Journal of Dispersion Science and Technology

ISSN: 0193-2691 (Print) 1532-2351 (Online) Journal homepage: https://www.tandfonline.com/loi/ldis20

Crude Oil Model Emulsion Characterised by means

of Near Infrared Spectroscopy and Multivariate
Techniques

Harald Kallevik , Susanne Brunsgaard Hansen , Øystein Sæther , Olav Martin

Kvalheim & Johan Sjöblom

To cite this article: Harald Kallevik , Susanne Brunsgaard Hansen , Øystein Sæther , Olav Martin
Kvalheim & Johan Sjöblom (2000) Crude Oil Model Emulsion Characterised by means of Near
Infrared Spectroscopy and Multivariate Techniques, Journal of Dispersion Science and Technology,
21:3, 245-262, DOI: 10.1080/01932690008913265

To link to this article: https://doi.org/10.1080/01932690008913265

Published online: 23 Apr 2007.

Submit your article to this journal

Article views: 93

View related articles

Citing articles: 19 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=ldis20
1. DISPERSION SCIENCE AND TECHNOLOGY, 21(3), 245-262 (2000)

Crude Oil Model Emulsion Characterised by means of Near Infrared

Spectroscopy and Multivariate Techniques

Harald Kallevik". Susanne Bmnsgaard an sen'. Oystein ~ ~ t h e r ' ,

Olav Martin Kvalheim' and Johan sjiiblom3

I Department of Chemistry
University of Bergen
Allegaten 4 1
N-5007 Bergen
Norway

Department of Chemistry
Technical University of Denmark
Building 207
DK-2800 Lyngby
Denmark

'Statoil Research Centre

Arkitekt Ebbells v 10
N-7005 Trondheim
Norway

* Comsponding author

Copyight 0 2000 by Marccl Orkkcr. Inc.

246 KALLEVIK ET AL.

Abstract
Water-in-oil emulsions are investigated by means of multivariate analysis of near
infrared (NIR) spectroscopic profiles in the range 1100- 2250 nm. The oil phase is a
paraffin-diluted crude oil from the Norwegian ContinentalShelf. The influence of
water absorption and light scattering of the water droplets are shown to be strong.
Despite the strong influence of the water phase, the NIR technique is still capable of
predicting the composition of the investigated oil phase.

Key words:
near infrared spectroscopy, emulsion characterisation, principal component analysis.
partial least squares

Introduction
Near infrared spectroscopy (NIR) is gaining ground as on-line instrumentation in
process control [I]. NIR provides fast and reliable information about the chemical
properties of samples in terms of overtones and combinations of fundamental
vibration modes. Absorption in the NIR region is much weaker than absorption in the
IR region, and this necessitates use of longer pathlengths. This is often is an
advantage in on-line applications. The broad absorption bands and thereby lack of
selectivity of the NIR spectrum often requires use of multivariate techniques for
analysis and interpretation. In addition to providing the chemical composition, several
authors have recognised NIR as a method for analysis of the sample structure [2-61.
Panicles within the sample scatters light, and this contribution to the light extinction
often give valuable information regarding the state of the samples. In previous work
we have used NIR for characterking the chemical composition and state of crude oil
systems, regarding their ability to form emulsions when mixed with a water phase [7,
8). In this report we aim at exploring NIR as a technique for characterisation of the oil
phase, in the presence of highly scattering water-in-oil emulsion droplets.
CRUDE OIL MODEL EMULSION 247

Table I
Assignments and positions (in nm) for some ovenone and combination bands.

0 - H First Ovenone 1400 - 1450

0 - H Combinations 1900- 1975
C - H Second ovenone 1125 - 1225
C - H Combinations first overtone 1350- 1450
C - H Fint overtone 1625 - 1775
C - H Combinations region 1950 - 2450

Theory

Near Infrared Spectroscopy (NIR)

Near infrared spectroscopy in the 1100 - 2250 nm region provides information about
overtones and combinations of the fundamental vibrations in the infrared region. The
main vibrational absorption bands are listed in Table I. Large aromatic fused-ring
systems possess electronic transitions in the NIR range. For crude oils this absorption
is strong at 1100 nm and gradually reduces towards 1600 nm. The light extinction in
the NIR region is severely influenced by scattering particles. If [he radius of the
particles is c 1 pm (1110 of the wavelength) the light scattering follows the theory of
Rayleigh scattering. Debye and Mie scattering applies for larger particles [9].

