This installment is the third in a three-part series focusing on major aspects of laser-induced breakdown spectroscopy (LIBS). The first installment set the stage with the basics of the LIBS measurement physics and standard applications. The second installment discussed the choices and trade-offs for LIBS hardware in detail, in particular focusing on lasers and spectrometers. This third installment discusses LIBS analysis in some depth, exploring the various ways to go from a LIBS spectrum to a solution.

One of the strengths of laser-induced breakdown spectroscopy (LIBS) is that it can be applied to multiple media. In various configurations, it is possible to measure solids, liquids, gases, and suspended aerosols. Measurements have been made on the ocean floor at depths of hundreds of meters, and on the surface of Mars in an atmosphere of 7 Torr, largely consisting of CO

. This flexibility is possible because the atomization and excitation source are combined in the LIBS laser pulse. This strength of LIBS is a complication, however, because the material ablation and plasma formation process profoundly impacts the analytical plasma, as we pointed out in the previous installments in this series (1,2). As a result, users must be careful when developing material classification and calibration methods for LIBS, in particular using "like" materials for calibrations and choosing methods consistent with the type and range of expected data. While it is impossible to cover the gamut of considerations in a short column installment, here I will attempt to cover some major features of the calibration and classification landscape.

Any discussion of calibration should start with univariate analysis of spectral peaks, the primary method of calibration of spectroscopic output since such analysis has been attempted. Spectral peak analysis in the context of LIBS spectra has a number of considerations. First, astute readers will note that the LIBS plasma is a time-dependent event. The acquisition period of the spectra relative to the plasma initiation becomes an important parameter in the quantification. This collection period is also a function of the laser energy density on the plasma, because the plasma lifetime scales with the deposited energy in the plasma, and the ablation event is important (primarily for solids) not only for exciting analytes, but also in determining the amount of laser energy channeled into the plasma. Hence, any calibrated LIBS system should take care to ensure consistent laser fluence (energy/area) on the samples. 

The influence of the data collection timing has been shown often, for example in an early paper by Fisher from Hahn's group (3). Figure 1 shows the chromium emission lines from particles in an airborne aerosol at three different settings of delay (after the plasma initiation) and gate width (shutter open time), collected with a Czerny-Turner spectrometer and an intensified charge-coupled device (ICCD) camera. At 2 μs delay and gate, the chromium triplet is not visible at all. At later times and longer collection periods, the signal-to-noise ratio (S/N) of the triplet is optimized. Fisher's paper goes on to point out that the optimum detector timing to maximize S/N is element-dependent under otherwise constant experimental conditions. 

Figure 1: Spectra of chromium aerosol collected at different delay/gate combinations, indicated by delay/gate in microseconds to the left of each spectrum. Spectra are offset for clarity, but are otherwise the same scale. Adapted from reference 3 with permission.			

After the data collection (laser energy, spot size, and collection timing) is fixed, data can be collected from known samples to form a calibration curve. Typically, a particular peak is selected for analysis of an element of interest. As shown in Figure 2, the peak (blue, solid fill) is defined and integrated, with the subtraction of the background, which can be defined by the area under the peak, defined by the peak edges (green, grid fill), or by nearby areas with no spectral interferences (orange, diagonal fill). Sometimes the peak area alone can be used for calibration, but in other cases the peak is divided by the background, sometimes called the "baseline," to form the "peak/base" ratio. If the spectrum is collected at proximate enough time to the plasma that there is a measurable continuum, this peak/base calibration tends to work well, because the baseline is indicative of the energy absorbed in the plasma. In other cases, normalization of the peak by the total plasma emission (Figure 3a) can be useful if the total plasma emission is fairly constant (for example, measurement of a low concentration alloying element in a steel matrix). If there is a known, fairly constant element, another similar approach is to use a line of the constant element to standardize the peak of the element of interest (Figure 3b). For groups of samples with a consistent or relatively simple background, univariate methods are often very successful calibration schemes for LIBS. 

Figure 2: Example of peak integration and baseline (background) definitions.			

Extensions to Basic  Calibration — Line Selection

Individual emission lines each have a particular range of applicability. Lines with low excitation energy and ending at or near the ground state tend to have a nonlinear response at higher concentrations because of self-absorption, but these lines are the most sensitive for measurements at low concentrations. Harder-to-excite lines, particularly those not ending on the ground state, are more useful for higher-concentration measurements. Lines not ending on the ground state are not as susceptible to self-absorption as lines that transition to the ground state (there is negligible population to absorb in upper states). The result of these factors is that calibration lines are selected based on the concentration range being measured, and often the strongest lines in the spectrum of a particular element are not used for higher-concentration measurements. Similar to inductively coupled plasma (ICP) methods, calibration over a large range of analyte concentrations may use a succession of lines, some for lower concentrations and some for higher concentrations. 

Figure 3: Univariate calibration of the (a) chromium 397.67-nm line divided by the total light intensity from the plasma, and the (b) chromium 397.67-nm line divided by the iron 404.58-nm line, in a range of low- to high-alloy steels. Adapted from reference 4 with permission.			

Multivariate concentration methods use multiple lines, or multiple parameters, to come to a concentration measurement. Broadly, these methods are lumped under the heading of "chemometrics." The most common include partial least squares (PLS), principal component regression (PCR), and support vector machines regression (SVR). Numerous other methods exist, and additional methods are continuously under development. Those mentioned here are the ones most likely to be found in commercial data analysis packages and implemented in mathematical programming languages such as R, Python, or Matlab. The description of the methods below is merely to guide use and implementation, and the reader is encouraged to find additional resources for the study of particular methods of interest. 

One commonality to all of these multivariate methods is the importance of the definition of a "training" set of data and a "test" set of data. This is more important for multivariate methods than for univariate peak integration because the predictive model is of arbitrary dimension using many of the multivariate techniques, and the number of sample spectra is relatively few compared with the number of data points in each spectrum. For example, an analyst may have 100 spectra for each of 10 different standards for calibration of a particular element in an alloy — seemingly a lot of data. However, each spectrum may itself have 10,000 or more data points, with each data point acting as an individual parameter, allowing arbitrarily complex modeling. Typically, as the number of predictive variables increases in the model, the better the performance is on the training data. The result is that complex models can be (and are, by hapless analysts) built using too many coincidently correlated variables. Such "over-fit" models may perfectly describe the training data, but will quickly fail on independent test data and on data that is slightly outside of the range of the training data. The remedy is rigorous testing with independent test data to ensure that the model is stable and accurate when presented with new challenges. During the model-building phase, one can and should use a cross-validation method (such as leave one out, k-fold cross-validation, and so on). One is reminded otherwise of the quotation attributed to John von Neumann: "With four parameters I can fit an elephant, and with five I can make him wiggle his trunk" (5).   

