Chemical acquisition and Raman spectral processing methods used with the Morphologi 4-ID

Introduction

The chemical analysis part of a Morphologi 4-ID measurement involves acquiring a Raman spectrum from each targeted particle. Acquired particle spectra are compared with library spectra and a correlation score is calculated to determine the chemical nature of each particle. Correlation scores close to 1 indicate a strong match to the reference spectrum, whereas scores close to 0 mean no match.

The Morphologi 4-ID software offers a number of spectral processing steps in order to improve chemical identification and measurement robustness. Figure 1 shows the processing steps available in the order that they are carried out.

Figure 1: Spectral signal processing options

The choice of steps depends on the application and the regions of the spectrum of interest. This technical note explains the different options and offers guidance on which options to use for different applications.

Spectral acquisition and creating a reference library

A reference library must be created for the correlation to be performed against. This can be updated, and data sets reanalyzed against different libraries after Raman acquisition.

Spectra for the library can be acquired in two ways:

1. By acquiring point spectra using the manual microscope.
2. By automatically gathering spectra from particles in an automated analysis and then selecting the spectra to be used.

Ideally, pure samples of the materials should be prepared in the same way as the full formulation. 

Acquiring reference spectra in the manual microscope allows optimization of the exposure time. The required exposure time depends on the Raman scattering efficiency of each material, the particle size range and the sample preparation method. If there is more than one component of interest, the exposure time should be determined by the weakest Raman scatterer.

The spot size of the laser with the 50x is approximately 2 μm. Acquisition from smaller particles will have a lower signal-to-noise ratio, thus the exposure time needs to be sufficient for the smallest particles of interest. If a wet sample preparation method is used, a longer exposure time will be required, as the laser has to pass through the coverslip.

Figure 2: Overlay of spectra of a drug particle and an excipient particle (in wet sample preparation)

Figure 2 shows an overlay of spectra from a pharmaceutical nasal spray formulation where the red line is the spectrum from the drug ingredient and the blue line is the spectrum from the excipient. The excipient is a very weak Raman scatterer and as a result its spectrum is similar to the background spectrum of the quartz substrate. However, the drug gives strong spectral features in the 1350 cm^-1 to 1800 cm^-1 region of the spectrum. In this case, it is only the drug component that is of interest so the acquisition parameters should be optimized for this.

Note: If the acquisition times need to be amended, measurements will need to be repeated.

Background correction and/or spectral range masking

The Morphologi software allows the removal of spectral contributions due to the sample substrate (typically high-purity quartz) to allow spectral features to be more easily identified for samples that have weak Raman spectra.

Alternatively, if the main spectral features of the material are well-defined and well-separated from any other features, spectral masking can be used instead of background correction.

In the nasal spray example, spectral range masking could be used so that the correlation is only carried out over 1350 cm^-1 to 1800 cm^-1 where the features of the drug are most prominent. Figure 3 shows the drug spectrum, masked so the correlation calculation will only be performed over the region marked in white and will ignore the shaded region.

Figure 3: Drug spectrum masked so the correlation calculation is only performed over the region marked in white

If background subtraction is preferred there are two options in the Morphologi software: simple subtraction and substrate subtraction.

Simple background subtraction is an un-scaled subtraction. The background spectrum used must be acquired with the same spectrometer settings as the measurement. Substrate background subtraction uses a background which is scaled to the signal based on its similarity. The more similar the signal and the background the greater the correction ratio applied. This means that a spectrum that has a lot of background signal in it (for example from very small particles) will be corrected more heavily than a spectrum with very little background in it (for example from large particles). It also means that the acquisition time for the background spectrum can be different from the acquisition time of the sample measurement. The substrate background subtraction is appropriate for most applications using the quartz substrates.

Figure 4: Overlay of raw drug spectrum (green), the quartz background spectrum (blue) and the corrected drug spectrum (red)

The overlay in Figure 4 shows a raw drug spectrum (green), the background spectrum which is used in the correction (blue) and the corrected drug spectrum (red).

Advanced preprocessing options - smoothing

To reduce noise in the spectrum but preserve the underlying signal Savitsky-Golay filtering is used. This process can be thought of as a weighted windowed average where each point on the smoothed signal is calculated by taking a weighted average of the corresponding source value and the values in a window around it. Figure 5 illustrates that the smoothed value for the point of the graph at the green line depends on the values of all of the data points within the blue box; the filter width determines the width of the blue box.

Figure 5: Illustration of windowed smoothing

The Morphologi offers three smoothing levels, narrow (over 7 points), intermediate (over 31 points) and wide (over 57 points). The smoothing width determines how many surrounding source values are included in the weighted average.

Using spectral range masking affects the smoothing, so this setting may be different when range masking is used compared to the entire spectrum. The signal is smoothed over a window of values on either side and there can be processing artifacts at the ends of the overall signal range (for example, a window of 7 data points will not be able to properly smooth the first and last three data points). Artifacts at the very edge of a spectrum are not an issue for correlation across the whole spectrum, but if spectral range masking is used to reduce the range over which that data is being processed, small artifacts can have a more significant effect relative to the total signal.

