Application of infrared spectroscopy and chemometric methods to identification of selected minerals

Acta Geodyn. Geomater., Vol. 8, No. 1 (161), 47–58, 2011 1. APPLICATION OF INFRARED SPECTROSCOPY AND CHEMOMETRIC METHODS TO IDENTIFICATION OF SELECTED MINERALS Michal RITZ 1)*, Lenka VACULÍKOVÁ 2) and Eva PLEVOVÁ 2) 1) VŠB-Technical University Ostrava, 17. listopadu 15, 708 33 Ostrava-Poruba, Czech Republic 2) Institute of Geonics of the AS CR, Studentská 1768, 708 00 Ostrava-Poruba, Czech Republic *Corresponding author‘s e-mail: [email protected] (Received November 2010, accepted March 2011) ABSTRACT A new way of identification of minerals was suggested. The identification was based on chemometric analysis of measured I R spectra of selected minerals. IR spectra were collected using diffuse reflectance technique. The discriminant analysis an d p rincipal component analysis were used as chemometric methods. Five statistical models were created for separation an d identification of clay minerals. Up to 60 samples of various mineral standards (clay minerals, feldspars, carbonates, sulphates and quartz) from different localities were selected for the creation of statistical models. The results of this study confirm tha t the discriminant analysis of IR spectra of minerals could p rovide a powerful tool for mineral identification. Eve n differentiation of muscovite from illite and identification of mixed structures of illite-smectite were achieved. KEYWORDS: minerals, identification, IR spectroscopy, discriminant analysis, principal component analysis Very useful for processing of IR spectra are multivariate statistical methods. The big advantage o f multivariate statistical methods is their capability to extract required spectral information from IR spectr a and explore this spectral information for qualitative o r quantitative applications. The most frequently use d multivariate statistical methods (often called chemometric methods) are factor analysis (FA), p rincipal component analysis (PCA), discriminant analysis (DA), principal component regression (PCR), multiple linear regression (MLR) and partial leastsquares regression (PLS). DA, FA and PCA are very useful tools for qualitative analysis. The discriminant analysis allows quantification of the probability with which an object can be assigned to an individual class. DA generates a discriminant function as a linear combination o f measured variables, which gives maximu m separability for objects from different classes. The ratio of between-class to pooled within-class sample variance is commonly used as a criterion Jfo r measuring the separability of different classes: T T aBa JaWa = (1) where ais a vector maximizing the discriminan t criterion and B and W are the between-class and the within-class covariance matrixes, respectively (Jiang et al., 2002). 1. INTRODUCTION Knowledge of mineral composition is essential to characterize the geochemical and physicomechanical properties of rocks. Nature and content o f minerals (especially clay minerals) present in rocks have a significant influence on the behaviour an d p roperties of rocks as well as on the whole roc k massif. In general, several conventional analytical methods exist that can be used to examine the mineral composition of rocks: optical microscopy, electron microscopy, X-ray diffraction (XRD), Fourie r transform infrared spectroscopy (FTIR spectroscopy), Raman spectroscopy, thermal analysis (TG/DTA) and b ulk chemistry analysis (Kodama et a., 1989; Chipera and Bish, 2001; Srodon, 2002; Vogt et al., 2002; Vaculíková, 2006). Unfortunately, the exac t identification of clay minerals by means of these methods is rather complicated and often inaccurate. The main analytical difficulties are related to variable chemical composition and common structural disorders of clay minerals. Another reason consists in the fact that the individual clay minerals occur in the