As suggested above, it is usually a good strategy to limit the number of predictive variables to the minimum number required to sufficiently model the data, thus ensuring that the model takes only "primary" features into account. One other means of accomplishing this is to preselect important lines or portions of the spectrum, potentially also pretreating the data, before sending the data to a multivariate technique for calibration. In this way, the physical aspects of the data can be emphasized (for example, several Mg emission peaks can be used to predict Mg), which tends to increase the accuracy of the model compared to models that use a general method to determine which variables are most correlated. An example of this is data treated using the PLS "dominant factor" method promoted by Wang and colleagues (6), in which preselected factors (for example, normalized peaks or plasma parameters) are emphasized in the PLS. Figure 4 shows PLS-based calibrations derived from 77 standard coals, in which the LIBS predicted value is on one axis, and the known value is on the opposite axis. Figure 4a is a prediction of carbon in coal using only PLS on the more than 13,000 values in the original spectra. Figure 4b shows preliminary data showing dominant factors PLS data, using primarily the roughly 200 most important lines in the spectra, standardized. Obviously, a significant improvement can be achieved with "smart" inputs to a multivariate method like PLS. 

Figure 4: Predictions of carbon concentration compared with actual values in 77 coal samples using (a) standard PLS on the raw coal data (by the author), and (b) dominant factors PLS (preliminary data from the laboratory of Professor Zhe Wang, Tsinghua University). In (a) "x" marks training data and dots indicate test data, while in (b) "+" marks the training data and squares indicate test data.			

, PLS attempts to find those factors, called 

, that best describe the variation in the training set. The first latent variable describes the maximum amount of variation in 

, and subsequent latent variables are orthogonal to previous latent variables, each describing the maximum amount of remaining variation in 

. Users should note that PLS is inherently a linear multivariate method, and so substantial nonlinearity in the relationship between 

 will be difficult to capture. However, PLS is excellent for datasets in which there are multiple related variables (as in many pixels related in a particular atomic peak), and in which there are many more variables than observations. As both of these conditions are true of LIBS, PLS is commonly used for LIBS calibrations. PCR is a related method. 

Support vector machines (SVM) can be used for classification or for regression, termed 

 (SVR). Data are projected into higher-dimensional spaces to achieve separation, and discrimination is made by determining the separation between data in those higher-dimensional spaces. Kernels for data projection can be either linear or nonlinear, enabling SVM-based methods to deal with both types of data, provided that the appropriate kernel is chosen. Figure 5 illustrates SVR used on the same coal dataset mentioned above, in which the line represents the model built with the training data, and the points represent the prediction on the test set. This prediction used a linear kernel on the raw LIBS data. 

Figure 5: SVM regression of the raw coal data from the dataset described earlier, by the author. The blue line is the model from the training data and the points are from the test set of data.			

Classification is often a simpler problem than quantification. These problems involve simple identification of a sample and placing it into one of a set of known classes. Multivariate classification methods are likely more numerous than multivariate quantification methods, and again readers are encouraged to find additional resources for self-study. Popular classification methods include clustering, soft independent modeling of class analogy (SIMCA), partial least squares discriminant analysis (PLS-DA), and the previously mentioned SVM.   

Of all of the methods, clustering methods are one of the simplest. In a typical clustering method, the centroid of a group of 

 points would be calculated by determining the average value of each of the 

 spectra would all be of the same type. With the centroid of every type (class) of sample calculated, the model can be run on unknowns. For each unknown, a "distance" such as the Euclidean distance (root-square calculated with each of the 

 pixels of the unknown spectrum) from the unknown to each of the class centroids is calculated. The class with the centroid that is closest to the unknown is deemed the class of the unknown. In this way, every unknown is assigned to a class. Variants include alternative methods of distance calculation, and definition of threshold distances that define when an unknown is a member of a class (and outside of which the unknown may be assigned to an "unknown" class). 

SIMCA relies on a principal components analysis (PCA) to separate known materials into classes. Because SIMCA is a supervised method that is given the class of each of the known training samples when building the model, it retains only those principal components that are useful in predicting class. When presented with unknowns, it can predict the probability that the unknown is a member of one of the classes, or it can predict that the unknown is not in any of the known classes, depending on the tolerances that are set during building of the model. 

PLS-DA uses the same methods as PLS to project the data into latent structures. Instead of continuous variables for quantitative prediction, discrete variables are assigned for classification. 

Most classification results are either presented in terms of a "confusion matrix" or a receiver operating characteristic (ROC) curve, which graphs the false positive rate (

-axis). An example of a confusion matrix is shown in Figure 6. In this figure, samples are given arbitrary class names 1, 3, 5, . . . 17, and 21. The actual data that generated this matrix were types of fertilizer and the test set were five samples of each type. A K-nearest neighbor (KNN) clustering algorithm was used to try to place unknowns into classes. Numbers on the diagonal of the matrix indicate correctly identified samples — that is, three samples that were actually in class "1" were also identified as class "1," and five samples that were actually in class "13" were identified as class "13." Misidentifications are off of the diagonal, such as the two samples of class "1" that were predicted to be class "3," or the one sample (each) of class "5" that were predicted to be class "3" and class "7." In this case, the samples were ordered by increasing nitrogen concentration and decreasing phosphorous concentration, so all of the confusions were from the correct class to the adjacent class. On the same data, a perfect classification (all "5's" on the diagonal) could be achieved with the SIMCA algorithm. 

Figure 6: A confusion matrix representing performance of a clustering algorithm on LIBS analysis of fertilizer samples.			

Methods can be combined to improve accuracy. As an example, a classification might be used to separate samples into logical groups before applying more specific quantitative calibrations developed for each group. Or a statistical method such as analysis of variance (ANOVA) could be used to determine the independent variables (intensities) in a spectrum that are most closely correlated to a dependent variable that one is trying to predict, and those variables could be fed into a quantitative method such as PLS. The possibilities for variation are endless, as are the pretreatment methods for spectra, such as normalization, averaging, and so forth. I encourage readers to explore and try different methods, particularly realizing that each dataset may require slightly different treatment.  

A discussion of LIBS calibration methods would not be complete without a mention of so-called "calibration-free" LIBS. These methods generally aim to derive fundamental parameters from the plasma, such as equilibrium temperature and electron density. They use these parameters in conjunction with the emission spectra to "close the loop" on overall element concentrations, assuming that all major elements in the sample can be measured.  

For example, if the plasma conditions are known, then for a single element the ratios of atomic lines in the spectrum can be predicted. For two elements in a binary system, the relative emission coefficients between various major lines of each of the elements are known, and so the relative concentration (which would add up to 100%) in the binary system is known. For 3, 4, . . . 

 elements, there are more equations but the method is the same, the closure condition being that the predicted concentration of the 

Great strides have been made in calibration-free methods, notably by researchers from Pisa (7) and with useful additions by Gornushkin and coworkers (8). Accuracies in known systems can be ±20% or better. Issues with the method include uncertainties in the relative emission coefficients, inter-element interactions, and uncertainties in the plasma parameters. Plasma conditions are not stationary during any detection interval of finite duration, nor are plasma temperature and electron density homogenous in the plasma volume. Adjustments for these time- and space-dependent distributions have been helpful in advancing the method, which remains an active area of research. 

