Jerry Sellors

This investigation describes how Fourier transform near-infrared (FT-NIR) spectroscopy can be deployed using a workflow approach to overcome a range of different challenges in raw material identification while also meeting the stringent requirements of regulated environments.

	Modern near-infrared (NIR) spectroscopy instrumentation and its interchangeable sampling modules offer a great deal of versatility to handle the diverse range of solid, liquid, and gel-based raw materials encountered at various stages of the pharmaceutical manufacturing process. Rapid characterization of a material by its NIR spectrum can be achieved by comparison against a series of reference spectra, which allows for unambiguous identification of the raw material, and discrimination from other similar compounds.

	This investigation describes how Fourier transform (FT)-NIR spectroscopy can be deployed using a workflow approach to overcome a range of different challenges in raw material identification while also meeting the stringent requirements of regulated environments. No sample preparation or dilution is required, and the measurements can be performed directly through glass vials or Petri dishes using an NIR reflectance sampling accessory.

	The NIR spectral bands of raw materials are caused by the combination and overtone bands derived from the fundamental vibrations in the mid-infrared (IR) region of the spectrum, and are often unique fingerprints for that material. Spectra of three typical raw materials used in the pharmaceutical industry are shown in Figure 1.


	Figure 1: Typical NIR spectra of pharmaceutical raw materials: diclofenac (black), poloxamer (red), and talc (blue).

	These spectra allow for the identification of the material by comparison against a series of reference spectra using appropriate decision-making algorithms, based on the complexity of the analysis. Many standard raw material identification methods use a simple spectral correlation algorithm, which will measure the correlation of an unknown spectrum against a series of reference spectra and report the closest material to the unknown (1). For example, a score of 1 is a perfect match, whereas a score of 0 shows no correlation between the sample and reference spectra. When analyzing materials, it is therefore important that the analysis is focusing on the spectral features of the materials. However, when using this method problems can arise calculating the similarity of spectra when encountering

		reproducibility of sampling errors (such as those associated with repacking of solid powdered sample),

	These effects can be decreased by using a number of mathematical filters that reduce the contribution where the information tends to be unreliable. The net result of these filters is to enhance material-specific spectral features while minimizing spectral contributions related to instrument, sampling, and atmospheric conditions.

	These methods allow pass–fail thresholds based on both correlation with the reference materials and discrimination from the second-best matches in the reference materials, ensuring that the correct material is identified and no false positives are obtained from other reference materials.

soft independent modeling class analogies

), which is a chemometric approach that models the variation within the collection of reference spectra for a given material as well as the difference between spectra of different materials (2). This algorithm helps overcome the issues described above. This capability allows SIMCA to be sensitive to small spectral differences produced by impurities, even in the presence of normal batch-to-batch or sampling variations. For ease of implementation, either of these algorithms is typically deployed using a workflow approach through the system software of the spectrometer.

	To show the practical implementation of these two algorithms, experiments were performed representing the typical challenges encountered in raw material identification measurements. The instrument used for this investigation was the Spectrum Two N FT-NIR spectrometer and its NIR reflectance module (PerkinElmer Inc.). The system's software was used for data modeling, troubleshooting, and method validation. The system includes workflow software for combining data collection, analysis, and reporting into simple software routines, with an enhanced security option, for assistance when working in a regulated environment.

	The three experiments, which require a different algorithmic approach to the data analysis, are described below.

	NIR Spectra of 34 chemically different solid raw materials were measured. The raw material "library" contained a small number of active ingredients, and the remainder were excipients. Since these materials are chemically different, their NIR spectra should be sufficiently different to allow easy identification of the material using the COMPARE algorithm supplied with the spectrometer.

	Spectra of seven different grades of Avicel (FMC), a microcrystalline cellulose compound used as an excipient, were measured. Since all of these materials are chemically the same, but varied in physical properties, a more powerful algorithm, SIMCA, was used to discriminate the materials.

	Identification of an unexpected failure material. In cases where a material fails the raw material identification, further investigative analysis is required using library searching against commercial spectral libraries.

	The operating conditions for all three experiments are shown in Table I.