Principal Component Analysis (PCA)

Near infrared spectroscopy provides spectra with broad absorption bands and low
selectivity. Direct qualitative and quantitative interpretation of such spectra by
analysis of the raw spectra is often impossible. and this has led to a tight connection
between NIR and multivariate techniques. NIR spectra of n samples with the
absorption recorded at m wavelengths results in a data matrix X with dimension
n x m. All spectroscopic data matrices contain correlated or redundant information
248 KALLEVIK ET AL.

due to the fact that every analyle in the sample absorbs radiation at several
wavelengths. Thus. the absorption at these wavelengths is highly correlated.

The problems with interpretation and the large amount of r&dundant data can be
overcomed by use of the principal component analysis (PCA) [lo, 1I]. In PCA the X
matrix is decomposed into principal components called score and loading vectors. The
scores and loadings contain information on how the samples and the variables in X
relate to each other, respectively. Mathematically, the procedure can be expressed as

In Eq. I . X is the original data matrix, ti are the score vectors and pi are the loading
vectors. E is the residual matrix. The direction of the first principal component (the
vectors 11and p,) is defined as the straight line that best describes the variation within
the X matrix. The direction of the second principal component is the one that best
describes the variation in X not explained by the first component. The third and founh
components and so on can be found in the same manner. This procedure results in k
orthogonal pairs of score and loading vectors. In decomposition of NIR spectral data.
usually more than 95 9% of the original variation in the data can be explained by 2-5
principal components. If thesc components are plotted against each other, the relation
between samples is easily detected.

Partial Least Squares Regression (PLS)

In the description of PCA, the lack of selectivity and the redundancy of NIR spectra
was commented on. This is also a problem for regression analysis. Regression
between the X data and a response y is often impossible by standard methods due to
these unwanted features of the NIR spectra. The solution to this is to decompose the X
matrix as shown in Eq. I and perform the regression between the resulting score
vectors and the response. PCA decomposition followed by regression is called
principal component regression (PCR) [lo. 111. The partial least squares (PIS)
CRUDE OIL MODEL EMULSION 249

technique [lo, 1I] is another and the one used in the present work. ln PCA the scores
and loadings are the vectors that best describes the variance of the X matrix. In PLS
thescores and loadings (called latent variables) are the vectors that have the highest
covariance with the response vector y. The decomposition is followed by a regression
between the latent variables and the response.

The number of latent variables used in the PLS regression is often assessed by the
cross validation technique (1 1. 121. The validation is imponant due to the risk of
overfitting of the experimental data.

Experimental

Chemicals :rl on !

The crude oil was pmduced'bn the ~ d k e ~ i Continental

an Shelf. The asphaltcne
content of the crude is - 4 wt. %. Exhol D-80 is the trade name for a C10 - C13
alkane I cycloalkane mixture (c 1 % aromatics) delivered by Exxon Chemical Noway
AS. The water was distilled before use.

The experimental design

The present experimental systems consists of 3 mixture variables (amount of crude
oil, exxsol and water) and one process variable (energy input in the mixing). The
amount of crude was limited to vary between 1 and 5 % and the amount of water from
5 to 20 46. Obviously, the sum of the mixture variables must be IGU % for each
sample, and the amount of the paraffinic constituent was .determined by this
constraint. A D-optimal design (13-171 with 8 experimental points and 3 centre points
was selected. The design was calculated in the Modde 4.0 (181 software, based on the
selection of a screening experiment with linear response surface. The G-efficiency
[I41 was selected as the criteria for the design evaluation. This procedure resulted in
Table 2.
D-optimal experimental design with 3 mixture variables and I process variable. 4
random samples are included. I - 11 = design variables (9 - I I = centre samples) and
-
12 15 = random samples.

Name Run Mixing (rpm) wt % C ~ d ewt % Disperse wt % b o l

1 7 700 1.OO 5.00 94.00
2 1 700 5.00 5.00 90.00
3 8 700 1.OO 20.00 79.00
4 9 700 5.00 20.00 75.00
5 10 1400 1.OO 5.00 94.00
6 4 1400 5.00 5.00 90.00
7 6 1400 1.OO 20.00 79.00
8 3 1400 5.00 20.00 75.00
9 11 1050 3.00 12.50 84.50
10 5 1050 3.00 12.50 84.50
11 2 1050 3.00 12.50 84.50
12 12 700 1.10 19.55 79.35
13 13 1200 2.31 10.69 87.00
14 14 1370 1.29 6.71 92.00
15 15 1050 4.32 6.62 89.06

Total mass of the samples: 100.00 g

the design shown in Table 2. In addition. we added 4 random samples for validation
of the prediction models.