There are obviously many methods to come to both quantitative calibration and qualitative classification with LIBS. I am often asked what to expect in terms of overall performance. With real-world samples, typical relative standard deviation (RSD) numbers are in the range of 2–3% for many types of analysis. For the most homogenous samples, RSDs less than 1% can be achieved, for example, on obsidian glass samples or gemstones. With both classification and quantification methods, one must be sure to avoid over-fitting and rigorously test any model before deployment. Following the general guidelines here, with a bit of experimentation, should get you a good way toward successful implementation of LIBS. 

As this is the last of the three tutorial installments planned in 

 on LIBS, I'd like to express my appreciation to the editors for suggesting this series. It has been a joy to write. For those readers who have additional questions, please feel free to use my e-mail address listed on the next page. Those who know me will attest that I always enjoy talking about LIBS! Best of luck to each of you in your spectroscopic endeavors! 

(3) B.T. Fisher, H.A. Johnsen, S.G. Buckley, and D.W. Hahn, 

(4) C.B. Stipe, B.D. Hensley, J.L. Boersema, and S.G. Buckley, 

(5) Attributed by E. Fermi, as quoted by F. Dyson in 

(6) Z. Wang, J. Feng, L. Li, W. Ni, and Z. Li, 

(7) E. Tongnoni, G. Cristoforetti, S. Legnaioli, and V. Palleschi, 

(8) I.B. Gornushkin, S.V. Shabanov, S. Merk, E. Tongnoni, and U. Panne, 

 is a Director of Market Development at TSI Incorporated, which acquired the LIBS business of Photon Machines, Inc., in 2012. Before cofounding Photon Machines, Steve was a tenured professor of engineering at the University of California at San Diego. He has been working on research and practical issues surrounding the implementation of LIBS since 1998. He can be reached at 

Articles

Laser and Optics Interface

Users must be careful when developing material classification and calibration methods for LIBS. By following some guidelines, one can achieve relative standard deviation values of 2–3% for many types of analysis, and below 1% for homogenous samples.

LIBS Basics, Part III: Deriving the Analytical Answer — Calibrated Solutions with LIBS

Macro attenuated total reflection (ATR)-Fourier transform infrared (FT-IR) spectroscopic imaging is a powerful analytical tool, but it is still underexploited in the spectroscopic imaging community. This article presents an overview of current macro ATR-FT-IR imaging approaches and opportunities for studying dynamic processes such as diffusion, sorption, crystallization, and dissolution. 

	Fourier transform infrared (FT-IR) spectroscopic imaging combines the chemical specificity of IR spectroscopy with the spatially resolved nature of a mercury cadmium telluride (MCT) focal plane array (FPA) detector. These detectors typically contain 4096 (64 × 64) or 16,384 (128 × 128) infrared sensitive elements or pixels. Unlike using single-element detectors, where the resulting spectrum is averaging information from different parts of the sample, each detector element in an FPA is focused on a different spatial region of the sample. For heterogeneous samples, this increases sensitivity because with a single-element detector, a small particle may be "averaged out" by the presence of other materials in a single spectrum, but with imaging, such a particle would occupy a few pixels in the array, whose spectra would then be that of only the particle and not the surrounding material (1). The spatially resolved nature of imaging also allows for location-based spectral processing from simple masking to improve average spectra, to clustering or classification techniques to identify regions of interest (2). It is important to note that the use of the term 

 indicates that spectra are obtained from all pixels simultaneously in contrast to 

, where a single-element detector is scanned over the desired area. The main advantage of imaging compared to mapping is the ability to study dynamic systems and samples as spectra can be obtained within a few minutes or less (3). Advances in computational power and optimization of scanning parameters have enabled images to be obtained within 50 ms, fast enough to study moving droplets within a microfluidic device (4).

	Imaging is commonly achieved in transmission, in which the IR beam passes through a sample that is on one or between two IR-transparent windows. The principal disadvantages of transmission are that the sample must be thin (<20 μm) and aqueous samples are particularly challenging to study because of the strong absorption of water. Attenuated total reflection (ATR) mode uses a suitably shaped crystal of high refractive index material, in which incident infrared light undergoes total internal reflection at the interface between the crystal and the sample. At this interface, an evanescent wave penetrates into the sample, whose depth of penetration (

) depends on the angle of incidence (θ), refractive indexes of the crystal (

), and wavelength (λ) of infrared light. The 

, where the intensity of the electric field of incident infrared light falls below e

For mid-IR, the depth of penetration is typically 1–5 μm and the crystal material used depends on the type of sample and material properties required of the crystal. Germanium and zinc selenide can be damaged easily with hard samples, with the latter also having a limited pH range for samples. The refractive indexes of these materials are 4 and 2.4, respectively, with germanium commonly used in microscope ATR objectives because of its high numerical aperture. Silicon and diamond are also commonly used materials, with diamond being particularly robust. They have refractive indexes of 3.4 and 2.4, respectively.

	The FPA detector is often mounted in a separate compartment or macrochamber attached to an FT-IR spectrometer in which different ATR accessories can be positioned. The type of accessory can vary depending on the field of view required and the angle of incidence of the incoming beam (which affects depth of penetration). The most suitable accessory depends on the experiment and different types are demonstrated in the following sections. The FT-IR spectrometer sends the IR beam through one of its side-ports, through the macrochamber and ATR accessory, to the FPA where the IR beam is focused on the surface of the crystal and FPA to obtain spatially resolved IR spectra. It should be noted that because of the exponential decay of the evanescent field, particles can appear smaller than their true size (5), but for many applications, the interest is in chemical and spatial changes at the surface. To obtain two-dimensional (2D) color images, univariate analysis, which involves selecting spectral bands unique to components of interest, can be performed on the spectral data cube. Plotting the integrated absorbance of spectral bands of different components for all pixels against a color scale produces images with respect to each component, where regions of differing concentration can be identified by different colors. A summary of this process is shown in Figure 1. Multivariate approaches are often used when many bands overlap such as in complex pharmaceutical systems (6,7). This versatile approach can also be applied in micro mode with an IR microscope (8), though the focus of this article is on macro-mode imaging (9). Macro-mode imaging has many applications (9,10), and this article provides examples of macro-ATR accessories and case studies where their different properties were used to reveal new information about the samples under investigation.

