent developments in cameras, and in the integration of MOEMS devices with cameras, might impact NIR spectroscopy.

Researchers at the University of Western Australia (Perth, Australia) have described the integration of MEMS optical filters with two-dimensional MCT-array detectors (58–62). This is based upon fabrication of MEMS-tunable Fabry–Pérot filters (Figure 7) (63) over each pixel in the MCT array, in a monolithic structure, and the goal is to produce low-cost, single-detector spectrometers and high-density, integrated imaging arrays. These devices could have low power consumption and low weight, be easy to cool, and be available in handheld packages, and the resulting spectroscopic imager could be the same size as today's CCD cameras. However, this approach does involve several significant challenges. 

MCT arrays do not tolerate the high temperatures of many semiconductor fabrication processes as a result of the volatility of mercury. Therefore, low-temperature (<100 °C) processes for the Fabry–Pérot construction must be employed. In addition, their goal is to move to mid-infrared (MIR) wavelengths, and this means that both the Fabry–Pérot cavity length and the mirror displacement must be larger than those employed for NIR applications. Recent work (64) reports on a tunable MEMS filter monolithically integrated on an MCT detector: a fully integrated filter with a center wavelength of 1950 nm gave a full width at half maximum (fwhm) of approximately 100 nm, with a tuning range of 400 nm. 

This group also has applied this concept to a single-element detector system, the goal being a rugged spectrometer for both agricultural and petrochemical applications. Figure 8 shows a drawing of a MEMS Fabry–Pérot tunable filter mounted over a detector, with this tunable detector packaged into a standard TO-8 can (62). A preliminary device was designed to operate between wavelengths of 1.6 μm and 2.3 μm, with a 1.2 μm spacing between the Fabry–Pérot mirrors. Although a 30-nm resolution was planned, only 100 nm was achieved as a result of mirror curvature. Mirror "snap-down" can be an issue in miniature Fabry–Pérot devices. This refers to nonlinear behavior in the voltage-displacement curve, where the moving mirror is attracted to a static element and can become permanently "stuck" due to the very strong Van der Waals forces that occur on this physical scale. This group is addressing both of these issues in new device designs. 

A number of companies (including Polychromix and Aspectrics), using different technologies, are working in the ~1300–2750 nm range, covering the first overtone (O-H, N-H, and C-H) and combination spectral regions. Extended InGaAs arrays can cover to ~2400 nm, and PbS and PbSe arrays have been used at longer wavelengths. However, the use of MCT arrays has been proposed for this region by researchers at VTT, Finland (65). In contrast to the use of MCT detectors in the MIR, where they require cooling to 77 K, operation at 200 K is sufficient for a 3-μm cutoff, and this can be achieved with a three-stage Peltier (thermoelectric) cooler. VTT has developed a photoconductive linear MCT array, 128 × 1 pixels, 1800 μm × 80 μm in size, with a 100-μm pitch. This has a new readout design, which uses synchronous (that is, lock-in) detection for each of the channels, and the detector achieves a detectivity (D*) of 1.5 × 10

 . It requires a chopped light source (800 Hz was employed), which allows discrimination against ambient light, including fluorescent illumination. The complete detector package, including Peltier cooler, has a footprint of approximately 6 cm × 5 cm. VTT has developed a demonstration spectrometer, incorporating this detector, with eight parallel A/D converters, each running at 250 kHz, with the lock-in detection implemented in software running on a PC. Because of the nature of MCT material, this approach can be extended to longer wavelengths.

Some MEMS-based FT spectrometers were described in Part I of this series (2), but we will revisit that area briefly. Kraft and colleagues (66) have described a number of miniaturized spectrometers, including a spectrometer with a MEMS interferometer (67–70). This employed a resonantly driven micromirror that is suspended on two long springs, driven by interlocking comb-structured electrodes, and contained in a vacuum chamber. The effective mirror area was ~1.6 mm

. The first prototype covered a spectral range of ~4500–1450 cm

, using a thermoelectrically cooled MCT detector, at a spectral resolution of better than 30 cm

, and a wavelength stability of better than 4 cm

. It had a footprint of 180 mm × 120 mm × 120 mm and a weight of about 1.5 kg, and has now been commercialized (71), covering the 2–6 μm range at 30 cm

  resolution. However, the developers found that the moving mirror suffered from a physical distortion (72), and a "pantograph" moving mirror (Figure 9) has now been designed, which also has a larger mirror area (7 mm

). These mirrors also are driven at their resonant frequencies, similar to the scanning grating described earlier. Again, this leads to a sinusoidal mirror velocity profile, in contrast to the efforts made in laboratory FT-IR spectrometers to achieve as uniform a mirror velocity as possible (73). The limited mirror displacement available in most MEMS implementations leads to thoughts of increasing this by various means; for instance, a "Genzel interferometer" (74) uses both sides of the moving mirror to double the optical retardation for a given physical mirror displacement. Alternatively, more than one moving mirror could be used together; however, MEMS devices will not have exactly the same resonant frequencies, which could lead to dephasing of the mirror motions. The goals of the IPMS/CTR team are to have a third generation spectrometer with an optical path difference of 1 mm, capable of 1-ms time resolution, in a package that is 50 mm × 50 mm × 50 mm. 

A Hadamard spectrometer (75–77) is a multiplex device; that is, it observes more than one wavelength at a time. This translates into improved S/N (or shorter measurement times) if the measurement is detector-noise limited. Given the brightness of NIR sources, and the low noise levels of NIR detectors, many NIR measurements are shot-noise limited, not detector-noise limited. However, detectors are noisier when the coverage extends to longer wavelengths, for instance beyond the 1.7-μm cutoff for "standard" InGaAs material. In addition, light levels are low, and large area detectors are beneficial for diffuse reflectance measurements in the combination region. In those cases, the measurements can be detector-noise limited, and therefore a multiplex technique is advantageous. 

A Hadamard spectrometer uses one or two masks, instead of slits. Harwit and Sloane (75) gave a detailed analysis of the possible types of Hadamard spectrometers and their theoretical S/N advantages. Because new technologies are enabling a new generation of Hadamard spectrometers, a highly simplified summary will be given here. Figure 10 (top) shows a version of a one-mask and a two-mask Hadamard spectrometer. In the single-mask design shown, light passes from the source through a sample and onto the entrance slit of a spectrograph. It is dispersed by a grating, and then the encoding mask selects 50% of the resolution elements and passes that light onto a single element detector. A typical mask is an array of zeros and ones, either allowing or not allowing the light to pass onto the detector. The position of the zeros and ones in the mask changes, and the detector is read out for each of these positions. Typically, the mask uses a cyclic "S-matrix" sequence, in which each row is obtained by shifting the previous row one place to the left, shown below for a 7 × 7 case (75): 

At the end of data collection, a simple matrix transform recovers the spectrum from the collected data. Typical commercial spectrometers use 128 resolution elements. In a single-mask system, the theoretical S/N gain is ½√

 is the number of resolution elements; for 128 elements, this equals ~5.6. In NIR, one practical point is that this advantage can be gained using a low-cost, single element detector, as opposed to a relatively expensive array detector. 

Figure 10 (bottom) shows a simplified dual-mask implementation. In this case, an entrance mask is used, giving a throughput advantage over an entrance slit, and a two-dimensional array detector acts as the second mask. In a 1993 paper (78), this design was applied to the weak emission from applications such as aurora spectroscopy, which were detector-noise limited, even with a CCD detector. The authors noted that the technique assumes that the coded aperture mask (entrance aperture) is illuminated uniformly, at least in the dispersion direction: this is a key caveat with important sampling consequences for analytical spectrometers. Today, the combination of array detectors and MOEMS devices enable the fabrication of doubly encoded Hadamard spectrometers in many spectral regions (79). A Hadamard-type, doubly encoded spectrometer design has been described recently (80–82), using a stationary mask, and having applications for weak signals (for example, Raman spectroscopy [83,84]) or large diffuse sources (85–87). At present, this technique is being applied mostly in Raman spectroscopy and will not be discussed in detail here. 

Returning to single-mask designs, the spectrometer from Aspectrics (88) was described in Part I (89), and this can be thought of as a Hadamard device using a sinusoidal mask, as opposed to a binary (on/off or zero/one) mask; alternatively, it can be thought of as a Fourier transform spectrometer using a rotating disk modulator instead of a Michelson interferometer. In the NIR, this spectrometer is configured to cover 1375–2750 nm, with 128 channels, using a quartz halogen source and a single-element PbSe detector. It can acquire 100 scans per second. 

Polychromix (90) has produced a handheld NIR spectrometer, using a MEMS chip as the Hadamard encoding device (91). Figure 11 shows the single-mask-style optical design. The Hadamard mask is a programmable MEMS diffraction grating (92–94), originally developed as the key element in a programmable correlation spectrometer for remote detection in the MIR (3–5 μm), with the ability to create "synthetic spectra" (95). The MEMS grating consists of a parallel set of reflecting ribbons, with individual rectangular pixels 20 μm wide and 1 cm long, which can be deflected vertically while remaining planar (Figure 12). With the telecommunications boom, this was adapted for use in the C-band (around 1550 nm) as the key element in a reconfigurable optical add drop multiplexer (ROADM) for dense wavelength division multiplexing (DWDM) applications. With the telecommunications bust, it was then applied to analytical NIR spectroscopy. 