form of mixtures (illite – montmorillonite, chlorite – montmorillonite, etc.) with various ratios of particula r clay minerals. The current FTIR spectroscopy makes it possible to analyze individual minerals, noncrystalline admixtures and, simultaneously, to detect the presence of organic matter. DSpace VŠB-TUO http://hdl.handle.net/10084/84629 21/06/2012 M. Ritz et al. 48 b ands represent spectral information positively correlated through particular principal component; the negative ones represent spectral information negatively correlated through particular principal component. Both of the above-mentioned chemometric methods (DA and PCA) are described in a lot o f statistical textbooks (Meloun and Militký, 2004; Hendl, 2006) in full details. Likewise, DA and/o r PCA of IR spectra are used in many studies (Tesch and Otto, 1995; Dahlberg et al., 1997; Kher et al., 2001; Tsuchikawa et al., 2003; Ritz, 2007). This paper describes a new approach to clay minerals and some accessories minerals identification b ased on chemometric analysis of IR spectra of clay minerals. All IR spectra were collected using diffuse reflectance technique (DRIFT). The aim of the stud y was the creation of statistical models for minerals identification by discriminant analysis of measured I R spectra. 2. MATERIALS AND METHODS 2.1. SAMPLES OF MINERALS The research has been focused on the basic clay mineral specimens (kaolinite, montmorillonite, nontronite, illite, chlorite and vermiculite) including micas (muscovite, biotite) and the associated minerals (quartz, feldspar, carbonates and sulphates). The standards of above-mentioned minerals were selected according to their authenticity and a degree of purity. These mineral standards were obtained from the collection of minerals available at the Institute o f Geonics, Nanotechnology Centre of VŠB-Technical University Ostrava and from Source Clays Repository of the Clay Minerals Society, USA. 60 samples of mineral standards were collected for the preparation of statistical models. 51 samples o f clay minerals (including micas) and 9 samples o f associated minerals (quartz, feldspars, carbonates an d sulphates) were used. Some of mineral standards were treated by jet mill; other ones were prepared by sedimentation to obtain fraction grain size less than 5 μm. Some standards were used as unmodified (“raw” standards). A list of all used samples o f mineral standards is shown in Table 1. 2.2. FTIR MEASUREMENTS Approximately 5-10 mg of sample of mineral standards was ground with approx. 400 mg dried KBr. This mixture was used to collect IR spectra. The IR spectra were collected using FTI R spectrometer Nexus 470 (ThermoScientific, USA). Diffuse reflectance measurement technique (DRIFT) was used. This technique was selected for its fastness and simplicity. The measurement parameters were as follows: spectral region 4000-400 cm-1, spectral resolution 8 cm-1; 128 scans; Happ-Genzel apodization. The discriminant analysis uses results of anothe r chemometric method – principal component analysis (PCA). PCA is one of the most common multivariate methods wide-spread also in infrared spectroscop y (e.g. Tesch and Otto, 1995; Dahlberg et al., 1997; Rusnak et al., 2003). IR spectrum is composed o f many points connected by the line; each point in the IR spectrum is defined by wavenumber and intensity (for example absorbance). In PCA, each spectrum is represented by a point in multidimensional space, whereas the number of dimensions corresponds to the number of points in the spectrum (in other words the number of variables). Because this multispace is very difficult to comprehend, PCA introduces a new coordinate system as a linear combination of the ol d coordinates. Most of the variances (it means the mos t of information from the IR spectrum) are concentrated in only a few of the new coordinates – called principa l components. This can be expressed as the decomposition