Overview of ATR Spectroscopic Accessories

Large field of view accessories use no magnifying optics and, thus, have a field of view between 2.5 mm × 3.6 mm (11) and 15.4 mm × 21.4 mm

 (12), depending on the pixel size and number of pixels in the FPA (64 × 64 up to 128 × 128), the angle of incidence, and the accessory used. The Specac Oil Analyser (13), FastIR (14), and Pike Veemax 2 (15) accessories (the latter of which has an adjustable angle of incidence) are examples of large field of view accessories. Schematics of different accessories are shown in Figures 2 and 3. In some accessories, an undesirable gradient of absorbance can be observed along the crystal as the optical configuration permits a spread of different angles of incidence, affecting the depth of penetration (14). In other accessories with movable mirrors, this gradient can be minimized or removed completely (10). Crystals used with such accessories are often zinc selenide, silicon, or germanium. Large field of view accessories are used when spatial resolution is secondary to field of view and, as such, are suitable for pharmaceutical studies (16), microfluidics (17), fingerprint analysis (18), depth profiling (15), protein crystallization (19), and other high-throughput approaches. Because the crystals are not as robust as diamond, they cannot be used for high-pressure experiments or in cases where materials need to be compacted against the crystal surface for good contact. The use of acidic or strong bases can also damage zinc selenide, leaving silicon as the only suitable option. However, silicon's higher refractive index than zinc selenide reduces the depth of penetration, resulting in a weaker probing depth, and its higher cut-off precludes the study of spectral bands below ~1000 cm

. The use of depth profiling by changing the angle of incidence is particularly useful for chemically imaging thin polymer films (15) and has potential in the future for depth profiling other materials, overcoming one of the limitations of transmission. The Oil Analyser has a field of view of 2.5 mm × 3.6 mm with a 64 × 64 pixel detector (11). The Pike VeeMax has a field of view of 7.0 mm × 9.8 mm with a silicon crystal (19) or 5.3 mm × 7.9 mm with ZnSe crystal (20), depending on the angle of incidence, when using a 96 × 96 pixel detector. An expanded field of view accessory as shown in Figure 4 is particularly useful for high-throughput studies in which a larger area (15.4 mm × 21.4 mm) of the ATR crystal can be imaged (12). For an angle of incidence of 45°, images will have an aspect ratio of 1:√2 unless correcting optics are used, and other angles will produce images of different aspect ratios as shown in Figure 3.

The imaging Golden Gate accessory, initially developed by Specac for nonimaging applications and, thereafter, containing aspect-ratio correcting optics for imaging, contains a diamond ATR crystal mounted in tungsten carbide (Figure 5). The diamond is cut to form a 45° inverted prism with a top surface of 2 mm × 2 mm. Lenses, normally zinc selenide, condense the incident beam into the diamond, and a similar lens on the opposite side expands the beam, resulting in an imaging area of approximately 656 μm × 579 μm for a 64 × 64 FPA detector, a magnification of approximately 4×. It is important to note that with noncorrecting optics, the imaging area is 820 μm × 1140 μm (21). This accessory allows for in situ compaction and dissolution of pharmaceutical tablets (22), high-pressure and high-temperature experiments (23), and, more recently, the use of apertures for variable angles of incidence to eliminate dispersion artefacts because of high refractive index samples (24). The use of apertures in this system allows for depth profiling in a similar manner to that of the large field of view accessories (25).

Dynamic Applications with Macro ATR-FT-IR Spectroscopic Imaging

Drug release from solid oral dosage forms is one of the most common methods by which drugs are administered to patients, and the release mechanisms that occur after a tablet is ingested determine the bioavailability of the drug. Drugs or active pharmaceutical ingredients (APIs) are often complex molecules with low aqueous solubilities, so one important role of other components or excipients within the tablet is to aid solubilization of the drug as it is released from the tablet. Conventional dissolution studies use United States Pharmacopeia (USP) dissolution tests, in which tablets are dissolved under well-defined flow conditions and the amount of dissolved drug is measured using ultraviolet–visible (UV–vis) spectroscopy. This approach has drawbacks such as not giving information about mechanisms within the tablet as the drug is released, and it only shows the amount of drug dissolved over time, not necessarily released in other solid forms such as particles or precipitates. This distinction is important as the drug can be released from the tablet, but in a form which is not bioavailable (that is, a solid or undissolved form). Other approaches such as UV imaging (26), Raman (27), and magnetic resonance imaging (MRI) (28,29) can also be used noninvasively to obtain information from a tablet as it dissolves. MRI can give three-dimensional (3D) images of the whole tablet under USP IV conditions, but its acquisition times for each 3D image are long, up to 30 min (28). FT-IR spectroscopic imaging of tablets can be done in transmission mode, but the tablets must be very thin to reduce the effects of strong water absorbance (although the use of D

O can aid with this) (30) and formulated to reduce scattering of the infrared light, limiting the applicability of this approach in studying pharmaceutically relevant tablets.

Macro ATR-FT-IR spectroscopic imaging is particularly suited to studying drug release from microparticles (31) and tablets (32) because they do not need to be thin and components within the layer in contact with the crystal can be imaged over time (33). Importantly, spectroscopic imaging can reveal drug crystallization because of changes in molecular interactions (16). Using a diamond ATR accessory, the tablet can be compacted in situ, guaranteeing good contact with the crystal and with a suitable flow cell (34), dissolution can also be carried out without removing the tablet from the crystal. A UV–vis detector connected in-line downstream can give the amount of dissolved drug over time in a comparable way to USP and using a transparent flow cell (with ex situ compaction), visible images can also be captured (32). With the latter approach, differences have been demonstrated between visibly observed dissolution fronts (that is, water penetration into the tablet, drug dissolution and diffusion, and so on) and those produced from FT-IR imaging. This result was particularly important because relying on changes in refractive index or color changes is not a reliable method to study dissolution and it would only be applicable to certain formulation types. It is also possible to use a large field of view ZnSe variable angle accessory for an overall view of the tablet as it dissolves. This approach has been combined with computer modeling of drug release, where information from FT-IR imaging was used for model validation and predictive capability testing (35). The combination of computer modeling and FT-IR imaging opens up the possibility of formulation optimization for tablets of different shapes and composition (20).

	The function of a protein is governed by its tertiary or 3D structure, and X-ray crystallography is usually used to obtain structural information. However, crystallization is a major bottleneck in protein screening because the conditions under which crystals will form and grow are narrow. Crystallization normally occurs as the solution is brought to super-saturation and nucleation followed by crystal growth. However, excessive nucleation can cause crystals to be too small and have insufficient space to grow into long-range ordered structures, or the protein may precipitate into an amorphous solid. Thus, it is necessary to optimize the crysallization conditions based on temperature, pH, and the presence of other chemicals or "leads" whose function is to initiate crystallization or stabilize growing crystals such that they become large enough for X-ray crystallography. One problem is that other species crystallize in addition to the protein, and determining if a particular crystal is proteinaceous is often chemically or physically destructive.

	Consequently, it is desirable to have a chemically specific, nondestructive method that can determine if a crystal is a protein and if possible, obtain structural information at the same time. UV fluorescence imaging can be used, but with wide absorption bands, it is not as chemically specific as IR spectroscopy. In addition, IR spectroscopy allows for the determination of secondary structure. The crystallization of proteins directly onto an ATR crystal, combined with imaging, allows for determining which crystals are indeed proteins and how quickly they grow over time. Unlike Raman spectroscopy, where individual crystals would have to be located and scanned, macro ATR-FT-IR imaging obtains spectra from all protein crystals simultaneously. The crystallization of lysozyme has been demonstrated on a ZnSe ATR crystal, and the use of this large field of view also opens up the possibility for high-throughput approaches, for example, different concentrations of NaCl and protein solution can be imaged over time in microwells (11). However, the spatial resolution suffers with a large field of view, and smaller protein crystals may not be detected. The use of a diamond ATR crystal (11) demonstrated protein crystallization at a higher spatial resolution, and combining this crystal with a microfluidic approach (36) would enable high-throughput screening of different conditions, considerably speeding up the discovery of the optimum set of parameters.