Preparation of the ~mulslona

The emulsion samples were preparedas illustrated in Figure 1 and the composition
nnd energy input were in accordance with Table 2.
 CRUDE OIL MODEL EMULSION 251

1 2 3 4 5

Addition of: Ultrasonic treatment Addition of Emulsification NIR sampling

1. Crude oil 1 minute water 1 minute 333 rpm mixer
2. Exxsol

Figure I.
Preparation of the emulsions and NIR measurements. The crude oil and exxsol is
mixed and treated with ultrasound, water is added and the sample is emulsified for 1
minute. The sample is then mixed with a rotor mixer at 500 rpm during the NIR
sampling.

Near Infrared Measurements

The emulsion was stirred at 500 rpm with a rotor mixer during the spectral accusation
in order to prevent sedimentation gradients of the water droplets. The near infrared
sampling was performed on a NirSystems 6500 spectrophotometcr, equipped with a
fibre optic sampling probe for transflectance sampling. The wavelength region was set
to 1100-2250 nm. The total path length was 2.4 mm. The total number of scan per
spectra was set to 32 and the sampling was done at25 T.

Data Analysis
The experimental design was constructed using the Modde 4.0 software from Umctri
AB [la]. The data analysis was performed in the Sirius 6.5 chcrnometric s o h a r e
form Pattern Recognition Systems AS [19].
252 KALLEVIK ET AL.

Figure 2.
The NIR spectra of the 15 emulsions: 1100 - 2250 nm.

Results and Discussion

NIR Spectra of the Emulsions

Figure 2 is the NIR spectra of the 15 emulsions. All spectra contain absorption peaks
in accordance with Table I, but the most prominent spectral future is the change in the
specual baseline. This must bc a result of light scattering of the emulsion droplets.
The droplet sizes of the samples were checked by means of a videoenhanced
microscopy technique. The average droplet radius for the samples are found to be in
CRUDE OIL MODEL EMULSION 253

the range from 1.5 to 2.2 pm. This means that the droplet sizes IT above the Rayleigh
limit (r > 1110 of the wavelength). Thus, the extinction of light due to scattering of the
water droplets is expected to be more complex than the monotonic increase with
radius within the Rayligh region. Spectra 9. 10 and I I are replicates and show the
excellent reproducibility of the sample preparation and NIR acquisition. In the
following discussion we focus on the first 8 samples since these are the design
samples.

The effect of increased energy input on the NIR spectra is twofold. The lowcnergy
input samples 1 , 2 , 3 and 4 have correspondingly high-energy input samples in 5 , 6 , 7
and 8, respectively. These 4 pairs of samples show the effect of increasing the energy
input for 4 different chemical compositions. For the low water content pain 1-5 and
2-6 the effect is an increased extinction of the light. For the high water content pain
3-7 and 4-8, the effect is opposite. This must clearly be related to the light scattering
of the water droplets and thus the radius, since the energy input is the only variable
changed. Increasing the amount of crude oil results in increased light extinction for all
pairs (1-2. 3-4. 5-6 and 7-8). The absorption due to electronic transitions of the
asphaltenes in the crude oil is not possible to observe just by looking at the spectm
Increased water phase increases the absorption in the 0-Habsorption regions (see
Table I). In addition the spectral baseline is shifted upwards for all sample pairs (1-3,
2-4,5-7 and 6-8). More water phase gives a larger number of scattering droplets.

PCA of the Spectra

Figure 3 shows the spectra plotted on the first score vector Venus the second. The
scores reveal the relations between samples, as explained in the theory section. Thus.
samples located close to each other in a score plot have common properties (e.g.
absorb light in a similar manner). From the score plot we conclude that the fint
principal component separates the samples according to water content. The first
loading vector is plotted versus wavelength in Figure 4. This plot shows the
importance of each wavelength in order for explaining the variation accounted for in
 KALLEVlK ET AL.

10 % w., phu

-1.00 .
-1.50 -1.00 -0.50 0.00 0.50 1.00
Comp. 1 (99.1%)

Figure 3.
Score plot of principal component I versus principal component 2.

the first principal component. The loadings can be interpreted much like a NIR
spectrum. We observe that the first principal component is explained by the spectral
baseline and by the regions related to absorption by the 0-Hgroups of water (Table
1). Thus, the first principal component describes the variation related to the amount of
light scattering and absorbing water droplets. The variance accounted for by the fint
principal component is as high as 99.1 90of the total.