Part II of this series described many of the miniature optical technologies that were developed as a result of the telecommunications boom, and Part III covered conventional small near-infrared (NIR) spectrometers. Here, in Part IV, we bring those themes together and see how the massive investment in telecommunications, microelectro- mechanical systems (MEMS), and micro-opto-electro-mechanical (MOEMS) is starting to impact NIR spectroscopy.

Miniature Optical Spectrometers: The Art of the Possible, Part IV: New Near-Infrared Technologies and Spectrometers

In contrast to the mid-infrared, where Fourier transform (FT)-IR spectrometers are the standard, and the UV–vis, where silicon photodiode-array (PDA) and charge-coupled-device (CCD) spectrometers dominate, there are a plethora of different technologies employed in laboratory near-infrared (NIR) spectrometers (3–5). Stark and Luchter (6) quote a total of 60 manufacturers, and the survey by Workman and Burns (3) lists close to 100 different instruments. There are several reasons for this. First, Fourier-transform instruments do not dominate NIR spectroscopy because commercial scanning grating instruments have been highly optimized for low-resolution applications, and FT spectrometers lose their multiplex advantage in this region, as they are generally shot-noise limited. FT spectrometers do, however, retain their throughput advantage and laser-based wavelength scale referencing in the NIR. Second, both one- and two-dimensional photoconductor array detectors are available at reasonable costs, although they are expensive compared with the PDAs and CCDs used in the visible and UV regions. Finally, many other wavelength-selective devices are available, including liquid-crystal tunable filters (LCTFs) and acousto-optic tunable filters (AOTFs). Each of these approaches has their particular advantages and disadvantages. 

There are at least three distinct threads in the evolution of NIR spectrometers. Research-grade instruments for molecular spectroscopy studies of liquids and gases, characterization of optical filters, and other applications, with the ability to do kinetic measurements at single wavelengths, have evolved from the Cary 14, introduced in the mid-1950s. Low-resolution instruments, for analytical solid-phase studies of heterogeneous food and agricultural products, following the work of Norris (7), started with linear variable filters and evolved via modification of a Cary 14 (8) to scanning filter and scanning grating instruments. Handheld and portable instruments generally are used for qualitative and quantitative analysis of solid materials, and therefore follow the latter thread. Third, with the development of small NIR spectrometers and small fiber-optic probes, application areas have opened up in biomedical spectroscopy: noninvasive diagnoses (9), minimally invasive diagnoses (10), and in-vivo studies (11). Optical coherence tomography (12) (OCT) is a closely related field that employs some of the same fiber-optic and NIR source technologies.  

Laboratory NIR spectrometers have been reviewed extensively (3–5,13–18). Some of the key components and approaches will be covered here briefly, as an introduction to miniature spectrometers, but the reader is referred to those reviews for further details about these instruments. Stark and Luchter (6) noted the high diversity of NIR technologies available a few years ago, and McClure and Tsuchikawa (5) classified these spectrometers as follows: diodes (emitting diode arrays, laser diodes, and diode-array detectors); filters (fixed, wedge, tilting, acousto-optic, liquid crystal); prism; grating; FT-NIR, and Hadamard.  

For technological reasons, there is another interesting division in the NIR spectrum, between the shortwave region (~740–1100 nm) and the conventional or midwave region (~1100–2500 nm). Silicon-based detectors are sensitive in the shortwave region, while lead sulfide (PbS), lead selenide (PbSe), and indium-gallium-arsenide (InGaAs) detectors work in the midwave region. The third overtone bands found in the shortwave region are considerably simpler and weaker than those observed in the midwave region (19,20). A weaker band (lower absorption coefficient) is not necessarily a disadvantage though, because it allows long pathlengths (for example, in transmission), or long effective pathlengths (for example, in diffuse reflectance or interactance) to be employed, and the simpler bands do not prevent measurements of properties (for example, octane (21)) or broad chemical constituents such as sugars (5) from being made. These long pathlengths discriminate against surface effects in transmission cells and against the skin or peel of natural products, such as in measurements on whole fruits. For example, the low cost and small size of PDA and CCD detectors enables handheld spectrometers, operating in the shortwave region, to be used for Brix (soluble sugar) measurements on intact apples and melons (5,22).  

In this article, we will first look at the critical elements of an NIR spectrometer (sources, wavelength-selective devices, detectors, and array detectors), multiplex techniques, and how diffuse reflectance sampling affects spectrometer design. Then we will survey existing small NIR spectrometers. 

The ubiquitous NIR source is a quartz-halogen light bulb, which operates around 3000 K (23), and whose emission peaks in the near-infrared; these are low-cost, reliable, very bright sources. Light-emitting diodes (LEDs) are an example of a semiconductor light source (24), and we will look at these in some detail because semiconductor light sources of various types show promise in NIR spectroscopy, especially for small, portable, and dedicated-function instruments. Historically, a small number of instruments (4,5,18,25) have used LEDs as sources, typically in the shortwave NIR region, the original telecommunications range. Because of their wide bandwidth (>40 nm), they have been used in conjunction with narrow-band filters. 

Schubert has covered the history, technology, and applications of LEDs (26). NIR (870–980 nm) LEDs were developed starting in the 1960s, using GaAs, followed by visible LEDs using GaAsP, deployed first as indicator lights on printed circuit boards. Yellow and green LEDs were developed in the 1970s, using materials such as GaP. In the late 1960s and 1970s, market drivers for these devices were indicator lights and numeric displays, in items such as telephones, calculators, and digital watches. Resonant cavity LEDS (RCLEDs) were developed in the 1990s using AlGaAs/GaAs materials; these have high infrared and red emissions from microcavities and are used widely in remote controls and telecommunications. Blue GaInN/GaN LEDs came on the scene in the 1990s, and today AlGaInP is the dominant system for high brightness LEDs used in traffic lights and automotive lighting. In telecommunications, AlGaAs/GaAs devices emit at 870 nm, and with the switch to the 1300-nm region, InGaAsP/InP LEDs were developed. Infrared LEDS are available commercially at wavelengths as long as 7 μm, based upon GaSb systems (27); however, their full width at half maximum (FWHM) increases at longer wavelengths and is around 120 nm at 1600 nm and almost 1 μm at 5 μm. Typical shortwave NIR wavelengths available are 940, 970, and 770 nm, with halfwidths (FWHM) of about 40–60 nm.  

White LEDs were developed in the mid- to late 1990s and usually are based upon a blue LED combined with a longer-wavelength phosphor; for instance, a GaInN/GaN LED and a cerium-doped yttrium aluminum garnet (Ce

:YAG) phosphor. These have been used as miniature broadband visible sources in some experimental studies (28,29). Their output is far from flat; there is typically a "double hump" — one at short wavelengths (~450 nm) corresponding to the output of the blue LED and a second at longer wavelengths (~550 nm) due to the phosphor. 

New mass-market applications (for example, camera autofocusing) for LEDs emerge rapidly, so that there is continuing development in this area. Today, LEDS are widely available at low cost from a variety of manufacturers (30), and the worldwide LED market was estimated at $4.2 billion in 2006, growing to $11 billion in 2011. Progress in solid-state sources is very rapid, and we shall see in Part IV the use of superluminescent light-emitting diodes (SLEDs), leading to the development of broadly tunable NIR lasers (31). 

Given the many review articles on conventional NIR instruments (3–6,14–18), this paper will not describe the wide variety of wavelength-selective devices available in the near-infrared. Instead, a few of the technologies that could yield, or have yielded, handheld devices will be discussed briefly.  

Fixed-filter NIR analyzers continue to be sold into a large number of process applications, ranging from gases (CO

, and so forth), to moisture in almost anything: building materials, foods, textiles, paper, and tobacco (32). Wetzel and colleagues (15) estimated that 85% of all NIR analyses are performed with fixed-filter instruments, but tilting-filter spectrometers have largely disappeared from the marketplace. Two tunable-filter technologies have been used: acousto-optical tunable filters (AOTFs) and liquid crystal tunable filters (LCTFs). These also have been reviewed in detail by Wetzel and colleagues (15). These two technologies have found different niches: AOTFs for scanning NIR instruments, and LCTFs for imaging instruments, both in the NIR and visible, especially in fluorescence microscopy. Commercial instruments based upon AOTFs are available, including a number of portable systems (33–35). As an example, one of these (Infrared Fiber Systems' PlastiScan-C) covers 1200–2600 nm, with an optical head size of 9 in. × 6 in. × 2 in., weighing 2.1 lb. This spectrometer is run from a PC via USB. 

