Absorbance spectra from near near infrared (NNIR; 800- 1050nm) spectroscopy of organic liquids are measured using a silicon diode array spectrometer. No sample dilution is necessary as the absorbance values are low. Multi-component analysis applied to complete NNIR spectra provides both quantitive and qualitive information, essential in industrial quality control. Acceptable accuracy (error < 1%) is achieved in the quantitative analysis of mixtures of non-interacting substances, and it is also possible to analyse alcohols in non-polar solvents, where intermolecular H-bonding occurs. Principal component regression was also performed on the NNIR data to give qualitative information.

Keywords:
Near near infrared spectroscopy; multi-component analysis; quality control; principal component regression; diode array spectrometer

Introduction
Spectrophotometry is widely used in industrial quality control. Quantitative analysis is carried out in the ultraviolet/visible (UV/VIS) region (190-820nm) on the basis that the Beer-Lambert law is obeyed, assuming intermolecular interactions are negligible. Concentrations are calculated from absorbance data at only one or two wavelengths. As absorbance values are high in this region, it is necessary to dilute samples accurately prior to measurement and analysis.

Lately, infrared (IR; 2500-15000nm) and near infrared (NIR; 1000-2500nm) spectroscopy have more frequently been used to obtain qualitative information. In the past, this has been achieved by visual comparison of the sample spectrum with a reference. A great improvement on this method was the use of factor analysis of NIR spectra, [1]. A ‘learning set’ of several reference spectra is required, against which the sample spectrum is compared. The results of the analysis are plotted in factor space, the spectra of each different substance forming a separate cluster of points. The sample can be identified by its relative position in the factor space.

Multicomponent analysis (MCA) is a simple least-squares regression algorithm, [ 2, 3]. It uses the data at all the wavelengths considered, if so desired. However, for accuracy, it is necessary that the Beer-Lambert law is obeyed at every wavelength taken. The method is currently being used in the UV/VIS region in this laboratory with increasing success. A standards list consisting of one spectrum of each of the relevant pure components is compiled by the user. The MCA then compares the test spectrum with the standards and constructs the least-squares fit from linear combination. The proportions required of each spectrum are converted into component concentrations. A chi-squared ‘error of fit’, normalized with respect to the system noise and sensitivity, is also returned. This is a measure of how closely the reconstructed spectrum fits the test data. It is a sensitive qualitative measure; a poor fit can indicate that the standards chosen are not appropriate to the sample composition, or that some intermolecular interaction occurs that influences the sample spectrum. The standards list can then be revised, or the wavelength range considered adjusted to improve the fit.

Sample preparation is not necessary for NIR measurements but quantitative analysis is complicated, as the spectra contain overlapping bands and are strongly influenced by intermolecular interactions. Multicomponent analysis using reference spectra is not viable^[*], and more complex algorithms such as principal component regression (PCR) and partial least-squares using factors rather than reference spectra, [4], are required. These involve tedious calibration procedures.

Spectra from near near infrared (NNIR; 800-1050 nm) spectroscopy of many organic liquids exhibit sharper bands with less overlap than the NIR spectra. The NNIR spectra are conveniently measured with silicon diode arrays, no sample dilution and almost no sample preparation is required as the absorbances are low. The aims of this investigation were:
    (i)  to establish whether the simple MCA algorithm is applicable despite the necessarily high sample concentrations; and
    (ii) to determinate whether NNIR spectra are sufficiently from another.

The NNIR absorbance spectra were measured using an HP 8452a extended visible diode array spectrometer (Hewlett-Packard). It was equipped with a multi-cell transporter holding two quartz cells of 10.0 mm optical pathlength.

To counteract instrumental drift, it proved necessary to simulate a double beam apparatus. One cell (the ‘blank’) was filled with solvent (CCl₄). The other was a flow through cell connected to an AMICA 5000 dynamic dilutor, [5], from which the sample, accurately diluted as required, was delivered. The dilutor was driven by the system software. The spectrum of the blank was always measured directly after the sample spectrum, and then subtracted from it mathematically.