Experiment 1: Raw Material Identification

	NIR reflectance spectra were collected for a series of 34 different powdered raw materials to be used as a reference spectra library for the raw material identification method using the COMPARE algorithm. A series of nine independent validation spectra were collected consisting of three samples of povidone from different batches, three samples of Avicel from different batches, and one sample each of calcium ascorbate, hydroxypropyl methyl cellulose (HPMC), and magnesium stearate. These validation samples were obtained from different suppliers to the original reference set. Spectra of these five very different material types are shown in Figure 2.


	Figure 2: Spectra of validation materials: Avicel (black), povidone (red), calcium ascorbate (blue), HPMC (purple), and magnesium stearate (green).

	The pass–fail criteria were set to a correlation value of 0.98 and a discrimination value of 0.05. All of the materials exceeded the correlation and discrimination thresholds and were therefore classed as pass materials. This is demonstrated in Table II.


	Table II: Comparison of results for validation materials

	Since the raw materials are chemically very different, their NIR spectra, combined with the COMPARE algorithms, were easily capable of identifying the materials.

	Further method validation was performed for the Avicel using three batches of the same grade of Avicel PH103. A sample from each batch was placed in a glass vial. Each glass vial was then measured in triplicate shaking the sample in the vial and replacing it on the NIR reflectance module. These nine measurements, with statistical data analysis, are shown in Table III.

Experiment 2: Discrimination of Closely Related Materials

	Avicel PH is a microcrystalline cellulose product with the PH designating that it is suitable for pharmaceutical use. It has excellent compressibility and is used for tablet compression and also for wet granulation. There are several different grades of Avicel PH that differ mainly in particle size and moisture content. Therefore, the correct choice of the right grade is essential to achieve the right compressibility required for making a tablet or low moisture content for use with moisture-sensitive active ingredients. Since the Avicel materials are all the same chemical compound, their NIR spectra are almost identical making it difficult to easily discriminate the different grades. However, NIR spectra are affected by particle size and spectral features because of moisture observed.

	The NIR spectra of a series of samples with different Avicel PH grades were used to develop a SIMCA model to allow confirmation of the grade of Avicel raw materials. Three samples were taken for each of the following grades of Avicel: PH101, PH102, PH103, PH105, PH113, PH301, and PH302. For each of the three samples, triplicate measurements were made on the NIR reflectance module, representing nine spectra per grade and a total of 63 total spectra for the seven grades. The spectra of all the samples are shown in Figure 3.

	The COMPARE algorithm correctly identified all of the materials as Avicel. However, it was not able to discriminate between the different grades with all grades exceeding the pass–fail correlation limit. SIMCA uses a chemometric approach to spectral modeling and is capable of accounting not only for the differences between different materials, but also for the variability within the same material because of batch differences. The 63 spectra of the Avicel samples were input into a SIMCA method to try to separate out the different grades of materials, and the resulting score plot is shown in Figure 4.


	Figure 4: Principal component scores plot showing separation of Avicel grades.

	Each of the spheres in the plot represents the data from each of the Avicel grades; the further apart the spheres (larger intermaterial distance), the better the chance of discrimination. A Coomans plot, as shown in Figure 5, shows the model distances between two classes of material-in this case, Avicel PH101 and PH102 exhibit clear separation. In this model, there are no overlaps between the grades, inferring that there is less chance of misclassification.


	Figure 5: Coomans plot showing separation of classes of materials.

	This experiment demonstrates the power of the SIMCA algorithm, which is capable of distinguishing between chemically identical substances having different physical properties. In this example, SIMCA successfully separated seven different grades of Avicel in which the grades differed only in particle size and moisture content. The COMPARE algorithm was capable of correctly identifying the samples as Avicel, but was unsuccessful at distinguishing between the different grades.

Experiment 3: Identification of an Unexpected Failure Material 

	Occasionally, an NIR raw material analysis can result in a failure requiring further investigation. This can occur, for example, when a new supplier is selected for a particular raw material or, in extreme cases, where the wrong material has been supplied. Fortunately, this is a rare occurrence, but it requires further investigation and correct identification of the suspect material. Commercially available pharmaceutical NIR spectral libraries are available containing more than 1300 spectra of excipients, drugs and active substances, and other supporting chemicals used in the industry. Using these spectral libraries, together with an optimized search algorithm, allows for the seamless identification of unknown materials.