Tunable filter technologies developed for telecommunications include linear variable filters and Fabry-Pérot tunable filters with a variety of tuning mechanisms. Telecommunications Fabry-Pérot tunable filters are described later, while linear variable filters (LVFs) were described in Part I of this series (2), and have been reviewed by Wilks (36). LVFs were the key component in one of the first miniature spectrometer "engines" to be seriously marketed to system integrators and instrument companies, as opposed to endusers. This was the MicroPac from Optical Coating Laboratory, Inc. (now JDSU [37]); it was described by Coates (38) and featured prominently in an early review of miniature spectrometers in 2000 (39). It was 0.63 in × 1.4 in. × 1.4 in. in size, and was offered, as an engine, for $799, in two versions: covering 400–700 nm with 5 nm resolution, or 600–1100 nm with 7 nm resolution. At that time, OCLI had very ambitious plans for this device, including driving the cost down so that it could be included in every color printer and copier as a measurement and correction sensor. This vision did not come to pass, but the concept of using an LVF and an array detector in a very low cost spectrometer survives today in the mid-infrared instrument from Wilks Enterprise (40) and the recently introduced series of visible and shortwave NIR spectrometers from microSpectralSensors (41,42). 

The responsivity of a particular detector usually is defined in terms of the noise equivalent power, or NEP (43). Simply put, this value is the power incident on the detector so that a signal-to-noise ratio (S/N) of unity is achieved. NEP is dependent upon the area of a detector, with larger detectors having larger noise. NEP is measured in units of W Hz

, and is defined formally for a particular modulation frequency, wavelength, and effective noise bandwidth; the lower the NEP, the better the detector. To compare detector material, independent of area, the specific detectivity, or D* (D-star), is defined as: 

. The higher the D*, the better S/N the detector will have. Formally, the D* is the S/N at a particular electrical frequency, in a 1-Hz bandwidth, when 1 W of radiant power is incident on a 1-cm

Scanning grating and filter-based systems traditionally have used photoconductive lead sulfide (PbS) detectors (44), with a range of 1–3.5 μm, a D* of about 5 × 10

, and a time constant of about 1 ms. Photoconductive lead selenide (PbSe) detectors have a longer wavelength cut-off (~4.8 μm), a typical D* of ~2 × 10

, and a time constant of ~2 μs. The telecommunications industry has driven the development of InGaAs PIN photodiode detectors because of their high sensitivity and low noise in the 1.3- and 1.55-μm regions, and their very fast response times — as low as 1 ns. These detectors are now used widely in FT-NIR and other fast data rate spectrometers. A PIN photodiode is a semiconductor positive–negative (P-N) structure with an intrinsic (I) region sandwiched between the other two regions (45).  

InGaAs detectors typically have a range from 0.8 μm to 1.7 μm, but extended InGaAs detectors are available with typical cut-off options of 1.9, 2.2, 2.4, and 2.6 μm (46). The D* for these detector materials drops steadily as the wavelength cut-off is extended, and typical room temperature values are 5 × 10

, and measured at 1 KHz. Large-area (up to 5 mm diameter) InGaAs detectors also are available, and these find applications in diffuse reflectance and integrating sphere applications. In the short wavelength, or third overtone, region of the near-infrared, silicon PIN photodiode detectors are used. These have a range of 300–1100 nm, a D* of about 10

, and a time constant of a few nanoseconds. 

In the shortwave NIR region, shorter than 1100 nm, silicon PDAs and CCDs are used. Stark (16) points out that while CCDs are used for low light level work, for instance in Raman spectroscopy, PDAs are used for analytical NIR transmission and reflection. Shortwave array detector costs have been driven down by widespread commercial applications such as cameras, cell phones, and scanners. There are many commercially available, compact, midwave NIR instruments, based upon spectrographs containing InGaAs linear array detectors (16). A few instruments employ lead sulfide and lead selenide linear arrays (47–49), with up to 256 elements. Detector arrays usually are specified differently from single-element detectors because of their different readout and amplification. A typical linear array has between 128 and 1024 detector elements, or pixels, and these are typically 25–50 μm wide and 50–500 μm high — designed for use with slit-based grating instruments. The pixel spacing is referred to as the pitch, and efforts are made to minimize any gaps between pixels. The ratio of total pixel width to length of the array is termed the "fill factor," and for modern arrays, this is close to 100%. One or more analog-to-digital (A/D) converters typically are located on the array package itself. These are commonly 14-bit, but recently, some 16-bit converters have been used (50). The key parameters that govern the S/N of an array are its dark current, the current through the detector pixel under normal operating conditions with no incident radiant power, and its read noise. In most situations, the dark current is the dominant noise. The typical dark current in InGaAs arrays increases from 0.08 pA for a 1.7-μm array to 1000 pA for a 2.6-μm array, which is a factor of 12,500 in increased noise (46). The dark current can be reduced by cooling the array, and a recently available spectrometer (46) using a 2.2-μm array and a 16-bit A/D quotes 800 counts dark current out of 65,536 (2

InGaAs arrays with 128 pixels were employed initially, but arrays with 256, 512, and 1024 pixels are now available (51,52). Although it is logical to assume that 256 pixels covering the range of 1120–2100 nm would imply a resolution of 4 nm, the practical resolution attained is actually at much worse — around 12–15 nm. This is for two reasons: at least three points are required to define and resolve a band, and these devices are not perfect; there can be leakage or crosstalk between pixels. That is, the signal on one pixel can affect the signal on its neighbors, with possible deleterious consequences in qualitative and quantitative analyses.  

Harwit and Sloane divided multiplex spectrometers into two classes (53): the first, which uses interference techniques and Fourier transforms, and the second, which uses masks and discrete transforms. The Michelson interferometer, which delivers both a throughput and a multiplex advantage, falls into the first class, and Hadamard-transform spectrometers into the second. Part I of this series (2) described why Fourier-transform spectrometers (Michelson interferometers), with their throughput and multiplex advantages, came to dominate mid-infrared instrumentation: because source temperatures are limited; suitable tunable sources do not exist; and detectors are noisy. It is important to note that the multiplex advantage applies because mid-infrared instruments are detector-noise limited; in the case of a shot-noise-limited spectrometer, the multiplex advantage is canceled by the increase in shot-noise. Hadamard spectrometers (53–55), which also have both a throughput and a multiplex advantage, have been applied in the mid-infrared, but historically, the practical construction difficulties relegated them to a footnote. For particulars of the theory of Hadamard-transform spectroscopy and a detailed analysis of noise for various designs, the reader is referred to the book by Harwit and Sloane (53). 

As multiplexing devices, Hadamard-transform spectrometers can be expected to produce higher signal-to-noise spectra than a sequential scanning instrument with the same optical dimensions and measurement time, when the system is detector-noise limited. Mid-infrared spectrometers are almost always detector-noise limited, while visible spectrometers are shot-noise limited. NIR spectrometers are an intermediate case, with instruments operating at low light levels, and with longer wavelength radiation, possibly detector-noise limited, while instruments operating at high light levels and shorter wavelengths (and with shorter-wavelength cut-off detectors) being shot-noise limited. Therefore, in some circumstances, a multiplex NIR spectrometer will have an S/N advantage; however, given the costs of NIR array detectors, there may well be a significant cost advantage in a Hadamard design.  

In a typical scanning dispersive spectrometer, light from a source and sample passes through a slit, is dispersed by a grating, and passes through a second slit onto a detector. Hadamard spectrometers replace one or both slits with encoding masks (56). In a typical single mask design, light passes through an entrance slit, is dispersed by a grating, then falls on the mask, returns to the grating, and is de-dispersed, and is then focused on a single element detector. If there are 

 points in the spectrum, then a signal-to-noise gain of 1/2√

The work done in the 1970s concentrated largely on single mask designs, because the masks used at that time were mechanical and array detectors were in their infancy. However, with the widespread availability of array detectors, construction of a doubly encoded spectrograph became possible (57). Here, one takes a conventional grating spectrograph, equipped with an array detector, and replaces the entrance slit with a Hadamard mask. The array detector acts as the second mask in this case. The spectrograph already enjoys a multichannel advantage, and the mask, with its increased area over the slit, provides a throughput advantage, potentially providing a total signal-to-noise gain of 

/4. In a 1993 paper (57), this was applied to the weak emission from applications like aurora spectroscopy, which were detector-noise limited, even with a CCD detector. The authors noted that the technique assumes that the coded aperture mask (entrance aperture) is illuminated uniformly, at least in the dispersion direction: this is a key caveat with important sampling consequences for analytical spectrometers. Today, the combination of array detectors and MOEMS devices enable the fabrication of doubly encoded Hadamard spectrometers in many spectral regions (58). 

The Implications of Diffuse Reflectance Sampling

Analytical NIR spectroscopy has been dominated by the study of natural products, especially in the food and agriculture industries (59). This has some very important implications in spectrometer and accessory design. These samples are heterogeneous, and therefore, to obtain an average spectrum of the sample, large sample sizes are required. Further, because these samples are not pure materials, and because each chemical constituent might be present in a wide variety of environments, the spectral features in these samples are very broad, typically many tens of nanometers halfwidth. Pure materials (hydrocarbons, fine chemicals, pharmaceuticals, crystalline materials, and synthetic polymers) have much narrower bands in the near-infrared, especially in the combination and first overtone regions. 