	Another example of a high-throughput approach is the surface modification of the ATR crystal itself (19). In this example, a gradient of hydrophobicity was created using a modified vapor deposition approach, in which octyltrichlorosilane is deposited on a silicon ATR crystal in one direction. It was shown that lysozyme adsorption and crystallization decreased as the hydrophobicity decreased along the crystal. The use of an FPA detector allowed for pixel binning in the direction normal to the gradient on the crystal surface, significantly improving the signal-to-noise ratio. Surface modification with nanoparticles for surface-enhanced infrared absorption spectroscopy (SEIRAS) can increase sensitivity by two orders of magnitude and enables the study of monolayers (37). Nanoparticle layer thickness and uneven deposition can present a problem with loss of spectral intensity and uneven enhancement. However, there exists great potential for imaging in this area and though this has been applied to polymers (10), it has yet to be applied to proteins.

	Another exciting application of macro ATR-FT-IR imaging is for the analysis of materials processed with supercritical fluids. Supercritical CO

 is a widely used supercritical solvent because it is nontoxic, environmentally benign, and relatively inexpensive. It can also be removed from systems at ambient temperature and pressure unlike many organic solvents because it is a gas under these conditions. An important effect is the plasticization of polymers in which the supercritical CO

 is absorbed into the polymer, causing it to swell and the glass transition temperature (

) to be lowered. This aids the diffusion of materials into the polymer and the processing of the polymer itself via extrusion or blending (38,39). Using macro ATR-FT-IR imaging with a high-pressure cell presents a unique opportunity to study polymer behavior under high pressures of CO

. In an article demonstrating this approach (40), two polymer systems were studied. The first was a polystyrene–polyvinylmethylether blend, which has been observed to have a lower critical solution temperature, thus increases in temperature result in phase separation. In this system, phase separation was observed after pressurizing the polymer blend with CO

 to 60 bar and back to atmospheric; the CO

 plasticized the polymers allowing phase separation to occur. The imaging approach allowed in situ observation of the formation of polystyrene- or polyvinylmethylether-rich regions, demonstrating the effect of supercritical CO

 in a similar way to raising the temperature. The second was a study into the sorption of supercritical CO

 into polymethylmethacrylate (PMMA) and polyethylene oxide (PEO) films simultaneously at different temperatures. It was observed that CO

 solubility decreases as a function of increasing temperature with PMMA. For PEO, little sorption was observed at 30 °C, but at 50 °C after the polymer had melted, the absorbed CO

 concentration increased significantly. This demonstrated a high-throughput approach for analyzing different polymers under identical environmental conditions and with imaging, it was also possible to observe polymer interdiffusion.

	Heating of crude oil for distillation accounts for a large fraction of energy use during oil refining, and fouling caused by precipitating species is a major source of inefficiency (41). Fouling materials are complex, containing insoluble impurities, inorganic species, and 

, which are polycyclic aromatic hydrocarbons, one of the heaviest fractions of crude oil. Asphaltenes are a major contribution to fouling and their composition and tendency to form depend on the type of crude oil, temperature, and pressure of the system. There are many analytical approaches for studying asphaltenes, each with their own advantages and drawbacks. FT-IR single-element spectroscopy has been used to study petroleum asphaltenes (42); however, with spectroscopic imaging it is possible to analyze crude oil behavior under different or industrially relevant conditions to observe asphaltene formation and obtain chemical information from their deposits in situ. Using a diamond ATR accessory, solid crude oil deposits were initially studied and the use of custom apertures in this accessory removed anomalous dispersion, enabling different chemical clusters to be observed within the deposit (24). Applying this method to study deposition in situ with crude oil was achieved by Gabrienko and colleagues (43), where 

-heptane was used to precipitate out asphaltenes. The approach demonstrated the viability of using in-situ macro ATR-FT-IR imaging to study these deposits as they formed over time, and it was subsequently used to study asphaltene deposition from different geographical sources of crude oil (44). From this, it appeared that the total asphaltene concentration increased with increasing sulfur content (45) with sulfoxides observed as the dominant sulfur species, but other sulfur species that may be present have spectral bands beyond the lower limit of the FPA detector (<900 cm

) (46). Further investigations into the stability of oils under different conditions such as temperature and pressure are possible with an imaging diamond ATR accessory (43,44), attesting to the versatility of this approach to studying these samples. Thus, in the future, the information gained can be used to find optimal physical or chemical conditions to prevent asphaltene deposition.

	Microfluidic technology continues to expand as an analytical tool because of the ability to manipulate very small volumes (up to microliters) of fluids under conditions where laminar flow and diffusive heat and mass transfer tend to dominate. The importance of label-free methodologies such as FT-IR imaging as a detection tool has recently been discussed by us in this publication (47). This miniaturization and control of flow allow for so-called "lab on a chip" devices to be manufactured. A popular material for manufacturing microfluidic devices is polydimethylsiloxane (PDMS), in which channels can be created using a variety of processes such as soft lithography (48). Direct printing using paraffin waxes has also been proven to be particularly effective for rapid prototyping of devices. With this method, it is possible to print microfluidic channels directly onto infrared transparent windows (49), enabling analysis of reactions by FT-IR spectroscopic imaging (50). However, there is the problem of high water absorption in these transmission devices. Printing wax directly onto an ATR crystal (49), or using a suitable PDMS device, removes this caveat and has allowed mixing of fluids such as PEG and water, or H

O to be observed (17). In these instances, a "Y"-shaped device was used with two fluids entering at the top of the "Y" and mixing, establishing steady-state conditions along the channels. Imaging allows the progression of chemical mixing (in the case of PEG and water) and reaction (in the case of H

O) along the channel to be monitored, with a gradient being established in the former case, or the reaction products (HDO) in the latter case formed at the interface between the two input fluids. Under steady-state conditions, the ATR crystal can be moved so as to map a larger area of the microfluidic device (17). Fast imaging in transmission has been achieved using "kinetic mode," where interferograms are obtained and saved without coaddition, enabling spectral images to be acquired in as little as 50 ms (4). The speed of acquisition is affected by spectral resolution (which determines the distance the moving mirror has to move in the interferometer), the spectral range, and the refresh rate of the FPA. The latter depends on the FPA window size, hence smaller window sizes (64 × 64) will scan faster than larger sizes (128 × 128). This fast scanning is important for imaging segmented flows, in which one phase is separated by another. This allows each droplet to become a "micro-batch reactor" and eliminates axial flow within the channel. In ATR mode, the droplet phase can become detached from the ATR crystal surface, although this can be mitigated by reducing the flow velocity. Monitoring these droplets as they progress through the device allows the reaction progress to be analyzed, and changing the input chemistry of each droplet as it is generated facilitates a high-throughput approach for measuring and optimizing reaction chemistry (47,50).