	An unlabeled powder material was tested in a glass vial using the NIR Reflectance Module. The NIR spectrum collected was tested using the COMPARE set of 34 reference standards described previously. The sample failed the COMPARE analysis with the best hit being dextrose with a score of approximately 0.48, well below the acceptance threshold of 0.98. This is exemplified in Figure 6, which shows the unlabeled powder spectrum (black), compared to the best hits dextrose (red) and Avicel (green).


	Figure 6: COMPARE algorithm results for the suspect sample (black), compared with the best hits dextrose (red) and Avicel (green).

	The same spectrum was then analyzed using the Search algorithm comparing the unknown against the NIR pharma spectral database from ST Japan (3), containing 1340 spectra. Using the Search algorithm, the material was correctly identified as D-mannitol with a Search score of 0.99. This is exemplified in Figure 7, which shows the spectrum for the unknown sample (black) and the D-mannitol (red).


	Figure 7: Search result identifying the suspect material (black) as D-mannitol (red).

	NIR spectroscopy has proven to be a fast and easy technique in the pharmaceutical industry for raw material identification. Materials can be measured within seconds using sampling techniques, such as NIR reflectance in glass vials, which require no sample preparation. NIR spectra are specific to the material being analyzed, allowing samples to be tested against databases of standard reference spectra. A choice of algorithms is available for routine identification of chemically different materials, advanced analysis of very similar materials, and identification of unknown materials. All of these approaches can be integrated into a workflow approach to both increase productivity of analysis and reproducibility of results.

	(1) National Cancer Institute, Division of Cancer Treatment and Diagnostics, COMPARE Analysis: 

https://dtp.cancer.gov/databases_tools/compare.htm

, B.R. Kowalski, Ed. (ACS Symposium Series 52, Washington, D.C., American Chemical Society, 1977), pp. 243–282.

http://www.stjapanusa.com/Products/SpectraDatabases/tabid/165/Default.aspx

 are with PerkinElmer Inc., in Seer Green, UK. Direct correspondence to: 

Articles

FT-NIR spectroscopy can be used to overcome a range of challenges in raw material identification while also meeting the stringent requirements of regulated environments.


Verification of Pharmaceutical Raw Materials Using FT-NIR Spectroscopy

When images are measured in reflectance, the observed structure is not necessarily just that of the sample surface. The two-dimensional (2D) images produced by near-infrared (NIR) and mid-IR imaging are representations of three-dimensional (3D) objects. In reflectance images the effective spatial resolution is affected by how far the radiation penetrates into the sample as well as by diffraction and the pixel size. NIR images are typically dominated by diffuse reflection with the result that in practice spatial resolution is limited by the penetration depth. Mid-IR reflectance images may contain both specular and diffuse components. These have different spatial resolutions as the specular component shows purely surface structure. In attenuated total reflectance (ATR) images the resolution has a component that depends on the distribution of components within a surface layer corresponding to the penetration depth, which is proportional to wavelength. We illustrate these different contributions to the observed resolution with images of materials with known 2D and 3D structures.

What do we measure in reflection? For the majority of samples, transmission measurements are not possible so spectra have to be obtained by reflectance. This involves either external reflection, often divided into specular and diffuse reflection, or attenuated total reflection (ATR). What you see in external reflection is the sum of what is reflected from the surface (specular or Fresnel reflection) and what penetrates into the sample before emerging (diffuse reflection). You see only surface reflection from a continuous uniform material where radiation that penetrates is not scattered back to the surface. Very strongly absorbing materials also show only surface reflection as the radiation that penetrates is all absorbed. With scattering materials some radiation that penetrates the material returns to the surface by encountering interfaces where there is a change of refractive index. The relative amounts of surface and diffuse reflection depend on the absorptivity and particle sizes of the sample and are wavelength dependent. Small particle size and weak absorption increase the amount of diffuse reflection. In practice, diffuse reflection dominates in the near infrared (NIR) whereas spectra in the mid-IR often show significant contributions from both surface and diffuse reflection.   