Figure 6: Example of wide, intermediate and narrow smoothing

Figure 6 shows an example spectrum with the various smoothing options applied. Choosing narrow smoothing reduces processing artifacts but will increase noise effects throughout the spectrum. More noise in the processed signal with narrower smoothing means the correlation scores will be less robust. In general, it is best to start off with wide smoothing but consider narrower smoothing if spectral range masking is used to exclude a significant portion of the signal.

Advanced preprocessing options - derivative

The smoothing option described above determines how much smoothing is applied to spectra, to suppress noise while retaining the signal, before the derivative processing is applied.

The derivative operation produces either the 1st or 2nd derivative of the smoothed signal. The 1st derivative shows the rate of change in the source signal, essentially the angle of the slope of the source signal. The 2nd derivative is the rate of change of the 1st derivative.

Figure 7: 1st derivative (red) of a synthetic signal (blue)

Figure 7 illustrates a synthetic signal (blue) with a constant upward slope, a blunt peak at 1100 cm^-1 followed by a constant downward slope then a steady baseline. The derivative (in red) starts at a constant positive value which matches the constant rate of climb of the signal. At the peak of the signal, the derivative drops through zero to a negative value which matches the downward slope of the source signal. When the source signal levels out, the derivative flattens out at zero, regardless of the height of the signal line.

On a real spectrum (soda glass) (Figure 8) the spectrum (blue) rises very gently from 200 cm^-1 to 1200 cm^-1 so the derivative is weakly positive. When the spectrum begins to climb more steeply after 1200 cm^-1, the derivative also climbs. As the spectrum levels out near the peak, the derivative drops. At the peak, the derivative is zero then goes negative as the spectrum slopes downward. The downward slope is gentler than the upward slope, so the derivative is less negative than it was positive on the upslope.

Figure 8: 1st derivative (red) of the spectrum of soda glass (blue)

2nd derivative processing applies the derivative operation twice. Figure 9 shows the source spectrum in blue, the 1st derivative in red and the second derivative in green. The green line is showing the rate and direction of change in the 1st derivative.

Figure 9: 1st (red) and 2nd (green) derivatives of a synthetic signal (blue)

This shows that 1st and 2nd derivatives are both effective in determining points of interest in a spectrum where the gradient of the spectrum is changing. The 1st derivative is sensitive to a varying baseline whereas the 2nd derivative is less sensitive to broad features in a spectrum.

Figure 10: 1st (red) and 2nd (green) derivatives of a soda glass signal (blue)

Figure 10 shows the 1st (red line) and 2nd (green line) derivatives of the soda glass signal. The soda glass spectrum has one broad peak and there are no sharp peaks. A 1st derivative would be more appropriate as it is more sensitive to broad features.

Figure 11: 1st (red) and 2nd (green) derivative of a polystyrene-soda glass signal (blue)

Figure 11 shows a spectrum from polystyrene on soda glass. The 1st derivative in red shows the polystyrene signature peak at just over 1000 cm^-1 but gives more weight to the soda glass signature as it is 'broader'. The 2nd derivative also shows the soda glass signature but gives more weight to the polystyrene signature because it is 'sharper'. In this case, if the polystyrene is the component of interest, it is the sharp peak at 1000 cm^-1 that is important and using the 2nd derivative is more appropriate than the 1st derivative.

In general, if the material of interest has sharp features, then the 2nd derivative is likely to be more effective, whereas the 1st derivative is likely to be more appropriate for samples which give a spectrum with broad features. Table 1 summarizes guidance as to which processing options should be used for different applications.

Summary of advanced preprocessing options

Table 1 gives suggestions for when the advanced preprocessing options should be used. 

Table 1: Guidance on when to use smoothing and derivative processing options

Classification based on correlation scores

The correlation scores can be used to classify particles according to their chemical identity. The Morphologi software allows particles to be classed according to their best correlation to the library spectra or according to specific values, or a combination of both. Once the appropriate combination of spectral processing (background correction, smoothing, derivative, and spectral range masking) has been applied it is important to look at the spectra of particles with different correlation scores to define which is most appropriate for the specific application.

Figure 12: Overlay of particle spectra with different correlation scores to the reference library component (red).

Using the same pharmaceutical nasal spray example from earlier, Figure 12 shows an overlay of the drug library spectrum (component 1 in red) and three particle spectra with different correlation scores. For this example, spectral masking has been applied so only the region between 1300 and 1800 cm^-1 was included in the correlation. A particle with a correlation score of 0.71 to the drug reference spectrum still gives a distinguishable peak in this region.

For this application particles were classed as being drug if their individual spectrum correlated to the drug library spectrum with a score greater than or equal to 0.7.

In practice, the value to use will depend on the relative scores of the library component of interest relative to the other components in the library. Where there are multiple spectra in the reference library the best correlation score can be used to classify particles and a combination of best correlation and a minimum correlation score can be the most robust method of classification. Visual verification of the correct classification of particle Raman spectra is always recommended during method development.

Conclusion

The Morphologi 4-ID offers a range of spectral processing techniques, including background subtraction, spectral masking, smoothing, and derivative processing, enabling users to optimize their analyses for a wide range of applications.

The guidance provided in this technical note describes how to select the appropriate processing options. Characteristics of the material spectra, the sample preparation method and the goal of the analysis are key to achieving accurate and reliable chemical identification.