	In this article, several case studies were presented that demonstrate the versatility of macro ATR-FT-IR spectroscopic imaging. The selection of the appropriate accessory, crystal, and angle of incidence depends on the system being studied and also on the required spatial resolution, field of view, sample type, and desired experimental conditions. As ATR-FT-IR spectroscopy relies on the evanescent field, it does not suffer from strong absorption of water, or other strongly absorbing samples, which can obscure weaker bands and enables a variety of aqueous or other systems to be quantitatively studied. The FPA detector increases sensitivity for heterogeneous samples as each pixel corresponds to a unique location on the sample surface. This feature is particularly powerful when combined with a macro-ATR approach for analyzing pharmaceutical drug or protein crystallization, asphaltene deposition, phase separation, and fluid analysis in microfluidic devices. This approach also facilitates high-throughput analysis, where many different samples can be analyzed under identical conditions, increasing experimental reliability and reducing parameter space exploration times. The time-resolved nature enables dynamic systems to be imaged in-situ, where depending on the spectral quality and range required, spectral data cubes can be obtained within 50 ms. As macro ATR-FT-IR imaging can analyze a variety of dynamic samples and systems with high resolution and chemical specificity, its applications in emerging areas such as biotechnology, supercritical fluids, microfluidics, and novel formulation and material design are far-reaching.

	S.G. Kazarian acknowledges research funding from the European Research Council under the European Community's Seventh Framework Programme (FP7/2007-2013)/ERC advanced grant agreement no. (227950).

	(2) P. Lasch, W. Haensch, D. Naumann, and M. Diem, 

	(4) K.L.A. Chan, X. Niu, A.J. deMello, and S.G. Kazarian, 

	(5) N.J. Everall, I.M. Priestnall, F. Clarke, L. Jayes, G. Poulter, D. Coombs, and M.W. George, 

	(6) J. Tetteh, K.T. Mader, J.M. Andanson, W.J. McAuley, M.E. Lane, J. Hadgraft, S.G. Kazarian, and J.C. Mitchell, 

	(7) J. van der Weerd and S.G. Kazarian, 

, (John Wiley & Sons, Ltd, 2014), pp. 1–26, DOI: 10.1002/9780470027318.a5609.pub2.

	(11) K.L.A. Chan, L. Govada, R.M. Bill, N.E. Chayen, and S.G. Kazarian, 

	(13) J. van der Weerd and S.G. Kazarian, 

	(14) E. Wessel, G. Heinsohn, H. Schmidt-Lewerkuehne, K.-P. Wittern, C. Rapp, and H.W. Siesler, 

	(17) K.L.A. Chan, S. Gulati, J.B. Edel, A.J. de Mello, and S.G. Kazarian,

	(18) C. Ricci, P. Phiriyavityopas, N. Curum, K.L.A. Chan, S. Jickells, and S.G. Kazarian, 

	(19) S. Glassford, K.L.A. Chan, B. Byrne, and S.G. Kazarian, 

	(20) J.A. Kimber, S.G. Kazarian, and F. Štepánek, 

	(22) J. van der Weerd, K.L.A. Chan, and S.G. Kazarian, 

	(23) O.S. Fleming and S.G. Kazarian, in 

 (Wiley-VCH Verlag GmbH & Co. KGaA, 2006), pp. 205–238.

	(25) T. Frosch, K.L.A. Chan, H.C. Wong, J.T. Cabral, and S.G. Kazarian, 

	(26) J. Pajander, S. Baldursdottir, J. Rantanen, and J. Østergaard, 

	(27) J. Østergaard, J.X. Wu, K. Naelapää, J.P. Boetker, H. Jensen, and J. Rantanen, 

	(28) C.A. Fyfe, H. Grondey, A.I. Blazek-Welsh, S.K. Chopra, and B.J. Fahie, 

	(29) P. Kulinowski, A. Mlynarczyk, P. Dorozynski, K. Jasinski, M.H. Gruwel, B. Tomanek, and W. Weglarz, 

	(30) C.A. Coutts-Lendon, N.A. Wright, E.V. Mieso, and J.L. Koenig, 

	(31) H. Keles, A. Naylor, F. Clegg, and C. Sammon, 

	(34) J. van der Weerd, K.L. Andrew Chan, and S.G. Kazarian, 

	(35) J.A. Kimber, S.G. Kazarian, and F. Štepánek, 

	(36) B. Zheng, L.S. Roach, and R.F. Ismagilov,

	(42) J.B. Cooper, K.L. Wise, W.T. Welch, M.B. Sumner, B.K. Wilt, and R.R. Bledsoe, 

	(43) A.A. Gabrienko, C.H. Lai, and S.G. Kazarian, 

	(44) A. Gabrienko, V. Subramani, O. Martyanov, and S. Kazarian, 

	(45) A.E. Pomerantz, D.J. Seifert, K.D. Bake, P.R. Craddock, O.C. Mullins, B.G. Kodalen, S. Mitra-Kirtley, and T.B. Bolin, 

	(46) T. Tveter, P. Klaeboe, and C.J. Nielsen, 

	(49) K.L.A. Chan, X. Niu, A.J. de Mello, and S.G. Kazarian, 

 is a professor of physical chemistry with the Department of Chemical Engineering at Imperial College London in London, England. Direct correspondence to:  s.kazarian@imperial.ac.uk

Peer-Reviewed Article

Macro attenuated total reflection FT-IR spectroscopic imaging is a powerful and underutilized tool. This article presents an overview of approaches and opportunities for using this method to study dynamic processes such as diffusion, sorption, crystallization, and dissolution.

Macro ATR-FT-IR Spectroscopic Imaging of Dynamic Processes

Metal oxides often occur in crystals where there are no molecular units. In contrast to Raman analysis of organic compounds in which there are molecular vibrating units with functional subunits, many of whose vibrations are isolated from the remainder of the molecule, the analysis of spectra of oxides requires understanding how atoms move in crystal lattices. These oxides occur in diverse materials such as paint and ceramic pigments, corrosion films, catalysts, and minerals. This column installment attempts to introduce analysts who often work with organic materials to the concepts that need to be understood in analyzing these materials.  