The broad features from natural products enabled grating instruments to be designed for these analyses with 10-nm resolution (7), and the much lower resolution optical-filter-based instruments to work well for moisture analysis, for example. These grating instruments, therefore, have much greater throughput than a high-resolution (~1 nm) laboratory spectrometer designed for molecular spectroscopy. When a solid material is examined in diffuse reflectance, as is commonly the case, the sample itself essentially becomes the source — and a large area, diffuse, dim source at that. This is because the sample scatters the incident light over a 2π solid angle. The problem is now to throughput match this large area source to the rest of the optics in the spectrometer. Therefore, the use of a large detector area is a distinct advantage. This is straightforward in predispersive (wavelength selection before the sample) spectrometers and Fourier-transform spectrometers. The solution to this problem devised by Norris (7) was to use a predispersive spectrometer with a very compact sampling accessory, equipped with two large area detectors inclined at 45° to the optical axis, and instruments using this concept are still available today (60). For this type of measurement, modern FT-NIR spectrometers use an integrating sphere, equipped with a large-area extended InGaAs detector. 

A spectrometer based upon a spectrograph and a linear array detector is, by definition, a postdispersive instrument. In an array-detector–based spectrograph, the individual pixels now become the limiting aperture of, and define the limiting angle in, the optical system. The trend in linear array detector development has been to increase the number of pixels and to decrease their size. There is an obvious mismatch between the detector pixel area with its acceptance angle, and the area of the sample (that is, the effective source) and the solid angle over which it reflects. This makes linear diode-array spectrometers fundamentally less suited to diffuse reflectance measurements than predispersive instruments. 

Some of the approaches to, and issues with, miniature spectrometers have been reviewed (61,62). In this section, we will describe some of the existing small or miniature spectrometers, fabricated using conventional techniques, before examining the impact of new technologies in this field in Part IV.  

Rosenthal has reviewed the use of LEDs as sources in small, portable NIR spectrometers, but focused on a few products only (18); if the definition of a spectrometer is broadened to include specific analyzers and sensors, then it is apparent that LED-based spectrometers are used routinely today in applications as varied as gas sensing, alcohol in wine, and blood oxygen saturation.  

In the products that Rosenthal reviewed, he noted that, because of their large bandwidth and the variability of LED sources, a bandpass filter, ~12 nm width, is placed in front of each LED. Although these are not true spectrometers, they are NIR analyzers, and LEDs have some advantages, especially in the construction of a portable, battery-powered device. They have low power consumption; a long life; enable a no-moving-parts system because they can be switched on and off rapidly; are "bright" sources (high power output per unit area); are optically efficient when their light output is confined to a narrow range; and because they are discrete wavelength sources, they enable the use of a low-cost, single-element detector in the analyzer. However, Rosenthal also notes that, "As of this writing, all IRED-based commercial NIR analyzers are individually calibrated." This is due to the variability of commercial LEDs and the lack of a suitable internal wavelength and power calibration system in these analyzers. However, he notes that, using short-wave NIR LED technology, it was possible to develop an NIR analyzer for body fat analysis with a sale price of less than $100, and that 100,000 of these were sold. In the 1990s, with longer wavelength, smaller, and more powerful LEDs, Zeltex (63) developed a portable octane analyzer that incorporated as many as 39 LED–filter sources. These instruments do highlight the possibilities of using semiconductor light sources in spectrometers, and taking advantage of telecommunications technology. 

Two 1998 papers surveyed the possibilities at that time for LED-based spectrometers for handheld sensors and process monitoring (64,65). They highlight the change in an LED's output, both power and wavelength shift, as a function of temperature, and means to compensate for that. They also describe visible-NIR LED-based device designs for the pulp and paper industry, and a prototype shortwave NIR spectrometer with LEDs operating at 850, 900, 935, 950, and 1020 nm. 

In 1995, Fateley and colleagues (66) described a different concept for an LED-based spectrometer. This consisted of a number of LEDs, each of which could be modulated at a different frequency. In this way, the intensity of each channel could be recovered via a Fourier transform. However, there will come a time when all the diodes are "on" simultaneously, and therefore, the detector signal will show a "centerburst." This can lead to dynamic range problems, and so the authors proposed a Hadamard-type scheme in which the modulation frequencies for each diode are the same but their phases are different, which eliminates the centerburst (67). 

McClure and colleagues (68) described a handheld NIR spectrometer for measuring moisture and chlorophyll. This used three unfiltered LEDs as sources, emitting at 700, 880, and 940 nm, was powered by four AA batteries, measured 10 cm × 19 cm × 5 cm, weighed 364 g, and cost around $300 to produce. The emphasis in this work was to produce a device for some basic, but critical, agricultural measurements, that would be affordable in developing countries. 

An instrument for measuring alcohol in wine from Anton Parr (69) uses an LED source and a small detector array. It operates between 1170 nm and 1200 nm (70): the C-H stretch second overtone region. Light from an NIR LED passes through a sample cell (with a Peltier thermostat control), then through a slit, is dispersed by a grating, and falls on a nine-element array detector. A relative baseline can be determined using the water-alcohol isosbestic points in this region. There are no moving parts in the spectrometer, and this is another example of a dedicated spectroscopic analyzer using an LED as a light source. 

A miniature carbon dioxide sensor from Vaisala (71) employs an LED light source, transmitting light through a miniature gas cell, and a miniature Fabry–Pérot interferometer, and onto a detector. The Fabry–Pérot can be set to transmit two wavelengths: one for a baseline and the other on a carbon dioxide absorption band. The company's data sheet claims that the sensors are so stable that they only require a calibration check every five years. A complete sensor on an electronics board measures 96 mm × 60 mm; the sensor module itself, including the gas cell, is approximately 20 mm × 20 mm. 

The pulse oximeter (72) also employs LEDs as light sources. Human blood contains oxyhemoglobin (HbO

), reduced hemogloblin (Hb), and usually small quantities of methemoglobin and carboxyhemoglobin. In a clinical situation it is important to know the oxygen saturation in arterial blood, defined as the oxygen content divided by the oxygen capacity. HbO

 absorbs strongly in the 900–1000 nm region and weakly in the 600–700 nm region, and the converse is true for Hb. The complication in this analysis arises from the human pulse; this causes the arteries to expand and contract, altering the pathlength under study. However, this pulsation also produces a modulated spectroscopic signal, and this is the key to the elegance of the modern pulse oximeter. Two LEDs typically are used, at 660 nm and 910 or 940 nm, and a photodiode detector. Four signals are measured, essentially the constant (or DC) component at each wavelength and the pulsatile (or AC) components. The AC component is only due to the change in absorbance due to HbO

 is calculated, which is therefore proportional to the oxygen saturation: 

where NIR is the signal at 910 or 940 nm. A stand-alone pulse oximeter can now be purchased by the public (on 

, for example), and costs around $200. It is therefore an excellent example of photonic components being used in a miniature, dedicated function analyzer, in a high-volume, low-cost market. Because it is a small, low-cost device, it can be integrated with other clinical analyzers (such as cardiac defibrillators), deployed in the field (by monitoring firefighters and first responders), and wirelessly networked for remote monitoring. Frost and Sullivan (73) estimated the market for pulse oximeters for the U.S. alone at $600 million in 2006, or hundreds of thousands of spectrometers. This dwarfs the unit sales of laboratory spectrometers. 

In Part I of this series, we examined recently developed miniature mid-infrared spectrometers (1). In Part II, we surveyed micro electro mechanical systems (MEMS), micro-opto-electro-mechanical systems (MOEMS), and some of the photonics technologies developed for optical communications (2). Here, in Part III, we summarize some of the conventional approaches to miniaturizing near-infrared (NIR) spectrometers, and in Part IV, we will bring these themes together and see how MOEMS and telecommunications photonics are poised to revolutionize NIR spectroscopy with a new generation of miniature instruments.

Miniature Optical Spectrometers, Part III: Conventional and Laboratory Near-Infrared Spectrometers

In Part I of this series (2), we saw that there are straightforward motivations for miniaturizing an optical spectrometer. If an instrument can be made smaller, it will often also consume less power, enabling it to be portable and eventually handheld. This allows the spectrometer to be taken to the sample, as opposed to taking the sample to the spectrometer, and this typically leads to increased productivity in traditional applications. Small size often also implies lower cost, and this opens up nontraditional applications. In addition, the recently developed technologies that enable miniature spectrometers can be regarded as disruptive, and instruments based upon those technologies find niches at the low-price end of the market, in new applications, where their performance is "good enough." However, because the performance of these newer products typically progresses faster than their mature counterparts, they rapidly become competitive with these mature products and eventually meet the performance requirements of the majority of the market. This scenario can lead to the demise of the previously dominant products and technologies.  

Part I of this series addressed the mid-infrared (MIR) region, where Fourier-transform (FT) spectrometers dominate the scene, and small versions of laboratory instruments have been developed (2). However, several other technologies are being implemented for niche applications in the MIR, including slab-waveguide spectrometers, lower cost array detectors, linear variable filters, tunable lasers, and photonic crystals (2). A number of FT-IR spectrometers using micro electro mechanical systems (MEMS) technologies also have been developed (2). The performance of spectrometers is highly dependent on their optical throughput, especially in the MIR, where sources are less bright and detectors are poorer, as compared with the NIR region. A miniature spectrometer implies low throughput, and this potentially compromises the applications of miniature instruments. 

In this article, we will examine the technologies developed during the telecommunications boom of the late 1990s, focusing on miniaturized optical techniques generally called micro-opto-electro-mechanical systems (MOEMS) (3). Some of these technologies only began to be applied to spectroscopy a few years ago, and we have yet to see their full impact. Part III of this series will survey conventional small NIR spectrometers, and Part IV will describe new miniature NIR spectrometers based upon MOEMS and telecommunications photonic technologies. 