Measurements are also sensitive to the effects of turbulence following the injection of the liquid into the cell. Turbulence causes fluctuations in the refractive index of the sample, which then appears as noise in the absorbance spectrum. This was overcome by a stabilization time of 180s prior to the measurement.

The spectra of pure organic liquids and of concentrations down to 40 % v/v in CCl₄ were recorded. Propan-1-ol was measured down to 0.05% concentration. Some two-component mixtures were also investigated over the range from 95 to 5% v/v.

Multicomponent analysis was performed within the instrument software, which was written in this laboratory. The standards list in each case consisted of spectra of the relevant pure substances and included a flat ‘dummy’ spectrum to account for any flat drift in the baseline of the instrument between measurements.

For treatment by PCR the data were transferred to the ICAP software package (Bran and Luebbe). The spectra were usually differentiated before analysis to eliminate the effects of any baseline offset between them (the substitute for the flat dummy in the MCA software package).

Spectral features in the NNIR are sharper that in the NIR. Pure organic liquids show absorbances (0.02-0.07 using a 10.0 mm light path) with a characteristic peak around 900 nm for aliphatic compounds. This is shifted to shorter wavelengths for aromatic compounds. Some typical spectra are shown in Fig. 1. The spectrum of water, by contrast, has a peak at 980 nm with an absorbance of 0.2.

Quantitative Analysis

Dilutions in CCl₄

The observance of the Beer-Lambert law in the NNIR by many organic liquids was confirmed. This was carried out by checking the correlation of the absorbance at a given wavelength with the concentration. The wavelength chosen was at a maximum in the spectrum, in order to minimize the influence of noise. (For alcohols, it was necessary to ignore the peak at about 970 nm, see below.) The standard error of estimate of the linear regression in terms of absorbance was S_y.x <0.0001 in most cases. Therefore, the Beer-Lambert law is closely followed.

Multicomponent analysis was carried out over the whole NNIR range on the spectra recorded. For toluene, benzene and heptane of concentrations from 100 to 40% in CCl₄ a calculation error (residual concentration) of better than 0.5% relative was achieved. Poorer accuracy was obtained with some other substances, e.g., cyclohexane and 1,4-dioxane, which have two possible molecular configurations. The proportions present of each configuration depend on the temperature, which was not necessary constant during the measurements. Problems with electronic interference have so far frustrated attempts at temperature stabilization of the samples. Improved results are expected upon implementation of thermostatic control. It will then be interesting to see how well MCA in the NNIR can differentiate between molecular configurations.

In studying alcohols diluted in CCl₄ an absorbance peak that did not obey the Beer-Lambert law was found. It occurs between 960 and 980 nm, depending on the alcohol (see Fig. 2). It is assigned to an overtone of the OH-stretching vibration, and its amplitude versus concentration characteristic is influenced by the degree of intermolecular hydrogen bonding, [6, 7].

This self-association of alcohols in non-polar solvents precludes the application of MCA over the entire NNIR wavelength range. The reliability of the analysis can be improved in two ways:
      (i)  by restricting the wavelength range so that the region of the OH peak is excluded; or
      (ii) by including in the MCA standards list the spectrum of a very diluted sample (e.g., 1%). Such a spectrum is dominated by the OH peak. The apparent concentration of this standard together with the calculated concentration of the pure standard gives a revised estimate of the alcohol concentration.

The relative accuracies of the above methods are illustrated in Fig.3 for dilutions of propan-1-ol in CCl₄. As the alcohol concentration decreases, the OH peak becomes more significant and corrections are more important. (In the following figures the true concentration is that prepared with the dynamic dilutor.)

Two-component mixtures

Some mixtures of two absorbing pure liquids were also investigated. The concentration of one component relative to the other ranged from 5 to 95% v/v. The MCA could be performed with an error of far better than 1%, if the components did not interact (e.g., by H-bounding). The result of MCA on toluene-heptane mixtures are given in Fig. 4.