The distinction between the vibrational analysis of an organic molecule, which can occur in a crystalline lattice, and that of a metal oxide, which normally occurs only in a crystalline lattice without a clearly defined molecular functional group, is an important one. The analysis of atomic motion is described by a field called 

; the unit cell of the crystalline form is defined and the motion of each atom in the unit cell is determined by solving the potential energy function of the lattice, which is the sum of all pairwise potential energies between each pair of atoms. To solve the equation, the sum is performed only over neighboring atoms (that is, in the unit cell) because the electric fields of distant atoms are assumed to be screened. In addition, the potentials are assumed to be harmonic, which will be true if the motion is small compared to the interatomic spacings. Another assumption is that the electronic motion can be separated from the atomic motion because, being much less massive, the electrons move much faster than the atoms. This is called the 

An important concept in the description of the motion of atoms in crystals (phonons) is the symmetry of the lattice. Group theory enables one to classify a crystal lattice according to what geometrical operations in space leave the lattice unchanged; if there is hexagonal symmetry, for example, and the lattice is rotated by 60°, the lattice will not have been changed. When the phonons are worked out, they will have symmetries consistent with the lattice symmetry. If a phonon leaves the lattice symmetry unchanged, it is described as being symmetric. However, a phonon can distort the lattice in a way that is consistent with its character; this type of phonon can be described by one of the "irreducible representations" of the point group of the lattice. Clearly, this short column cannot cover lattice dynamics, but if readers are sufficiently interested they can take a look at numerous texts that are available (1,2). In addition, two of David Tuschel's recent installments of "Molecular Spectroscopy Workbench" considered some of these concepts in more detail (3,4).   

For analysts who need information on this class of materials, I can recommend a text by Ross (5) that may be difficult to locate, and a more recent text from the European Mineralogical Union (6). There are also emerging resources on the web such as the RRUFF data base (7) and the Romanian Data Base of Raman Spectroscopy (8). 

Before going on, I want to add some remarks on amorphous oxides and "soft modes." Metal oxides do not always occur in a crystalline form. When in an amorphous state, the Raman bands are quite broad, but often can be derived from one of the crystalline forms of the same material. In fact, in the case of ceramics, the transition from the amorphous to crystalline form (devitrification) can be an important phenomenon in terms of the performance of the ceramic. The assumption is that in the amorphous form there is a distortion of interatomic bond angles. Long-range order is lost, but nearest-neighbor interactions are affected so as to shift a particular vibrational frequency; because there is a population of such distortions, the observed band will be broad.   

There is one important case where atomic motion may not be small compared to the interatomic spacings. There are cases where there can be a (crystalline) phase transition in which the lattice changes symmetry. The most cited example is that of BaTiO

, which is cubic at any temperature above 393 K, but is tetragonal at lower temperatures. In the cubic phase, every atom sits on a center of symmetry, but in the tetragonal form the Ti atom in the center of the unit cell is displaced in one of the cubic directions and the octahedral cluster of O atoms are displaced in the opposite direction. It turns out that there is a normal mode that, if frozen, exactly induces this change. When one follows this mode as a function of temperature, its frequency is observed to go lower as the atoms in the lattice move toward their new coordinates. Figure 1, adapted from reference 9, shows the unit cell of BaTiO

Figure 1: BaTiO3 in the distorted tetragonal phase. Adapted with permission from reference 9. 			

Metal oxides occur in crystals in which the O

  anions are often in close-packed sites and the metal cations fit into tetrahedral or octahedral sites, with two tetrahedral sites per O

. To a large extent, this follows from the fact that the ionic radii of anions are significantly larger than that of the cations. In fact, relatively simple geometric arguments based on ionic radii enable predicting whether a particular anion will occupy tetrahedral or octahedral sites (10). Of course, the stoichiometry of the oxide will determine how many metal cations will be present and details regarding the electronic states will play a role determining the overall symmetry of the lattice. 

But the structure for a particular oxide is often not unique. A common example is TiO

, a white pigment that is found in paper, polymers, and paints, and as photoreactive coatings on glass. TiO

 occurs in two common forms known as anatase and rutile, as well as a form known as brookite that sometimes occurs in minerals. As shown in Figure 2, the Raman spectra of these forms are quite different from each other. Aside from the limited intensity in the low 140 cm

  region in the spectrum of rutile that overlaps with the strong band in anatase, these spectra would not be suspected as originating from the same chemical entities.  

Figure 2: Raman spectra of crystalline anatase (top) and rutile (bottom), two forms of TiO2.			

Another metal oxide that has multiple forms is ZrO

. The high temperature phase (>1500 Â°C) is cubic and the first order phonons are not strictly Raman active because every atom sits on a symmetry site, although broad bands do often appear. The normal room temperature phase is monoclinic, but the ceramics engineer might want to create a powder or compacted part in the tetragonal phase (that normally exists at a temperature above room temperature but lower than that of the cubic phase) by alloying with a small amount of CaO or Y

. Figure 3 shows the spectra of the tetragonal and monoclinic phases of ZrO

Figure 3: Raman spectra of the tetragonal (top) and monoclinic (bottom) phases of ZrO2.			

Some years ago Professor William Wright of Penn State University described to me the phenomenon of topotaxy (11), which could be quite interesting, although I have never observed it. In this case, in a system like that of ZrO

, which both have two or more crystal phases, it is possible for the two phases to exist simultaneously in a continuous "crystal." In this case, there are specific crystal planes where the atomic placements are the same for the two phases, and this is what enables one phase to "grow" out of the other.   

Another phenomenon that is sometimes observed in oxide powders is the effect of particle size. There is often an interest in engineering materials of particular dimensions, especially with the current activities in nanotechnology. When the dimensions of crystals get small enough, the phonons experience "quantum confinement." Because the crystals are so small, the normal assumption of translational symmetry does not apply. When this happens, the phonon wavevector is no longer constrained at 0; that is, there is uncertainty in the value of the scattering wavevector. Raman scattering has contributions away from the center of the Brillouin zone.  

Figure 4: Hypothetical dispersion of acoustic and optic phonons in a lattice with two atoms in the unit cell (for example, silicon).			

The concept of the wavevector has not yet been introduced in this column. The 

 describes how the vibration propagates through the crystal. As a result of translational symmetry, there has to be wavevector conservation. That means that the difference between the wavevectors of the incident and scattered light in the crystal has to be transferred to the phonon. For crystals larger than 1 μm, this difference will always be much smaller than the dimension of the Brillouin zone, which scales as the reciprocal of the lattice constant. That is why you may hear it said that Raman scattering in crystals always occurs at the Brillouin zone center. When the crystal becomes much smaller than  1 μm, there is uncertainty in the phonon wavevector because the mean free path of the phonon is larger than the nanocrystal's dimensions. So, there will be contributions to the Raman scattering from phonons away from the center of the Brillouin zone. In most crystalline systems, the phonon energy decreases with the phonon wavevector so one tends to observe the Raman frequency decrease in nanocrystals; the band tends to be broadened and asymmetrical, with a low frequency tail because one is observing contributions from particles of varying sizes. This phenomenon has often been observed in crystal silicon that had been annealed from the polycrystalline state (12,13). 

Figure 5 shows spectra of elemental silicon in various stages of crystalline order. Although this material is not an oxide, the spectra illustrate the phenomena that we have been discussing. The bottom spectrum is that of amorphous silicon, the top of crystalline material. The strong band assigned to the optical phonon in the crystal at 520 cm

  is broadened and down-shifted in the amorphous material. The amorphous phase also shows a broad band centered near 150 cm

, which can be assigned to scattering by phonons that are not restricted by symmetry arguments of the crystal because there is no translational symmetry.  