of the data matrix (D) into two othe r matrixes - the score matrix (P) and the loading matrix (T). For better understanding of the decomposition o f the data matrix we can use the following graphical representation (Geladi and Kowalski, 1986): D matrix has the dimensions of number of samples (n) b y intensity (at particular wavenumbers) of IR spectra (m). The score matrix (P) has the dimensions o f number of samples (n) by principal components (a) and this matrix represents the location of the samples in the new principal component space. The loading matrix (T) has dimensions of intensity (m) by principal components (a); loading matrix represents the recipes by which each of the principal components was constructed from the old variable - intensity a t each wavenumber. The most important information is included in the first two or three principal components. The rest of the principal components contain only minor characteristics of samples an d noise, so that can be ignored. The main advantage o f PCA consists in reducing the influence of noise and exploiting the subtle differences in the IR spectra o f samples to obtain relevant and significant spectral information. In PCA of spectral data, the loading matrix is very important. This matrix shows us which spectral information (regions or bands) are important in the p rincipal component space. The most significant wa y of projection of loading matrix in spectral analysis is so-called loading spectra. They often look like “typical spectra” and positive and/or negative ban d can be present at the loading spectra. The positive DSpace VŠB-TUO http://hdl.handle.net/10084/84629 21/06/2012 APPLICATION OF INFRARED SPECTROSCOPY AND CHEMOMETRIC METHODS … 49 Table 1 List of samples. Sample Mineral Locality Description Bio1 Biotite Ktíš, Czech Republic fraction ≤ 5 μm Bio2 Biotite Ktíš, Czech Republic “raw” standard Bio3 Biotite Dolní Bory, Czech Republic fraction ≤ 5 μm Mus1 Muscovite India “raw” standard Mus2 Muscovite Měděnec, Czech Republic fraction ≤ 5 μm Mus3 Muscovite Otov, Czech Republic fraction ≤ 5 μm Mus4 Muscovite Měděnec, Czech Republic “raw” standard Mus5 Muscovite Rožná, Czech Republic fraction ≤ 5 μm Mus6 Muscovite-Chlorite Měděnec, Czech Republic “raw” standard Ill1 Illite Hungary fraction ≤ 5 μm Ill2 Illite (IMt-1) Montana, USA fraction ≤ 5 μm Ill3 Illite Karlov, Czech Republic “raw” standard Ill4 Illite (IMt-2) Montana, USA “raw” standard Ill5 Illite (IMt-1) Montana, USA jet mill Ill6 Illite (IMt-2) Montana, USA “raw” standard Ill7 Illite-Smectite (ISCz-1) Czechoslovakia “raw” standard Ch1 Chlorite (CCa-2) California, USA fraction ≤ 5 μm Ch2 Chlorite Letovice, Czech Republic fraction ≤ 5 μm Ch3 Chlorite Orlické hory, Czech Rep. fraction ≤ 5 μm Ch4 Chlorite (CCa-2) California, USA “raw” standard Ch5 Chlorite (CCa-2) California, USA jet mill Kao1 Kaolinite Božíčany, Czech Republic “raw” standard Kao2 Kaolinite Horní Bříza, Czech Republic “raw” standard Kao3 Kaolinite (KGa-1b) Georgia, USA fraction ≤ 5 μm Kao4 Kaolinite Kadaň, Czech Republic “raw” standard Kao5 Kaolinite Kaznějov, Czech Republic “raw” standard Kao6 Kaolinite Sedlec, Czech Republic “raw” standard Kao7 Kaolinite Únavov, Czech Republic “raw” standard Kao8 Kaolinite (KGa-1b) Georgia, USA “raw” standard Kao9 Kaolinite (KGa-2) Georgia, USA “raw” standard Kao10 Kaolinite Sedlec, Czech Republic “raw” standard Kao11 Kaolinite Sedlec, Czech Republic fraction ≤ 5 μm Dic1 Dickite Krásno, Czech Republic fraction ≤ 5 μm Hal1 Halloysite Bílá Hora, Slovakia “raw” standard Mon1 Montmorillonite Ivančice, Czech Republic fraction ≤ 5 μm Mon2 Montmorillonite Jelšový potok, Slovakia fraction ≤ 5 μm Mon3 Montmorillonite (SWy-1) Wyoming, USA “raw” standard Mon4 Montmorillonite (SWy-1) Wyoming, USA fraction ≤ 5 μm Mon5 Montmorillonite (SWy-2) Wyoming, USA “raw” standard Mon6 Montmorillonite (SAz-2) Arizona, USA fraction ≤ 5 μm Mon7 Montmorillonite (SAz-2) Arizona, USA “raw” standard Mon8 Montmorillonite (STx-1b) Texas, USA “raw” standard Non1 Nontronite (Nau-1) Uley Mine, South Australia “raw” standard Non2 Nontronite (Nau-2) Uley Mine, South Australia “raw” standard Hec1 Hectorite (SHCa-1) California, USA “raw” standard Ver1 Vermiculite China expanded Ver2 Vermiculite Aldrich standard expanded Ver3 Vermiculite Aldrich standard expanded, jet mill Ver4 Vermiculite Letovice, Czech Republic not expanded Ver5 Vermiculite Brasil not expanded Ver6 Vermiculite Brasil not expanded, jet mill Acc1 Quartz Otov, Czech Republic “raw” standard Acc2 Siderite Dúbrava, Slovakia “raw” standard Acc3 Magnesite Nižná Slaná, Slovakia “raw” standard Acc4 Calcite Štramberk, Czech Republic “raw” standard Acc5 Gypsum Nothern Bohemia, Czech Rep. “raw” standard Fel1 Albite West. Bohemia, Czech Rep. “raw” standard Fel2 Feldspar (Na, K) West. Bohemia, Czech Rep. “raw” standard Fel3 Feldspar (Na, K) West. Bohemia, Czech Rep. “raw” standard Fel4 Feldspar (Na, Ca) West. Bohemia, Czech Rep. “raw” standard DSpace VŠB-TUO http://hdl.handle.net/10084/84629 21/06/2012 M. Ritz et al. 50 Model 1 was created for the separation of clay minerals from accessory minerals like feldspars, quartz, carbonates (calcite, siderite and magnesite) and sulphates (gypsum). General discrimination o f clay minerals was performed in Model 2. The clay minerals were classified into particular groups of cla y minerals in this model: biotite, muscovite-illite, chlorites, kaolin-serpentines, smectites an d vermiculites. The remaining models were created fo r identification of individual members of selected clay mineral groups. Model 3 has allowed identification o f minerals of the kaolin-serpentine group (kaolinite, dickite and halloysite), Model 4 has allowe d identification of minerals of the smectite group (Namontmorillonite, Ca-montmorillonite, nontronite and hectorite) and Model 5 has allowed identification o f micas (biotite, muscovite) and illite. 3.2. MODEL 1 – SEPARATION OF CLAY MINERALS FROM ACCESSORY MINERALS This model was created for the selection of clay minerals from accessory minerals. There were three classes used in Model 1 (Clay minerals, Feldspars and Accessory minerals). The first class included all used samples of clay minerals (biotite, muscovite, illite, chlorite, kaolinite, dickite, halloysite, montmorillonite, notronite, hectorite and vermiculite). The first class (Clay minerals) involved 51 samples as members of this class. The second class (Feldspars) had 4 members; three samples of mixed feldspars an d one sample of albite. The third class (Accessor y minerals) included 5 members (quartz, siderite, calcite, magnesite and gypsum). The Mahalanobis distance plot is shown in Figure 1. All minerals were clearly distributed into three clusters whereas the cluster of Feldspars is relatively close to the cluster o f Clay minerals. The cluster of Accessory minerals is evidently isolated from the other two clusters. The principal component analysis was used as integral part of the discriminant analysis. Ten p rincipal components (PCs) were calculated in Model 1. These ten PCs described 99.3 % of variance of spectral information, whereas almost 95 % o f variance was involved in the first four PCs. Loadings spectrum of the first principal component (PC1) includes spectral bands of kaolinite only: stretching bands of inner-surface hydroxyl groups (3695 cm-1, 3668 cm-1, 3652 cm-1 and 3620 cm-1), Si-O stretching band (1100 cm-1), in-plane Si-O stretching bands (1030 cm-1 and 1010 cm-1 ), deformation band o f inner-surface hydroxyl groups (915 cm-1), Si-O stretching bands (795 cm-1, 755 cm-1 and 695 cm-1) and deformation bands Al-O-Si, Si-O-Si and Si-O, respectively (540 cm-1, 470 cm-1 and 430 cm-1). Assignment of IR spectral bands of minerals according to literature (Russel et al., 1994; Madejov á and Komandel, 2001; Vaculíková and Plevová, 2005) was used in this paper. PC1 described 76.3 % o f variance. The second principal component (PC2) Every sample was prepared and consequently measured several times (3 – 5 times). The mean I R spectrum of every sample was calculated fo r subsequent statistical processing. 