How does this affect spatial resolution? In an imaging experiment, separate spectra are measured for each pixel while the incident radiation covers the whole area. If reflection comes only from the surface, the spatial resolution should depend simply on diffraction and the pixel size. However, any radiation that penetrates a scattering sample takes a random path through the material and will not all return to the point at which it entered the sample. Thus, the radiation contributing to any pixel in the image is a combination of rays from some volume underneath the point on the surface that is being observed. The deeper the radiation can penetrate into the sample and return without being absorbed, the larger this volume will be. If we assume isotropic scattering, the average lateral displacement from the point of incidence to where the radiation emerges will be similar to the average depth to which the radiation penetrates. This will cause blurring of the image by an amount that depends on the average penetration depth. There have been several attempts to estimate the depth that the radiation penetrates (1,2). These suggest average values in the range of 50–100 µm for NIR observations of materials such as pharmaceutical tablets. As a result, the spatial resolution achieved in NIR reflection images is likely to be in this range. This is in contrast to what is observed with the metal-on-glass reflection targets normally used where diffraction limited spatial resolution is better at the shorter wavelengths of the NIR region than in the mid-IR. 

 Figure 1: Diffuse reflection and transmission.			

One way to estimate the penetration depth is by comparing the band intensities in transmission and reflection for a scattering sample. The absorbance is proportional to the distance traveled inside the sample. In transmission, this distance is proportional to the thickness of the sample, but in reflection it is proportional to twice the average penetration depth (Figure 1). One sample used was an acetaminophen tablet ground down to a thickness of 1.1 mm. The ratio of the absorbances at 8800 cm

  measured in transmission and reflection is 18:1. From this, the average penetration depth is calculated as 30 µm (Figure 2). Similar measurements on pressed pellets of other materials give average depths of 40 µm to 60 µm (Table I). Because these are average depths we can conclude that some rays must penetrate significantly further. This will produce blurring of 50–100 µm. 

Figure 2: Spectra of an acetaminophen tablet.			

Measuring Spatial Resolution in Scattering Materials

We prepared a sample by pressing a pellet of microcrystalline cellulose on top of a pressed pellet of polyethylene, giving a flat boundary between the two materials. Microtoming one edge of the pellet in a direction parallel to the interface gave a flat sample with the boundary perpendicular to the surface. Images were obtained with a PerkinElmer Spotlight system with a pixel size of 6.25 µm over the 6000–1400 cm

  range (Figure 3). The spectra include dispersive features characteristic of Fresnel surface reflection such as the polyethylene CH

, strong absorptions such as the broad cellulose OH band at 3500 cm

Table I: Estimated NIR penetration depths			

As expected, an image based on the polyethylene CH

  (Fresnel reflection) has a sharp boundary. In addition, it reveals an anomaly on the surface of the polyethylene layer at a distance of about 50 µm from the interface. An image of the strong cellulose OH absorption at 3600–3400 cm

  shows a similar sharp boundary and indicates that the anomaly results from a flake of cellulose, something that is confirmed from the individual spectra. The sharp boundary illustrates how the strong absorption limits radiation that penetrates the sample returning to the surface. In contrast, the weak absorptions of the NIR region lead to much more diffuse images with features from one layer visible up to 100 µm away from the boundary. The polyethylene layer has significantly less scattering than the cellulose layer, which probably explains why the cellulose features appear to extend farther beyond the boundary. 

Figure 3: Images from a two-layer sample.			

What Is the Penetration Depth in ATR Imaging?

The effective penetration depth is related to the wavelength λ, the refractive indices of crystal and sample 

 is the refractive index of the crystal, θ is the angle of incidence, and 

 is the ratio of refractive indices of sample and crystal. 

In the converging beam of a microscope there is a range of angles of incidence, so it is appropriate to determine the effective penetration depth experimentally. We have compared the absorbance of liquid paraffin measured by ATR and by transmission of a film with known thickness. In the Spotlight system, using a germanium crystal the penetration depth with an organic material of refractive index 1.4 

 is expected to vary from about 0.6 µm at 3000 cm

. Comparison with a transmission spectrum of paraffin oil demonstrates that the equivalent pathlengths for ATR are fairly similar to this (Table II). 

Table II: Effective pathlengths for ATR imaging			

When the boundary between regions is not normal to the surface, the penetration depth will affect the spatial resolution. To demonstrate this we prepared samples consisting of two layers, polystyrene and an acrylic resin. In one the boundary was normal to the surface and in the other the boundary was at 10° to the surface (Figure 4). In the latter, the polystyrene layer 10 µm from the boundary at the surface is 1.8 µm thick. Images of the two were obtained with a pixel size of 1.56 µm.  