In the late 1990s, a vision of an annual telecommunications equipment market of $110 billion caused around $20 billion to be invested in optical communications companies, spawning a multitude of developments (4). These devices are all described in standard texts on photonics (5), and include pump laser diode modules, semiconductor tunable lasers (6), optical gain media and fiber amplifiers, optical fibers, optical switches, detectors, gratings, fiber Bragg grating sensors, photonic crystals, and tunable filters. The size of many of these components is particularly interesting, because fiber optics used in communications today have a core diameter of ~8 μm and operate in single mode. The small size of these single-mode fibers in turn mandates that all the other optical components be on that scale; therefore, a huge industry grew up that involves devising and manufacturing miniature optical components. Furthermore, telecommunications applications require that devices be incredibly reliable and rugged and operate under a wide range of temperature, vibration, and humidity conditions; those properties make them very attractive for many applications outside telecommunications. The massive amount of investment in this market also spurred intense competition, resulting in very low prices for these components, especially after the dot-com and telecomm bubbles burst about five years ago. After that market crash (4), there was interest in applying these technologies in other applications. Even after the market crash and resulting business consolidation, developments in optical communications and communications technology continue, and the long-awaited deployment of fiber-to-the-home (FTTH), now with an installed base of 2 million subscribers in the US (7), and increasing bandwidth demand from downloadable music and video, are spurring new investment. 

Because of the nature of optical communications, most of these telecomm devices work in the NIR region. These developments are now fueling a new generation of small NIR spectrometers; to understand those developments and anticipate the next generation of instruments, we need to follow Deep Throat's advice to Bob Woodward (8) and "follow the money" — follow all the developments in telecommunications, photonics, and MOEMS, funded by those billions of dollars of investment. 

MEMS processing technologies were developed during the 1980s in response to a demand for rugged, miniature sensors, and during the 1990s micro-optics technology was developed (9), both in industry and at national laboratories; Motamedi has given a historical outline (10). Key applications were optical switches for telecommunications, components for wavelength division multiplexing (WDM), digital mirror devices (DMDs), and optical scanners. MOEMS and optical communications have their own terminologies, abbreviations, and acronyms, and we will describe a few of them here. MEMS techniques are widely used in nonspectroscopic areas of analytical chemistry, principally in microfluidics, but those applications (11) will not be covered here. 

A microprocessor chip has all-electrical connections and no internal moving parts. A MEMS chip has, by definition, moving parts, and a MOEMS chip adds an optical interface as well. Given that very small contaminants can cause a chip function to fail, packaging (12,13) — the placement of the chip in a dry, clean, hermetically sealed, but interconnected, environment — is a critical issue. Telecommunications MOEMS devices typically employ optical fibers and fiber feed-throughs in and out of device packages, so that many devices (such as pump lasers and filters) are referred to as fiber-pigtailed, while MEMS display devices require optical windows in the package. Devices and other materials in a package tend to outgas, and this can cause performance problems; therefore, getters (reactive materials used to remove traces of gas) are included within the packages. Packaging is not widely described in the literature; instead, it is frequently regarded as a proprietary technology, but interested readers can refer to references 12 and 13.  

In process control terminology, an actuator is a powered mechanism that causes motion of a part. The motion of many MOEMS devices is enabled by a comb drive actuator, silicon-based linear motors, which operate using electrostatic forces. Figure 1 shows a comb drive developed by Sandia National Laboratories (Albuquerque, New Mexico), and the interdigitating teeth in the combs can be seen clearly; the teeth are designed so that they can slide past each other. A potential difference between the two combs causes electrostatic attraction between them. The magnitude of this force depends upon the voltage difference, the number of comb teeth, and the length of the teeth; the force decreases when the combs are further apart. 

A large class of MOEMS devices are optical scanners (14), which can control the position of an optical light beam; these find a multitude of applications in imaging, barcode scanners, and printing. Optical scanners use movable mirrors that move around one or two axes: motion of a mirror, or an array of mirrors, can manipulate light by reflection or by diffraction. Replacing the mirror in an optical scanner with a MEMS grating leads to a miniature scanning grating device, which clearly has spectroscopic applications. Very complex MEMS and MOEMS devices have been fabricated using Sandia National Laboratories' SUMMiT V process (15): a five-layer polycrystalline silicon surface micromachining process, with one ground plane–electrical interconnect layer and four mechanical layers. In spectroscopy, this process is being used for the Block Engineering (Marlborough, Massachusetts) MEMS FT-IR program (16). As examples of Sandia's MOEMS technology, Figure 2 shows a gear reduction system, and Figure 3 shows a mechanism that erects a silicon mirror. 

Optical Fiber Communications Technologies

Saleh and Teich (17) give a condensed history of the evolution of optical fiber communication systems. At a very basic level, efficient optical communication requires powerful light sources (LEDs and lasers) that can be modulated rapidly, and then fast-response detectors for the light wavelengths employed. In the 1970s, the first systems operated around 870 nm, using AlGaAs LEDs and laser diodes, with silicon positive-intrinsic-negative (PIN) diode and avalanche photodiode (APD), detectors. However, at longer wavelengths, both fiber attenuation (that is, absorption) and dispersion (18,19) (change of refractive index with wavelength) in silica fibers are smaller, and the next systems were designed for 1310 nm operation ("O-band"), which is the point of minimum dispersion. Dispersion causes light at different wavelengths to travel at different speeds in the fiber; this broadens pulses of light, and ultimately limits the information transmission rate. These O-band systems used InGaAsP lasers with InGaAs PIN or APD detectors. Silica fibers have their lowest attenuation (but nonzero dispersion) around 1550 nm ("C- and L-bands"). The dispersion issue was addressed using graded index optical fibers, distributed feedback (DFB) lasers, based upon InGaAsP, and again InGaAs detectors were used. Over distance, optical communication signals require periodic amplification, and this has been done with semiconductor optical amplifiers (SOAs) and optical fiber amplifiers (OFAs), the most common of which is the erbium doped fiber amplifier (EDFA). Today, many telecomm lasers are vertical-cavity surface-emitting lasers (VCSELs) (20), which are multi-quantum well devices, based upon GaAs–InGaAs, and light is emitted from their top face. They can be fabricated in the form of arrays, with individual elements emitting at different frequencies. 

To carry more information down a single optical fiber, many channels, operating at different optical frequencies in the "C-band" (1530–1565 nm) and "L-band" (1565–1625 nm) regions, are used (21,22). This is termed wavelength division multiplexing (WDM), and dense wavelength division multiplexing (DWDM). DWDM systems use channel spacing as small as 50 GHz, which, in more familiar spectroscopic units, is 1.66 cm

  (everywhere in the spectrum, because they are both frequency units), and at 1550 nm this corresponds to ~0.4 nm. Clearly, means are required to separate and detect a large number of closely spaced wavelengths, requiring the development of miniature high-resolution wavelength-selective devices: gratings, variable filters, interferometers, and rapidly-tunable filters, as well as linear detector arrays.  

Miniature lenses are also required, and one means of accomplishing this is the use of graded index materials: GRIN lenses. In these devices, the refractive index of the material is varied gradually to bring light to a focus. These lenses are manufactured with a small cylindrical form factor, with a radial index gradient, and are self-aligning. They are often referred to under their trade name of SELFOC lenses (NSG America, Inc., Somerset, New Jersey) (23). A typical diameter is 1–2 mm, and length is a few millimeters. Miniature lenses and lens arrays have been fabricated by a number of other techniques (9,24,25), and one interesting method for materials such as GaP and InP is mass transport. In this technique, a lens preform with the desired profile expressed digitally at low resolution is etched in the material and is subsequently annealed; the material has a finite vapor pressure at elevated temperatures and redistributes to form a smooth profile. Figure 4 shows a mass transport lens preform, and Figure 5 a finished array of GaP lenses. 

WDM and DWDM require an optical multiplexer to mix the signals together and a demultiplexer to separate them (5), as well as devices to measure the power in each channel — optical power monitors (OPMs) or optical channel monitors (OCMs) — and subsequently rebalance those powers, using a variable optical attenuator (VOA). A number of MOEMS devices have been developed for use as VOAs to control optical power in networks, cross-connects, and add–drop multiplexers. We shall see in Part IV of this series how Axsun Technologies' (Billerica, Massachusetts) NIR spectroscopy products evolved from their OCM business. Finally, there is a need to switch optical network traffic remotely, without converting optical signals back to electrical signals: the key MOEMS component is called an optical cross-connect, which has the ability to switch N input channels to N output channels, in any configuration (26). The complete devices are called reconfigurable optical add–drop multiplexers (ROADMs) (27), and a number of different optical devices have been developed for this application, including 'active gratings.' One device of this type is now the key component in Polychromix's (Wilmington, Massachusetts) handheld NIR spectrometer (see Part IV). 