2.3. DISCRIMINANT ANALYSIS The discriminant analysis (DA) was performed using TQ Analyst software (ThermoScientific, USA). Two spectral regions 4000-3000 cm-1 and 1300-400 cm-1 of each IR spectrum were used fo r DA. A onep oint baseline was used in every spectral region. A multiplicative signal correction o f p athlength was used for calculation. The Mahalanobis distance was used to formulate a distance between clusters. Before DA, p rincipal component analysis (PCA) carried out. The task of PCA was reduction of redundant spectral information and finding o f important spectral features which had a significan t influence on spectral variance. The number o f principal components used for the preparation o f model was 10. Validation of statistical models has been carried out by means of a validation spectra set (TQ Analys t software does not allow cross-validation fo r discriminant analysis). The IR spectra of minerals were split into two groups in each class used in DA – calibration group of spectra and validation group o f spectra. The group of calibration spectra was used to create a discrimination model; the group of validation spectra was used to verify this model. Most of the I R spectra in each class were used as calibration spectra; up to three IR spectra in each class were used as validation spectra. There were no validation spectra used in one-member classes only. The selection o f calibration and validation spectra has been performed b y TQ Analyst software. The following samples were used for collection of validation spectra: Bio2, Mus1, Mus4, Ill2, Ill4, Ch3, Kao2, Kao7, Kao8, Mon3, Mon6, Non2, Ver4, Fel3 and Acc3. The models in this study were considered correct when all validation spectra occurred inside their class clusters. All discrimination models mentioned in this paper fulfill this condition. 3. RESULTS AND DISCUSSION 3.1. STRATEGY OF MODEL CREATION The creation of a unique statistical model for all used samples was a primary idea of the authors. During the process of creation of the unique model serious problems occurred. A lot of minerals were classified into incorrect classes because o f considerable variability of spectral bands. The solution to this problem was a creation of partial models; five partial models were created in this study. These partial models have allowed separation an d subsequent identification of clay minerals in a few steps. A brief summary of the created models is in Table 2. DSpace VŠB-TUO http://hdl.handle.net/10084/84629 21/06/2012 APPLICATION OF INFRARED SPECTROSCOPY AND CHEMOMETRIC METHODS … 51 Table 2 Brief summary of statistical models. Name of model Used minerals Classes Model 1 All mineral samples Clay minerals Feldspars Accessory minerals Model 2 Clay minerals Biotite Muscovite-illite Chlorite Kaoline-serpentine Smectite Vermiculite Model 3 Minerals of kaoline-serpetine group Kaolinite Dickite Halloysite Model 4 Minerals of smectite group Na-Montmorillonite Ca-Montmorillonite Nontronite Hectorite Mixed illite-smectite structure Model 5 Micas Biotite Muscovite Illite Mixed illite-smectite structure Fig. 1 Mahalanobis distance plot of three clusters of minerals (Model 1). the hydroxyl group of gypsum (3550 cm-1 and 3400 cm-1). The bands of quartz and gypsum had normal orientation (e.g. they did not have negative intensities of bands). The fourth principal component (PC4) described 2.8 % of variance and PC4 loadings spectrum included the same bands as PC 3 loadings spectrum but entire bands had negative intensities. The loadings spectra of the above-mentioned PCs are shown in Figure 2. described 9.3 % of variance and its loadings spectru m included again the kaolinite bands only. Contrary