Figure 5: Raman spectra of crystalline (top) and amorphous (bottom) silicon as well as two samples with the crystalline silicon peak down-shifting because of particle size (middle two spectra), with evidence for a mixture with the amorphous phase in the spectrum second from the bottom.			

One other curious thing can be said about observations in nanocrystals. It is possible in some instances for the optical phonon dispersion curve to curve up before it goes down, and this has been observed in anatase (14). In this case the phonon will be shifted up in frequency and broadened on the high frequency side.   

To add to the usefulness of this article, I will discuss the oxides of iron along with some related phases. Iron is the metal element that is used most prevalently for mechanical structure. However, it easily forms a surface oxide and the most common surface oxides ("rust") do not stick to the metal. That means that the metal continues to oxidize, reducing the metal content continuously. For uses that can support additional cost, the metal is alloyed with additional metals, especially chromium, which has the benefit of "passivating" the surface. That means that a layer of Cr

 forms on the surface and sticks. Figure 6 shows the spectra of four common oxides of iron as well as Al

, is the common form of rust. When it is highly crystalline, the peak near 225 cm

, crystallizes in the spinel form, which can accommodate a range of metals in solid solution, especially transition metals; it is easily distinguishable from hematite, but its spectrum is generally weaker and the bands are broader and they will be shifted when it occurs as a solid solution. A third oxide, FeO, crystallizes in the rock salt structure in which every atom sits on a site of symmetry, and is thus Raman-inactive. Geothite and lepidocrocite are two forms of FeOOH, and they are easily distinguishable in their spectra. The spectrum of Cr

 was recorded from a powder from a bottle with no information on elemental composition. It is well known that metal oxides can form solid solutions, so the frequencies shown in this spectrum may not be accurate for pure Cr

has a very strong photoluminescence doublet at 693 nm that is used as a strain calibrant. The spectrum of Al

. The ability to differentiate these species enables the use of Raman microscopy to aid in the elucidation of corrosion mechanisms. Also, because measurements can be made in situ, one can follow corrosion in water, brine, or at elevated temperatures. 

Figure 6: Raman spectra of the two common iron oxides, two common iron hydroxides, Cr2O3 and Al2O3, which are isomorphic to Fe2O3. From top to bottom they are Cr2O3, hematite, lepidocrocite, goethite, magnetite, and sapphire (Al2O3).			

It is fairly well known that many heterogeneous catalysts are prepared by creating monolayers of defined transition (or other) metals on silica, alumina, ceria, zirconia, titania, or magnesia. Although there are numerous groups working in this field using Raman spectroscopy, in conjunction with other techniques, to elucidate the catalytic reaction, I am going to cite the work from Lehigh University because that is what I know best (15).  

It is understood that a heterogeneous catalyst works by providing electronic sites on the surface that react with organic molecules that come into contact with these surfaces. In his review of the field in 2002, Wachs (15) described the surfaces as having monoxo, dioxo, and trioxo metal ions, and the geometry and electronic properties of these species will then determine the reactivity of the surface.   

In that article (15), Wachs proposed that exchanging the oxygen on the surface with 

O would enable differentiating the possibilities. Mono species would show two Raman bands in isotopically substituted samples, dioxo species would show three bands, and so on. Thus, Raman spectroscopy has the ability to uniquely describe the surface configuration of a heterogeneous catalyst! However, to do this the catalytic surface has to be controlled by controlling the temperature and chemical state of the surface in a chemical reactor. Bañares and Wachs (15) argued that the catalytic performance can be understood by controlling the temperature and chemistry of the environment while making measurements (Raman, infrared, gas chromatography, and so on). In particular, molecular level information regarding the surface metal oxides are revealed by Raman spectroscopy. In this article, Bañares and Wachs systematically studied the behavior of numerous catalysts as a function of temperature cycling, reduction, and hydration to put together a story about the structure and reactivity of the catalytic surfaces.   

In contrast to a Raman study of a bulk crystalline or amorphous form of a metal oxide in which spectral bands correspond to atomic motions representing long or mid-range order in the solid, in the case of catalysts, the Raman spectrum represents the structure of a surface. That means that without nanometer spatial resolution one can infer the geometrical or chemical properties of atoms at a surface by interpreting the Raman spectrum! With this information one can understand the reactivity of a surface, and then engineer a catalytic surface to perform a desired task. This takes catalyst engineering from empirical-try-everything to rational design.   

We have reviewed some of the major characteristics of Raman scattering in bulk oxides (crystalline, and to some extent amorphous) and in heterogeneous catalytic materials. Depending on the application, the spectrum can provide identification of pigment, elucidation of a corrosion process or surface oxide treatment, characterization of amorphous or nanocrystalline phases, and characterization of heterogeneous catalysts in which the active phase is a properly prepared surface metal oxide. 

(1) A.A. Marududin, E.W. Montroll, and G.H. Weiss, "Theory of Lattice Dynamics in the Harmonic Approximation," 

Crystal Symmetry, Lattice Vibrations, and Optical Spectroscopy of Solids, A Group Theoretical Approach 

(World Scientific Publishing Co. Ptc. Ltd., Singapore, 2014). 

Raman Spectroscopy Applied to Earth Sciences and Cultural Heritage

, J. Dubessy, M.C. Caumon, and F. Rull, Eds., Vol 12 of the 

European Mineralogical Union Notes in Mineralogy

 (European Mineralogical Union and the Mineralogical Society of Great Brittain and Ireland, London, 2012). 

Structural Phase Transitions - I.Landau Theory, Advances in Physics 

Inorganic Chemistry, Principles of Structure and Reactivity

(11) Professor William Wright, private communication. 

(12) G. Kanellis, J.F. Morhange, and M. Balkanski, 

(14) S. Mamedov, T. Moriyama, T. Numata, N. Nabatova-Gabain, S. Yanagida, and L. Yan, 

 The original, published version of this article contained an error in this paragraph: Another metal oxide that has multiple forms is ZrO

. The normal stable room temperature phase is cubic, which is not Raman active because every atom sits on a symmetry site. However, for chemical or physical reasons a ceramics engineer might want to create a powder or compacted part with one of the other two phases. ZrO

can be stabilized in the tetragonal or monoclinic phases by alloying with a small amount of CaO or Y

Molecular Spectroscopy Workbench

Metal oxides often occur in crystals where there are no molecular units. Here, we provide an introduction to the concepts that need to be understood when analyzing metal oxides in materials such as paint, ceramic pigments, corrosion films, catalysts, and minerals.

Raman Spectra of Metal Oxides

Issue PDF

Click the title above to open the Spectroscopy October 2014 regular issue, Vol 29 No 10, in an interactive PDF format.

Spectroscopy-10-01-2014

LIBS Basics, Part III: Deriving the Analytical Answer — Calibrated Solutions with LIBS

Macro ATR-FT-IR Spectroscopic Imaging of Dynamic Processes

Raman Spectra of Metal Oxides

Vol 29 No 10 Spectroscopy October 2014 Regular Issue PDF