LIGA (28–31) is a German acronym for X-ray lithography (lithographie), electroplating (galvanoformung), and molding (abformung), and was developed at the Forschungszentrum Karlsruhe (Germany) in the early 1980s (32). An X-ray mask defining the desired part is placed over a photoresist, such as polymethylmethacrylate (PMMA), and the combined part is exposed on a synchrotron beamline, which produces an intense, highly parallel beam of X-rays (Figure 6). A well-defined hole is produced in the photoresist, with smooth, vertical sides; the roughness of the sidewalls can be less than 20 nm. The dimensions of details in the 

 direction can be as small as 200 nm, while a typical depth (

-direction) of the hole is 150 μm to 1 mm. The resultant devices, therefore, can have a high aspect ratio: the ratio of the width of details in the structure to its depth. The hole is then filled in via a metal electroplating step, typically with nickel; the PMMA is then dissolved and the part freed. In practice, of course, hundreds or thousands of parts are made simultaneously. Some examples of Axsun Technologies' LIGA optical mounts are shown in Figure 7. The photoresist holes can be overfilled during the plating process, in which case the resulting metal part can be used as a master for precision embossing or injection-molding processes, and then final parts can be made of a plastic material. Multilevel structures also can be created using LIGA, by bonding a second layer of PMMA to the first, after the electroplating process is complete. 

Texas Instruments' (Dallas, Texas) Digital Light Projector (DLP) is probably the most common MOEMS device, with 10 million units in computer-based projectors (Figure 8) shipped by 2006, after its commercial introduction in 1996 (33). The underlying technology, a digital mirror device (DMD), was invented in 1987. The design and packaging of the DLP chip are described in detail elsewhere (34,35), but a few details will be given here, because, as we shall see, the DLP chip can be used as the key component in a Hadamard transform spectrometer. The DLP chip consists of an array of mirrors that can be driven ±10° from a rest position, and these angles are determined by position stops or "landing sites," so they are very precise (Figure 9). One position is "on" and light will reach the screen, and the other is "off" and light is directed into a trap. The mirrors are aluminum-coated silicon, and are 16-μm square on 17-μm centers, leading to a fill factor of over 90%. The fill factor is high because all the hinge mechanisms are below the mirror itself. Each mirror is capable of switching on and off up to several thousand times per second. When a mirror is switched on more frequently than off, it reflects a light pixel; a mirror that is switched off more frequently reflects a darker pixel. The available resolution (number of pixels) in these devices has increased steadily over the past ten years. An office projector today is often SXGA or 1280 × 1024 resolution (1,310,720 pixels), and cinema projectors are 2048 × 1080 resolution. 

The DMD array chip is attached to an alumina substrate, which is brazed to a Kovar ring. The optical window is attached to a Kovar frame with a fused seal, and finally the window frame is welded to the seal ring, creating a hermetic package. There are also getter strips inside the package. Texas Instruments now offers the DLP chip with a choice of three windows: Standard visible, IR, and UV. The IR window transmits to about 2500 nm, has greater than 90% transmission from 900–2100 nm, and has less than 50% transmission at shorter than 500 nm (36). 

Although many different wavelength separation techniques have been used in optical communications, the dominant one is the use of Fabry-Pérot interferometers (37), devices that found few widespread commercial applications for almost 100 years (38). A Fabry-Pérot filter consists of two mirrors, either plane or curved, facing each other and separated by a distance 

. There are two basic versions: an interferometer, where 

 is fixed. The condition for constructive interference within a Fabry-Pérot interferometer is that the light forms a standing wave between the two mirrors, in which case the optical distance between the two mirrors must equal an integral number of half wavelengths of the incident light. When a Fabry-Pérot cavity is on resonance, constructive interference within the cavity allows transmission of essentially 100% of the light through the filter. When the cavity is off resonance, the Fabry-Pérot filter reflects nearly all the incident light. 

These interference effects are most familiar to MIR spectroscopists because of the fringes, or channel spectra, frequently observed when using transmission cells, measuring spectra of thin polymer films, or silicon wafers (39). These are samples with low-reflectivity, plane parallel surfaces (such as etalons), and the resulting fringes resemble a cosine wave in the spectrum. These fringes can be used for thickness measurement, and dedicated interferometric instruments have been used for many years in the semiconductor business for exactly this purpose (40). 

The difficulty of fabricating a Fabry-Pérot with highly reflecting parallel faces and a small spacing, along with its limited free spectral range, has meant that this technique has not been used for analytical spectroscopy and has been limited to a few specialist high-resolution applications. However, it is worth noting that a Fabry-Pérot interferometer, with a circular aperture, does have a throughput advantage over a grating, similar to a Michelson interferometer. This was realized by Jacquinot in his early work on high-resolution spectroscopy (41).  

Because of the widespread use of Fabry-Pérot interferometers in telecommunications and their increasing use in a number of innovative miniature spectrometers (Part IV), we will cover the theory of the Fabry-Pérot here, and contrast it with the more familiar Michelson interferometer. 

In the Fabry-Pérot interferometer, for normal incidence and an air gap, we have the following relationships: 

λ = wavelength of light resonant in the interferometer 

Therefore, changing the mirror separation (

) tunes the transmitted wavelength (λ) of a Fabry-Pérot interferometer.  

Transmission through a Fabry-Pérot filter is periodic with wavelength, and the distance between two wavelength transmission orders is the filter free spectral range (FSR): 

Alternatively, this can be stated as the frequency spacing between modes: 

Note that the FSR is a constant in frequency space, but not in wavelength space (Figure 11). 

The spectral resolution, defined as the full-width-half-maximum (fwhm) at the peak transmission, depends upon the finesse (

) of the Fabry-Pérot, which, in turn, is determined by the reflectivity (

The author examines NIR spectrometers and the technologies developed during the telecommunications boom of the late 1990s, focusing on miniaturized optical techniques generally called MOEMS.

Miniature Optical Spectrometers: Follow the Money Part II: The Telecommunications Boom

Miniature spectrometers can be regarded as a "disruptive technology" as described by Christenson (10). He gives several examples of products based upon innovative technologies whose initial performance was significantly less than that of mature and established products. He also notes that the performance of mature products could overshoot the market and give customers more performance than they really require. Innovative products found niches at the low-price end of the market, in new applications, where their performance was "good enough." But because the performance of these newer products progressed faster than their mature counterparts, they rapidly became competitive with these mature products and eventually met the performance requirements of the majority of the market. This scenario can lead to the demise of the previously dominant products and technologies.  

Translating this into spectroscopic terms, we can observe that the performance of laboratory spectrometers has been enhanced steadily over the last 60 years, by some estimates improving an order of magnitude every 10 years. Today's instruments appear deceptively simple compared with those of 20 or 30 years ago, but that apparent simplicity hides massive sophistication in optics, electronics, firmware, software, and manufacturing techniques. But have these spectrometers overshot the market requirements? As an example, in Fourier-transform infrared (FT-IR) spectroscopy, the majority of industrial applications, which in turn represent the majority of the market, can be satisfied at 8-cm

  resolution, whereas a typical laboratory instrument might be capable of a factor of 10 better resolution.  

Optical spectrometers employing devices based upon photonic and micro electro mechanical systems (MEMS) technologies will be described here. Funded by the telecommunications boom, many of these devices were developed only in the late 1990s and are still going through rapid performance enhancements. Therefore, we need to be aware of these new miniature spectroscopic technologies, not just because of the new applications they enable, but also because they might supplant traditional instruments in some sectors in the future. We can anticipate that there will be rapid growth in the number of miniature spectrometers and their applications because, as Feynman noted in his celebrated talk, there's "plenty of room at the bottom" (11). 

Let us take as a benchmark a scanning grating spectrometer. Although these have been supplanted in some sectors by other technologies (Fourier-transform spectrometers in the mid-infrared (MIR), and charge-coupled device [CCD]–based spectrographs in the visible, for example), this gives us a well-understood basis upon which to compare all the existing technologies. The two factors upon which to focus are throughput and limiting noise. 

Throughput, or étendue, is defined as the product of the area of an optical element (

) and the solid angle (Ω), subtended by a limiting stop at that element (12): 

 is a measure of the optical power transmitted through the system. In optics, this also is known as the optical invariant. Ideally, 

 is a constant throughout the optical system; it cannot be increased, but it is easy to decrease, for instance by overfilling an optical fiber, a sampling accessory, or a detector. If we take an existing spectrometer and shrink its linear dimension by a factor of 

, then the area of a typical mirror or lens will diminish by a factor of 

, and the throughput will be affected by the same factor. This clearly will have an effect on the signal-to-noise ratio (S/N) performance of the spectrometer; exactly how much will depend upon the limiting noise of the system. 

Optical spectrometers fall into two broad classes, those that are detector-noise limited and those that are shot-noise limited. As a generalization, MIR spectrometers are detector-noise limited; that is, the detector noise is greater than the sum of all other noise sources in the system (13). However, shorter wavelength spectrometers (near-infrared [NIR], visible, and ultraviolet, including Raman spectrometers [14]) are, again as a generalization, shot-noise limited. In an MIR spectrometer, decreasing the optical power by a factor of 

; in a visible spectrometer, that factor would be √

Putting aside developments in electronics, which are very significant, the approaches taken to improve the performance of existing spectrometers or to miniaturize the spectrometer without dramatically degrading its performance have therefore focused in a small number of areas. These include reducing detector noise, making the source brighter (increasing the useful power emitted per unit area), using a different optical technique to improve the throughput, and using a technique that observes more than one resolution element at a time (multiplexing and multichannel methods). It is instructive, then, to look at each spectral region and see which of these techniques is most appropriate and has been successful. This article concentrates on MIR spectrometers; subsequent articles in this series will examine NIR and Raman spectrometers. 