to the PC1, the entire spectral band in PC2 loadings spectrum had negative intensities of bands. The thir d p rincipal component (PC3) described 5.9 % o f variance. The PC3 loadings spectrum included also negative bands of kaolinite and moreover a doublet of Si-O stretching bands of quartz (800 cm-1 and 780 cm-1) and a doublet of stretching bands o f DSpace VŠB-TUO http://hdl.handle.net/10084/84629 21/06/2012 M. Ritz et al. 52 Fig. 2 Loadings spectra of the first four principal components (Model 1). Ten principal components (PCs) were calculated in Model 2. These ten PCs described 99.9 % o f variance of spectral information, whereas almost 97 % of variance was involved in the first four PCs. PC1 loadings spectrum included spectral bands of kaolinite only as well as PC1 in Model 1. PC1 described 79.4 % of variance. PC2 described 10.8 % of variance an d its loadings spectrum included negative bands o f kaolinite and “positive” broad and very weak band a t 3400 cm-1 (stretching vibration of water). PC3 described 4.0 % of variance and its loadings spectru m contained negative bands of kaolinite and “positive” b ands of stretching vibration of water (higher intensity than in PC2). PC4 described 2.3 % of variance an d PC4 loadings spectrum included also negative bands of kaolinite and “positive” band of water, whereas the intensity of the last-mentioned band was higher than in PC3 loadings spectrum. The band at 1645 cm-1 (in PC2, PC3 and PC4 loadings spectrum) belongs to deformation vibration of water. This band very probably presents moisture of samples. The band at 1645 cm-1 has not been used for the creation of Model 2. The loadings spectra of the above-mentioned PCs are shown in Figure 4. 3.3. MODEL 2 – IDENTIFICATION OF CLAY MINERALS GROUPS This model was created for classification of clay minerals into particular groups of clay minerals. Classification of clay minerals according to Weiss and Kužvart (Weiss and Kužvart, 2005) was used in this study. Model 2 involved six classes (Biotite, Muscovite-illite, Chlorite, Kaoline-serpentine, Smectite and Vermiculite). In this model the first two classes represented the micas group of clay minerals; the other classes represented the remaining groups o f clay minerals. The micas group had to be split into two classes (Biotite and Muscovite-illite) by reason o f b ad creation of clusters. In the Biotite class were samples of biotite only; in the Muscovite-illite class were samples of muscovite and illite. In the Chlorite class were chlorite samples only. In the Kaolineserpentine class were samples of kaolinite, dickite and halloysite. In the Smectite class were samples o f montmorillonite, nontronite and hectorite. In the Vermiculite class were samples of vermiculite only. With exception of sample Ill7 all minerals were clearly classified into appropriate classes afte r performance of discriminant analysis. The abovementioned sample (mixed illite-smectite structure) was classified between the class Muscovite-illite and class Smectite (Fig. 3). DSpace VŠB-TUO http://hdl.handle.net/10084/84629 21/06/2012 APPLICATION OF INFRARED SPECTROSCOPY AND CHEMOMETRIC METHODS … 53 Fig. 3 Mahalanobis distance plot of classes Muscovite-illite and Smectite (Model 2). Fig. 4 Loadings spectra of the first four principal components (Model 2). Halloysite). All minerals were clearly distributed into used classes by discriminant analysis. Ten principal components (PCs) were calculated in Model 3. These ten PCs described 99.9 % o f variance of spectral information, whereas almost 97.5 % of variance was involved in the first three PCs. PC1 described 82.5 % of variance, PC2 describe d 11.3 % of variance and PC3 described 3.6 % o f 3.4. MODEL 3 – IDENTIFICATION OF MINERALS OF KAOLINE-SERPENTINE GROUP This model was created for identification o f individual members of the Kaoline-serpentine group of clay minerals. Kaolinite, dickite and halloysite were