The second (and higher) score vector is difficult to interpret. In Figure 3 we sec no
clear relation between the positions of the samples and the design variables in Table
2. The comsponding loading vector (Figure 4) is difficult to relate to any spectral
features of the components in the samples. The explanation for these difficulties in
CRUDE OIL MODEL EMULSION

Wavelength

Figure 4.
Loading plot of principal component 1 versus principal component 2.

relating spectral propertics to the design variables is that the scattering effect
dominates the spectrum. and that the scattering result in a non-linear relation betwccn
the design variables and the spectra

PLS Regression on the NIR Spectra

Partial least squam regression is like PCA a linear regression technique. It has,
 KALLEVIK E7 AL.

2.00 2.50 3.00 3.50 4.00

Measured (Mixing)

Figure 5.
PIS prediction model for the energy input (mixing).

-I
1 .OO 2.00 3.00 4.00 5.00
Measured (%Crude)

Figure 6.
PLS prediction model for the amount of crude oil in the emulsion samples. 12. 13. 14
and 15 are random samples fitted to the model (they are not included in the
modelling).
CRUDE OIL MODEL EMULSION

Wavelength (nm)

Figure.7.
The regression coefficients for the PLS model in Figure 6.

however, the capability of handling some non-linearity's by inuoducing more latent

variables. Figure 5 shows the result of regression between the NIR spectra and the
mixing variable of the I I design samples. It is clear that PLS in this case can not
handle the non-linearity introduced by the light scattering. Several non-linear
regression techniques can handle such challenges, but this is not further investigated
in the present work.

In order to determine the composition of the emulsion samples the PLS regression
technique was used to build the prediction model. The I1 design samples with the
composition specified in Table 2 were used ar; calibration samples (sample number I -
11). The 4 random samples were used for validation. Figure 6 is the PLS model for
the amount of crude oil in the samples. The cmss validation technique suggested 6
principal components, and the R~ of the model is 0.976. Figure 7 shows the regression
coefficients for the model and shows that the region between 1100 and 1200 is
 KALLEVIK ET AL.

2.00 -

n
1 1.00 -

0.00 5
' -10'
0.50 1.OO 1SO 2.00
Measured (%Water)

Figure 8.
PLS prediction model for the amount of water phue in the emulsion samples. 12. 13.
14 and 15 are the random samples fitted to the model (they are not included in the
modelling).

Wavelenglh (nm)

Figure 9.
The regression coefficients for the PIS model in Figure 8.
CRUDE OIL MODEL EMULSION

Measured (% Eusol)

Figure 10.
PLS prediction model for the amount of exxsol in lhe emulsion samples. 12, 13. 14
and 15 are the random samples fitted to the model (they are nor included in lhc
modelling).

Wavelength (nm)

Figure 1 1.
The regression coefficients for lhe PLS model in Figure 10.
260 KALLEVIK ET AL.

important for the prediction of this variable. This is explained by the electronic
absorption of large aromatic molecules. Figure 6 also shows the result of fitting the 4
random samples to the prediction model (sample number 12 - 15). The results are
good.

Figure 8 is the PLS prediction model for the amount of water in the emulsion samples.
R' of this model is as high as 0.998. Figure 9 gives the regression coefficients for the
model. The most important region is the first 0-Hovertone region around 14M) nm.
Figure 8 also shows the result of fitting the 4 random samples to the model. The fit to
the regression line is excellent.

The prediction model for the amount of exxsol in the emulsion samples is presented in
Figure 10. The interpretation of the regression coefficients in Figure I I is not straight
forward, but the regmsion model performs well on the 4 random samples as shown in
Figure 10.

Conclusion
In this work we have shown the influence of water emulsion droplets (radius 1.5-2.2
pm) on the near infrared spectra in the region 1100 - 2250 nm. Despite the strong
influence of scattering and water absorption, we have shown that the composition of
the oil phase can be predicted from the NIR spectra.

Acknowledgements
H d d Kallevik would like to acknowledge the Norwegian Research Council (NFR)
for a Ph.D. grant. The Near Infrared Specvophotometcr was also financed by NFR.
Financial support from the industrial participants in the Flucha program is also highly
appreciated.