Qualitative Analysis

Multicomponent analysis algorithm

The software used for the MCA has the advantage that quantitative and qualitative information are delivered simultaneously. The concentration is calculated, and the normalized error of fit can be used as a qualitative indicator as it measures how closely the test spectrum can be matched by a linear combination of the standard spectra (see above).

Step can be taken to improve a bad fit (e.g., by restricting the wavelength range considered, as with the alcohols, see under Dilution in CCl₄), but it is necessary to decide the extent to which this is reasonable. A persistently poor fit indicates that the sample is incompatible with the standards selected.

Principal component regression

The NNIR spectra are also suitable for analysis by PCR. The spectra of two-component mixtures are separated into well defined clusters in factor space, as shown in Fig. 5. Here, the clusters have degenerated into lines as the data cover wide concentration ranges. The end of a line corresponds to 95% concentration of one of the components. Note the near-coincidence of two of the lines at one end: this occurs where the mixtures both contain 95% propan-1-ol.

It is also possible to construct a quantitative model within the PCR software. This requires more work than with the MCA in order to build a set of calibration spectra. The spectra of 0.05-100% propan-1-ol in CCl₄ were examined in this way. The process yielded two factors: factor one with an eigenvalue of 0.2829 and factor two 0.0024. The former correlates well with the concentration (r = 0.999, relative standard deviation = 0.7%). Factor 2 improves the concentration prediction for very weak solutions (relative standard deviation = 0.6%) where the OH peak is dominant in the spectrum (see under Dilutions in CCl₄). It describes the evolution of intermolecular interactions with the alcohol concentration.

Conclusions

Spectrometry in the NNIR region (800-1050 nm) is suitable for routine industrial quality  control. The NNIR spectra are readily measured using silicon diode arrays, and time-consuming sample preparation is eliminated as the absorbance values are low.

Unlike NIR spectra, NNIR spectra of organic liquids contain distinct bands. The simple, rapid MCA algorithm is applicable and delivers quantitative and qualitative information simultaneously. For liquids with no intermolecular interactions, this method can be applied directly to the entire NNIR spectrum. It is also possible to study substances such as alcohols in non-polar solvents if the interactions are taken into account.

The NNIR spectra are also suitable for qualitative analysis by PCR. Distinct clusters are formed in factor space by the spectra of the different substances investigated. Construction of a quantitative model with this algorithm is also possible, but requires a calibration set of many spectra. Similarly, more complex algorithms are also  available.

It is hoped to improve the present system further by stabilizing the sample temperature. The use of cells with longer optical pathlengths should improve the signal-to-noise ratio, although it might be necessary to extend the stabilization time to allow for the greater turbulence.

[1] Gemperline, P.J., and Weber, L. D., Anal. Chem., 1989, 61, 138.

[2] Brown, C. W., Lynch, P. F., Obremski, R. J., and Lavery, D. S., Anal. Chem., 1982, 54, 1472.

[3] Perkampus, H. H., UV-VIS-Spektroskopie und ihre Anwendungen, Springer-Verlag, Berlin, 1986.

[4] Fuller, M. P., Ritter, G. L., and Draper, C. S., Appl. Spectros., 1988, 42, 217.

[5] Valser, P. E., and Bartels, H., Am. Lab., 1982, 12(2), 32.

[6] Shinomiya, K., and Shinomiya, T., Bull. Chem. Soc. Jpn., 1990, 63, 1093.

[7] Kunst, M., van Duijn, D., and Bordewijk, P., Ber. Bunsenges. Phys. Chem., 1979, 83, 840.

[*] "MCA using  .......  is not viable!":
     As you can see, the whole remain of the paper is an absolutely evidence against this item.
     What have I told you?:  “Too many cooks spoil the soup”!

[1] Infralytic GmbH, Dr. H. Freitag: www.infralytic.de       NNIR-References: www.infralytic.de/g_referenz_en.htm