used in this study as representatives of the Kaolineserpentine group of clay minerals. Thus the Model 3 included three classes (Kaolinite, Dickite an d DSpace VŠB-TUO http://hdl.handle.net/10084/84629 21/06/2012 M. Ritz et al. 54 Fig. 5 Loadings spectra of the first three principal components (Model 3). Fig. 6 Loadings spectra of the first three principal components (Model 4). DSpace VŠB-TUO http://hdl.handle.net/10084/84629 21/06/2012 APPLICATION OF INFRARED SPECTROSCOPY AND CHEMOMETRIC METHODS … 55 vibration (840 cm-1), Si-O stretching band of silica (800 cm-1), coupled out-of-plane vibration band of Al-O and Si-O (625 cm-1) and deformation bands Al-O-Si, Si-O-Si, respectively (525 cm-1 and 470 cm-1). PC1 described 86.0 % of variance. PC2 described 7.2 % of variance and its loadings spectrum included negative bands at 3640 cm-1, 3440 cm-1 and 1650 cm-1 and positive bands a t 800 cm-1, 625 cm-1, 525 cm-1 and 470 cm-1. PC3 described 4.3 % of variance and the loadings spectru m contained negative bands at 3640 cm-1, 3440 cm-1, 1650 cm-1, 525 cm-1 and 470 cm-1. All bands in the loadings spectra of PC2 and PC3 belonged to montmorillonite too. The band at 1650 cm-1 has no t b een used for the creation of Model 4 by reason o f p otential interference intensity of this band by samples moisture. The loadings spectra of the above-mentioned PCs are shown in Figure 6. 3.6. MODEL 5 – IDENTIFICATION OF MINERALS IN GROUP OF MICAS This model was created for identification o f individual members of micas. Biotite, muscovite an d illite were used in this study as representatives of this group of clay minerals. Model 5 included four classes (Biotite, Muscovite, Illite and Mixed illite-smectite structure). All minerals were clearly distributed into used classes by discriminant analysis. The Mahalanobis distance plot of Muscovite and Illite classes is shown in Figure7. The clusters of muscovite and illite were relatively close to each other but they were clearly separated. IR spectra of muscovites and illites can be very similar (on account of similar chemical composition and structure) and common identification variance. All three PC loadings spectra included negative bands of kaolinite. The particular loadings spectra differed only in intensities of bands. Bands o f the highest intensities were in PC1 loadings spectrum. The band at 1440 cm-1 (negative intensity in PC2 loadings spectrum and positive intensity in ureC3 loadings spectrum) belongs to stretching vibration o f carbonates present as impurity in the sample o f dickite. This band has not been used for the creation of Model 3. The loadings spectra of the abovementioned PCs are shown in Figure 5. 3.5. MODEL 4 – IDENTIFICATION OF MINERALS OF SMECTITE GROUP This model was created for identification o f individual members of the smectite group of cla y minerals. Na-montmorillonite, Ca-montmorillonite, nontronite and hectorite were used in this study as representatives of this group of clay minerals. Model 4 included five classes: Na-montmorillonite (Mon1 - Mon5), Ca-montmorillonite (Mon6 - Mon8), N ontronite, Hectorite and Mixed illite-smectite structure. All minerals were clearly distributed into used classes by discriminant analysis. Ten principal components (PCs) were calculated in Model 4. These ten PCs described 99.9 % o f variance of spectral information, whereas almost 97.5 % of variance was involved in the first three PCs. The loadings spectrum of the first principal component (PC1) included spectral bands of montmorillonite only: stretching bands o f structural hydroxyl groups (3640 cm-1), broad stretching band of water (3440 cm-1), deformation band of water (1650 cm-1), band of Si-O stretching (1040 cm-1), deformation bands of Al-Al-OH vibration (915 cm-1), deformation bands of Al-Mg-OH Fig. 7 Mahalanobis distance plot of Muscovite and Illite classes (Model 5). DSpace VŠB-TUO http://hdl.handle.net/10084/84629 21/06/2012