:hi
CRUDE OIL MODEL EMULSION

References
McClure, W.F., Near-Infrored Spectroscopy - The Giant is Running Strong.
Analytical Chemistry. 1994. 66(1): p. A43A53.
Gonzalez. F. and Pous, R., Quality-Control in Manufacturing Process by
Near-Infrared Spectroscopy. Joumal of Pharmaceutical and Biomedical
Analysis, 1995. 13(4-5): p. 419-423.
Frake, P., Gill, I., Luscombe, C.N., Rudd, D.R., Waterhouse. J. and
Jayasomya, U.A.. Near-infrared mass median particle size determi~tionof
lactose monohydrate, evaluating several chemometric approaches. Analyst.
1998. 123(10): p. 2043-2046.
ONeil. A.J.. Jee, R.D. and Moffat, A.C.. The application of multiple linear
regression to the measurement of the median particle size of drugs and
pharmaceutical excipients by near-infrared spectroscopy. Analyst, 1998.
l B ( 1 I): p. 2297-2302.
Gosscn, P.D.. MacGregor, J.F. and Pelton, R.H., Composition Md Panicle
Diameter for Siyrene-Methyl Methacrylate Copolymer Latex using LN and
NIR Spectroscopy. Applied Spectroscopy, 1993.47(11): p. 1852-1870.
Bull. C.R., Compensation for Particle Size Effects in Near-Infrared
Reflectance. Analyst, 1991. 116(8): p. 781-786.
Kallevik, H., Kvalheim, O.M. and Sjohlom, J.. Quantitarive Determi~tionof
Asphaltme Content in Crude Oils and Prediction of the High Electric Field
Emulsion Stability by means of Near Infrared Spectroscopy and Multivariate
Regression. Submitted to Joumal of Applied Spectroscopy, August 1999.
Kallevik, H., Kvalheim, O.M. and Sjoblom, J.. Quantitative Determination of
Asphhlrenes and Resins in Solution by means of Near Infrared Spectroscopy.
Correlations to Emulsion Stability. Submitted to Joumal of Colloid &
Interface Science, May 1999.
Bohren, C.F., Absorption and scattering of light by small panicles /Craig F.
Bohren, DonaldR. Huffman., ed. D.R. Huffman, New York.: Wiley.
 KALLEVIK ET AL.

Kvalheim, O.M. and Karstang. T.V.. Interpretation of latent-variable

regression models. Chemometrics and Intelligent Laboratory ~ ~ i i e i n 1989.
s.
7: p. 39-51.
Martens, H. and Nes, T., Multivariate calibration, ed. T. NZS. 1989.
Chichester.: Wiley.
Efron, B. and Gong. G..A hisurely look at the Bootstrap, the Jack!aife, and
the cross-validation. The American Statistician, 1983. 37: p. 3748.
Eriksson, L.. lohansson. E. and Wikstrorn. C., Mirrure design-design
generation, P U analysis, and model usage. Chemometrics and Intelligent
Laboratory Systems, 1998.43: p. 1-24.
deAguiar. P.F., Bourguignon, B.. Khots, M.S., Massan. D.L. and
PhanThanLuu. R., D-optimnl designs. Chemometrics and Intelligent
Laboratory Systems. 1995. 30(2): p. 199-210.
Lindgren. A,, Sjostrom, M. and Wold. S., PLS modelling of detergency
performance for some technical nonionic surfactants. Chemomeuics and
Intelligent Laboratory Systems, 1996.32(1): p. 111-124.
Duineveld, C.A.A., Smilde, A.K. and Doornbos, D.A.. Comparison of
Experimental-Designs Combining Process and Mixiure Variables. I. Design
Construction and 'I%eoretical Evaluation. Chemometrics and Intelligent
Laboratory Systems. 1993.19(3): p. 295-308.
Duineveld, C.A.A., Smilde. A.K. and Doornbos, D.A., Comparison of
Experimental-Designs Combining Process and Mixture Variables. 2. Design
Evaluation on Measured Data. Chemometrics and Intelligent Laboratory
Systems, 1993.19(3): p. 309-318.
Modde 4.0. Umetri AB, Box 7960. S-907 19 UMEA. SWEDEN. '

http:/lwww.umeui.sel
Sirius 6.5, Pattern Recognition Systems AIS. Bergen High-Technology Center.
N-5008 BERGEN, NORWAY, ht~p:llwww.prs.no

Received: 8/1/99
Accepted: 8/24/99