Conventional and Laboratory MIR Spectrometers

In MIR laboratory spectrometers, sources have been heated wires and ceramics, commonly operating at 1300–1500 K in air or the purge gas of the spectrometer (15). Applying Wein's displacement law (12) (expressed in frequency, not wavelength, units), the peak in the emission source curve for a 1300 K black body is at 2550 cm

, so that the majority of the light output is in a useful range for that spectrometer. Higher operating temperatures shift that peak to higher frequencies and increase the power emitted in the fingerprint region only slightly; furthermore, higher temperatures cause rapid degradation of the source as a result of the reaction of the source material with oxygen or nitrogen. Operation of a source in a vacuum is impractical in even a modest-sized instrument due to the problem of sealing infrared-transparent windows into a small vacuum chamber. Continuously tunable lasers, operating over the usual spectral operating range, do not exist. Moving to a mercury–cadmium–telluride (MCT) detector operating at 77 K might reduce detector noise, but this is impractical for a portable instrument. Finally, MIR array detectors have significant problems, including cooling requirements, very high cost, and problematic reliability.  

Therefore, the emphasis has been on enhancing throughput and on multiplex techniques. Harwit and Sloane (16) divided multiplex spectrometers into two classes: the first uses interference techniques and Fourier transforms, and the second uses masks and discrete transforms. The Michelson interferometer falls into the first class and Hadamard transform spectrometers fall into the second. The interferometer delivers both a throughput and a multiplex advantage, which has led to the dominance of FT-IR instruments over grating instruments. First, for a given size of optics and the same spectroscopic resolution, an interferometer can accept a larger solid angle of illumination than a grating. This is the well-known throughput or Jacquinot advantage (13,17). Second, the interferometer is a multiplexer: all the wavelengths of interest fall on the detector all the time. This is the multiplex or Fellget's advantage (13,17), and it provides an advantage factor of the square root of the number of resolution elements. It is important to note that the multiplex advantage applies because the instruments are detector-noise limited; in the case of a shot-noise-limited spectrometer, the multiplex advantage is cancelled by the increase in shot-noise. Hadamard spectrometers (18), which also have both a throughput and a multiplex advantage, have been applied in the MIR, but historically the practical construction difficulties relegated them to a footnote. Virtually every MIR laboratory spectrometer on the market today employs an interferometer rather than a grating. 

It is also worth noting that laboratory Fourier transform instruments have been designed with the interferometer between the source and the sample. This makes it a "pre-multiplexing" device, in the same way that a dispersive instrument with the grating between the source and the sample is a "predispersive" device. For a Fourier transform –infrared (FT-IR) instrument, this has an advantage in that any radiation emitted from the sample itself is not modulated, and therefore will not appear in the spectrum. It has a second, much more practical advantage in that it allows an unlimited choice of sampling accessories, detector optics, detector sizes, and detector materials. This has made the laboratory FT-IR spectrometer an extraordinarily versatile device. 

New Technologies and Fabrication Techniques

During the 1990s, billions of dollars were invested in optical communications companies, spawning a multitude of developments in pump laser diode modules (commonly operating at 980 nm and 1480 nm), semiconductor tunable lasers (19), optical gain media and fiber amplifiers, optical fibers, detectors, gratings, fiber Bragg grating sensors, photonic crystals, and tunable filters. These devices are all described in standard texts on photonics (20). Telecommunications applications require that devices be incredibly reliable and rugged, and that they operate under a wide range of temperature, vibration, and humidity conditions. The properties of these components make them very attractive for many applications outside telecommunications. The massive amount of investment in this market also spurred intense competition, resulting in very low prices for these components, especially after the dot-com and telecommunications bubbles burst about five years ago. After that market crash, there was interest in applying these technologies in other applications, and analytical spectroscopy started to see an impact. Because of the nature of optical communications, most of these devices work in the NIR, but it is likely that they will also have an impact in the MIR. 

One class of high-precision miniature devices is MEMS, and for optical devices the term micro optical electrical mechanical systems (MOEMS) is also used (21). The promise of MEMS-based manufacturing is analogous to the integrated circuit industry: there is considerable R&D expense upfront, but the manufacturing process can be both automated and highly reproducible, expanding to large volumes with little additional expenditure. With large volumes (>1000), piece part costs can be low. 

MEMS refers to the manufacture of miniature devices using semiconductor fabrication techniques. Instead of trying to machine small parts, processes such as lithography, chemical vapor deposition (CVD), wet etching, and reactive ion etching (RIE) are used; however, details of these processes are well beyond the scope of this article. MEMS devices are constructed most commonly in silicon, but there are processes for metal and plastic MEMS as well. On the microscopic scale, silicon suffers no hysteresis when flexed, and therefore, it dissipates little energy, does not suffer from fatigue, can run for billions of cycles, and provides very reliable devices.  

Miniature metal devices can be fabricated using a technique called LIGA (22), which is a German acronym (lithographie, galvanoformung, abformung) for lithography, electroplating, and molding. A metal mask is created lithographically and is placed in front of a sacrificial material like polymethylmethacrylate. This is then exposed to an intense beam of collimated X-rays (from a synchrotron) or UV light, which etches the sacrificial medium. The resulting cavities are filled using an electroplating technique, and the remaining medium is then dissolved away.  

MEMS devices are now commonplace (23), if not very visible, despite MEMS being a $6.5 billion dollar industry in 2007 (24). The largest applications are inkjet printer heads, pressure sensors, microphones, accelerometers, gyroscopes, and computer projectors (23). In the automotive industry, every airbag contains a MEMS accelerometer, tire pressure sensors are MEMS devices, and MEMS accelerometers are used in electronic stability control applications (25). Virtually every computer projector uses the Texas Instruments (Dallas, Texas) digital light projector (DLP) chip (26), which is a MOEMS device. Consumer electronics are seen as the next frontier in MEMS devices, enabling items such as image stabilization in cameras, improved microphones (27) in cell phones, and motion sensors in game controllers (28). Along with the growth and maturation of the market, there are now well-established MEMS foundries that can undertake design and fabrication (24); therefore, this is a function that a spectroscopy company can outsource — it does not have to develop the manufacturing technology itself (29), and does not need to invest in the considerable amount of capital equipment required. This will enable much more rapid adoption of MEMS techniques than if companies had to be vertically integrated. 

A number of prototype miniature spectrometers using MEMS techniques also have been constructed or proposed (30–34), including Fourier-transform spectrometers (35–39), and some of the gratings in the dispersive spectrometers are fabricated using the LIGA process (40–42). Given the low cost of high-volume MEMS devices, the vision of many of these developers is for low-cost spectrometers (under $1,000) manufactured in high volumes (many thousands per year). MIR instruments will be discussed in the following sections, and others in subsequent articles.  

Laboratory FT-IR spectrometers are high-throughput devices, but in many cases they overfill sampling accessories, losing that throughput. A key observation is that an instrument dedicated to one purpose, using a well-defined sampling accessory, can use lower throughput optics with no loss in performance. Also, for a dedicated analytical instrument, modest resolution (4 or 8 cm

) is all that is required. The first commercial FT-IR to embody these principles was the Hewlett-Packard (Palo Alto, California) gas chromatography (GC)–IR detector, introduced in 1986. Here, the sampling accessory was a 1-mm diameter lightpipe, and no provision was made for any other sampling mode. The IR detector contained a modest (4 cm

 ) resolution interferometer, and the complete system was significantly smaller than contemporary laboratory spectrometers, occupying a space about 8 in. wide. Most laboratory interferometers at that time used a laser referencing system running through the center of the circular beamsplitter. This is a poor match to a small-aperture accessory because there is effectively a "hole" in the middle of the IR beam: the central, on-axis rays that can best couple to the accessory are not present! However, the IR detector used a square beamsplitter, with the laser reference in one corner. This "systems engineering" approach to a GC–IR system resulted in performance that was better than that obtained by larger laboratory instruments with GC–IR accessories. 

This concept has been taken up by other manufacturers, most notably by SensIR (acquired in 2004 by Smith's Detection, Danbury, Connecticut) (43), with other dedicated systems incorporating low-throughput sampling devices: the IR microscope and diamond-attenuated total reflectance (ATR) accessories. Figure 1 shows an integrated FT-IR–diamond ATR spectrometer. SensIR produced a small interferometer matched to these accessories and again obtained performance equal to that of larger instruments. Today, other small interferometers (44) and dedicated FT-IRs are available (45,46). However, outside the laboratory, small interferometers have been used for decades in dedicated remote sensing applications, going back to the Block Engineering (Marlborough, Massachusetts) model 196 (47); a current example is the D&P Instruments (Simsbury, Connecticut) (48) MicroFT (Figure 2) (49). This employs a pair of potassium bromide sliding prisms, and the higher refractive index of that material enables a larger solid angle to be accepted into the interferometer (and therefore greater throughput), as compared to a mirror-based Michelson design of the same size. This is an example of a field-widened interferometer (50), and a similar approach can be taken in a grating instrument (see below). 