Figure 4: ATR images of two-layer samples.			

Images of the two samples show that the edge appears much more diffuse in the sample where the boundary is at 10° to the surface (Figure 4). Penetration through the polystyrene layer a lows the acrylic material to be detected up to 10 µm away from the apparent boundary on the surface. Intensity profiles of the acrylic resin at different wavelengths clearly show how longer wavelengths detect the subsurface layer at greater depths. The profile at 1137 cm

  is very unsymmetrical, extending about 15 µm beyond the surface boundary (Figure 5). 

In looking at reflection images it is important to remember that samples are three dimensional and that spectra may be generated from a finite volume, not just from the immediate surface. The boundaries between regions may, therefore, appear broader than those seen in visible images. This is most apparent in near-IR images where the resolution is likely to be limited by lateral scattering within the sample rather than by diffraction or the pixel size. In ATR images, the penet-ration depth allows spectral contributions to be seen from subsurface material. 

 are with PerkinElmer in Seer Green, Bucks, United Kingdom. Direct correspondence to: 

 (2) S.J. Hudak, K. Haber, G. Sando, L.H. Kidder, and E.N. Lewis, 

When images are measured in reflectance, the observed structure is not necessarily just that of the sample surface.

Spatial Resolution in IR Reflectance Imaging — NIR, Mid-IR, and ATR

What do we measure in reflection? For the majority of samples, transmission measurements are not possible so spectra have to be obtained by reflectance. This involves either external reflection, often divided into specular and diffuse reflection, or attenuated total reflection (ATR). What you see in external reflection is the sum of what is reflected from the surface (specular or Fresnel reflection) and what penetrates into the sample before emerging (diffuse reflection). You see only surface reflection from a continuous uniform material where radiation that penetrates is not scattered back to the surface. Very strongly absorbing materials also show only surface reflection as the radiation that penetrates is all absorbed. With scattering materials some radiation that penetrates the material returns to the surface by encountering interfaces where there is a change of refractive index. The relative amounts of surface and diffuse reflection depend on the absorptivity and particle sizes of the sample and are wavelength dependent. Small particle size and weak absorption increase the amount of diffuse reflection. In practice, diffuse reflection dominates in the near infrared (NIR) whereas spectra in the mid-IR often show significant contributions from both surface and diffuse reflection.

Figure 1: Diffuse reflection and transmission.			

 measured in transmission and reflection is 18:1. From this, the average penetration depth is calculated as 30 µm (Figure 2). Similar measurements on pressed pellets of other materials give average depths of 40 µm to 60 µm (Table I). Because these are average depths we can conclude that some rays must penetrate significantly further. This will produce blurring of 50–100 µm. 

 range (Figure 3). The spectra include dispersive features characteristic of Fresnel surface reflection such as the polyethylene CH

 (Fresnel reflection) has a sharp boundary. In addition, it reveals an anomaly on the surface of the polyethylene layer at a distance of about 50 µm from the interface. An image of the strong cellulose OH absorption at 3600–3400 cm

 shows a similar sharp boundary and indicates that the anomaly results from a flake of cellulose, something that is confirmed from the individual spectra. The sharp boundary illustrates how the strong absorption limits radiation that penetrates the sample returning to the surface. In contrast, the weak absorptions of the NIR region lead to much more diffuse images with features from one layer visible up to 100 µm away from the boundary. The polyethylene layer has significantly less scattering than the cellulose layer, which probably explains why the cellulose features appear to extend farther beyond the boundary. 

When the boundary between regions is not normal to the surface, the penetration depth will affect the spatial resolution. To demonstrate this we prepared samples consisting of two layers, polystyrene and an acrylic resin. In one the boundary was normal to the surface and in the other the boundary was at  10° to the surface (Figure 4). In the latter, the polystyrene layer 10 µm from the boundary at the surface is 1.8 µm thick. Images of the two were obtained with a pixel size of 1.56 µm. 

 is very unsymmetrical, extending about 15 µm beyond the surface boundary (Figure 5).  

(2) S.J. Hudak, K. Haber, G. Sando, L.H. Kidder, and E.N. Lewis, 

In these types of IR imaging, various factors contribute to the observed resolution. These factors are illustrated with images of materials with known structures.