There are several papers in the literature describing MEMS FT-IR spectrometers, but to the best of the author's knowledge, these have not yet resulted in MIR commercial products from a major or traditional vendor, although two research-company spinoff products are available (see the following). In 1999, Collins and colleagues (51) reported on a miniature, microfabricated spectrometer. This used dovetail microjoints etched from silicon for the bearing of the moving mirror. Although this device was configured for the NIR, it did incorporate a large mirror displacement, providing an optical resolution that was unusually high for a miniature device. The bearing travel was as long as 8 cm, and the measured halfwidth of a laser line was 1 nm. This bearing design has been used to construct a brassboard system for Earth atmospheric studies (37), weighing only 0.5 kg but with a resolution of 0.1 cm

 . Knipp and colleagues (52,53) described the concept of a standing wave Fourier spectrometer for the visible region, a device that combines aspects of both a traditional Fourier spectrometer and Fabry-Pérot spectrometers. A miniaturized LIGA Fourier-transform spectrometer was described by Solf and colleagues (33), who reported spectra for laser emission and a bandpass filter in the NIR. 

Articles from the University of Neuchâtel (Neuchatel, Switzerland) have described a number of miniature spectrometers. In 1999, the authors presented a MOEMS-based miniaturized Fourier-transform spectrometer (Figure 3), 5 mm × 4 mm × 0.5 mm in size, based upon a silicon comb drive (31), with a resolution of ~10 nm in the visible region. The moving mirror was 75 μm × 500 μm and could incorporate a laser reference. The authors comment that the instrument's resolution is good enough for applications such as color sensors. Further details were given in a second article (54). A third article (35) describes a miniature lamellar grating (55) interferometer, again with dimensions of 5 mm × 5 mm (Figure 4). This had a resolution of 70 cm

  or 2.8 nm at 633 nm, and the emission spectrum of a xenon lamp was measured. The authors stated that the advantage of this technique was that it required neither a beamsplitter (as compared to a Michelson interferometer) nor an array detector (as compared to a spectrograph) for these measurements. A series of commercial instruments are now available based upon this lamellar-grating technology (56), with 30-cm

 resolution, and operating in the visible to NIR regions 500–2600 nm). Ataman and colleagues have also described a MEMS lamellar grating spectrometer operating in the visible and NIR (57). 

Kraft and colleagues (58–61) have described a number of miniaturized spectrometers, including a spectrometer with a MEMS interferometer. This device employed a resonantly driven micromirror, suspended on two long springs, driven by interlocking comb-structured electrodes (Figure 5), and contained in a vacuum chamber. The effective mirror area was ~1.6 mm

. The first prototype covered a spectral range from ~4500 to 1450 cm

, using a thermoelectrically cooled MCT detector at a spectral resolution of better than 30 cm

. It had a footprint of 180 mm × 120 mm × 120 mm and a mass of about 1.5 kg. This has now been commercialized (62), covering the 2–6 μm range at 30-cm

Yu and colleagues (63,64) also described a silicon-MEMS Fourier-transform spectrometer. The silicon beamsplitter, micromirrors, MEMS actuators, and fiber-optic positioners were all fabricated simultaneously by micromachining techniques: they used a combination of deep reactive ion etching and anisotropic wet etching. The footprint of the micromirror assembly was 4 mm × 4 mm. The performance was characterized using an external-cavity tunable laser, and a resolution of about 50 nm was obtained around 1500 nm in the NIR. The authors comment that devices such as this could find application in lab-on-a-chip and optical coherence tomography applications.  

Keränen and colleagues (65,66) described a number of miniature spectrometer designs. One MIR spectrometer was composed of three silcon micromachined devices: a pulsed infrared emitter, a tunable Fabry-Pérot interferometer, and a detector. Saari (67) described a second device, a Michelson interferometer that used a series of micromirrors to generate the optical path difference. Five micromirrors are arranged in a two-over-three configuration so that their displacements are additive. 

Finally, in 2004, Block Engineering reported (39) on the start of an effort to design, fabricate, and test a MEMS-based FT-IR spectrometer for use in the 2–13.5 μm spectral range (Figure 6). The design described uses an internal source at 800 K and an interferometer with 10-cm

  resolution coupled to a 1-m White cell. The resolution requires mirror motion on the order of 0.5 mm, which would give a retardation of 1 mm and therefore a spectroscopic resolution of about 10 cm

. This is in contrast to some of the other designs described earlier, which had mirror displacements around 40–50 μm and resolutions of about 100 cm

. Block anticipates using a torsional ratchet actuator, converted into linear motion, to achieve this retardation. 

It is clear that for analytical spectroscopy, resolution and throughput are the major issues with these devices. The Block Engineering project has specified a resolution familiar to analytical spectroscopists, but the others are around a factor of ten worse. Throughput remains the largest issue; to overcome this, the data given in most of these papers employs lasers and arc lamps as sources, and not standard analytical transmission samples, let alone diffuse reflectance samples. A typical laboratory FT-IR spectrometer has 30-mm diameter mirrors in the interferometer, giving an area of ~2800 mm

. One of the MEMS-device mirrors noted above had dimensions of 75 μm × 500 μm, giving an area of 0.0375 mm

, a factor of 75,000 less in area. In the best case of operating in the shot-noise limit in the NIR, this implies a factor of ~275 poorer in S/N. The instrument described by Kraft quoted 1.6 mm

, or a factor of 1750 less in area. It is reasonable to assume that it is detector-noise limited, leading to a factor of up to 1750 poorer S/N. As Block Engineering notes in their article, for MIR applications, a MEMS device must have the largest optical component sizes practical. A clear distinction should therefore be made between, on the one hand, the use of MEMS actuators and MEMS fabrication techniques, and on the other hand, the desirable size of the optics. 

We have already noted that Hadamard transform spectrometers are multiplex devices, and therefore can have an advantage in the detector-noise-limited regime (68). Most designs have used linear or rectangular masks, but Harwit and Sloane (69) described a design using a rotating mask with transmitting and reflecting sections (a "0/1" design), and they note some difficulties with this "0/1" implementation. A recent addition to the class of multiplex spectrometers is Aspectrics' (Pleasanton, California) encoded photometric infrared (EP-IR) spectrometer (70,71), and the overall concept and some applications have been described by Coates (6). The optical arrangement is source to sample, to spectrograph with encoder, to single-element detector (Figure 7). This technology can be applied in different spectral regions, but we will consider the MIR version here. Inside the spectrograph, a grating disperses the light from the source and sample onto a rotating encoder disk. This has a series of reflective tracks, each patterned to give a different modulation frequency for the wavelength region falling on it. After reflection, all the reflected light is directed onto a single detector. The tracks on the disk are constructed to give a sinsusoidal, as opposed to "0/1," modulation to the light. A Fourier transform converts the detector signal into a spectrum. The spectrometer has an overall volume of 180 in.

This system therefore bears some resemblances to FT-IR and Hadamard spectrometers, but there are also significant differences. The motion of the modulator is circular and continuous, as opposed to reciprocating, which has both mechanical design and duty cycle advantages. The number of wavelength regions to be examined apparently is limited to 128 currently; the choice of the grating determines the wavelength region to be examined, and the resolution is therefore a function of that range and the number of tracks. As a multiplex technique in a detector-noise-limited regime, this will accrue an advantage, or improvement in S/N, of ~11 (√128), as compared with a scanning grating instrument of the same dimensions and resolution. Unlike an FT-IR spectrometer, this does not have (or does not require) an internal reference laser. 

Visible spectroscopy has profited from the CCD detector, whose technical progress and low cost have been driven by consumer applications, while small NIR spectrometers have been enabled by indium-gallium-arsenide (InGaAs) array detectors, whose rapid development was fueled by the optical telecommunications boom. Unfortunately, there have not been comparable drivers for MIR array detectors, outside the specialist defense arena. Two-dimensional MIR array detectors are expensive; a complete MCT camera can cost in the range of $30,000. Uncooled cameras (72) are somewhat lower cost but are still prohibitive for portable spectrometers. Therefore, the use of array detectors has been largely confined to research-grade spectroscopic imaging systems, which have employed both two-dimensional and linear MCT arrays. However, a very limited number of nonimaging spectrometers with multiplexed linear arrays are available commercially.  

Lead sulfide (PbS) arrays cover the range from 1–3 μm, while lead selenide (PbSe) arrays cover from 1–5 μm, and these are available commercially from a number of vendors (73). PbS and PbSe are photon detectors, and uncooled photon detectors perform poorly at wavelengths longer than 5 μm. For that reason, longer wavelength uncooled detectors are thermal devices: bolometers, thermopiles, and pyroelectrics (72). Pyroelectric arrays (74) are available commercially (75), and depending upon the window material employed, these can have coverage to as long a wavelength as 40 μm (250 cm

). In general, pyroelectric detectors have no dc response; therefore, they require a modulated (chopped or pulsed) infrared source. Microthermopile and microbolometer detectors are still somewhat specialized items, available from a limited number of defense-related companies (3). A monograph by Kruse (71) compares the principal types of uncooled thermal detectors. 