1674 Energy & Fuels 2007, 21, 1674-1680

Partial Least-Squares (PLS) Correlation between Refined Product

Yields and Physicochemical Properties with the 1H Nuclear
Magnetic Resonance (NMR) Spectra of Colombian Crude Oils
Daniel Molina V.,*, Uriel Navarro Uribe, and Juan Murgich
Escuela de Qumica, UniVersidad Industrial de Santander, Apartado Ae reo 678, Bucaramanga,
Colombia, Ecopetrol-Instituto Colombiano del Petro leo, Piedecuesta, Colombia, and
Centro de Qumica, IVIC, Apartado 21827, Caracas 1020A, Venezuela

ReceiVed July 20, 2006. ReVised Manuscript ReceiVed February 5, 2007

A method was developed, in which the integral areas of resonance signals in 12 regions of the 1H nuclear
magnetic resonance (NMR) spectra of 5 widely different Colombian crudes and a crude mixture were correlated
with selected physicochemical properties and standard distillation cut yields. The crudes used ranged from
very light to heavy aromatics, thus covering a wide range of chemical compositions. The relationship between
the NMR spectra and the properties of interest was obtained with the partial least-squares (PLS) algorithm.
High correlation values were obtained for all of the PLS models used. The R2 values of the prediction models
for the physicochemical properties and SimDis results were >0.95, while the CV-q2 values fluctuated between
0.74 and 0.99. The properties of an unknown crude or crude mixtures can be obtained from their 1H NMR
spectra and the use of the PLS models developed in this work. The correlation method is very fast (1 h instead
of several weeks) and much less costly than the conventional ones.

Introduction structure-activity relationship in hydrocarbons have been

proposed using 13C and 1H NMR spectra.8 Properties such as
One of the main goals of refineries is to maximize the yield
boiling point, specific gravity, refractive index of middle distil-
of valuable products using different types of crudes and their
lates, density, number, cetane index, sulfur content, final boiling
mixtures while maintaining the quality and the specifications
point of the diesel fractions, and the octane number of gasoline
of the refining products. Usually, to know the yield and the
have been predicted with high accuracy.2-10 Therefore, it seemed
quality of the resulting fractions, it is necessary to study and
interesting to extend these type of correlations based on 1H NMR
evaluate each crude1 in distillation units at the laboratory level.
spectra to the prediction of different properties of the crudes.
There, the different cuts are obtained; their physicochemical
In this work, the widest available variety of Colombian crudes
properties are measured; and the operational scheme of the
was studied, from the most light paraffinic to the heaviest
refinery that generates the best yields is, then, determined. The
aromatic ones. A typical naphtenic crude was also included to
time required for these studies is between 3 and 4 weeks and
have a wide set. It is expected that other Colombian crudes will
could cost up to 15 000 dollars/crude, depending upon the
fall within this set. Using their 1H NMR spectra, a method was
number of cuts and the type of physicochemical analysis
developed that is able to predict the yield of the typical products
required. Faster and less expensive methods are thus of interest
at standard cuts and the most important physicochemical prop-
to optimize valuable refinery outputs at reasonable prices.
erties of an crude or mixture. The crudes used had a very wide
Nuclear magnetic resonance (NMR) spectroscopy have been
distribution of chemical compositions, thus allowing the deter-
extensively applied to characterize the different fractions of
mination of the role of each major hydrocarbon family on the
crude oils through the determination of averaged molecular
crude properties and standard cut product yields. Therefore, these
parameters of asphaltenes, resins, vacuum residues, etc. or the
prediction models may be applied to crudes falling within these
quantification of other hydrocarbon families, such as paraffins,
compositional limits. Moreover, these methods may also allow
naphtenics, and aromatics.2-7 Predictive models based on the
for the selection of the most convenient set of crudes or crude
* To whom correspondence should be addressed. E-mail: dmolina@ mixture for a specific refinery setting before they are acquired
uis.edu.co. in large quantities. In this way, costly studies and refinery trial
Universidad Industrial de Santander.
Ecopetrol-Instituto Colombiano del Petro
runs may be avoided with the corresponding savings in cost.1,11
leo.
IVIC.
Experimental Section
zum, B. Petroleum Refining Processes; Dekker: New
(1) Speight J. G.; O
York, 2002. To have a wide range of crude chemical and molecular
(2) Gillet, S.; Delpuech, J.-J.; Valentin, P.; Escalier, J.-C. Anal. Chem. composition, a selection of paraffinic, intermediate, naphtenic, and
1980, 52, 813-817. aromatic (heavy) Colombian crudes were used in this study. Table
(3) Gillet, S.; Rubini, P.; Delpuech, J. J.; Escalier, J-C.; Valentin, P. Fuel
1981, 60, 221-225. (8) Hue, Z.; Ring Z.; Liu, S. Energy Fuels 2005, 19, 152-163.
(4) Brown, J. K.; Ladner, W. R. Fuel 1960, 39, 87-96. (9) Kapur, G. S.; Ecker, A.; Meusinger, R. Energy Fuels 2001, 15, 943-
(5) Rongbao, L.; Zengmin, S.; Bailing, L. Fuel 1988, 67, 565-569. 948.
(6) Kapur, G. S.; Singh, A. P.; Sarpal, A. S. Fuel 2000, 79, 1023-1029. (10) Meusinger, R. Fuel 1996, 75, 1235-1243.
(7) Sarpal, A. S.; Kapur, G. S.; Mukherjee, S.; Tiwari, A. K. Fuel 2001, (11) Speight, J. G. The Chemistry and Technology of Petroleum, 3rd
80, 521-528. ed.; Dekker: New York, 1999.

10.1021/ef060335n CCC: $37.00 2007 American Chemical Society

Published on Web 03/31/2007
1H NMR Spectra of Colombian Crude Oils Energy & Fuels, Vol. 21, No. 3, 2007 1675

Table 1. Properties of Colombian Crudes

crude ICa API type
crude 1 17.93 45.1 light paraffinic
crude 2 19.60 43.0 paraffinic
crude 3 32.26 29.0 paraffinic intermediate
crude 4 44.00 23.8 naphtenic
crude mixture 48.13 22.7 naphtenic intermediate
crude 6 72.18 12.3 aromatic heavy
a IC is a correlation index defined as: IC ) (48 640/T
50%) + 473.7F -
456.8, where F the specific gravity at 15 C and T50% is the temperature at
50% of the distillation curve (see refs 11-13).

Table 2. Hydrogen Types Present in the NMR Spectra of Crudes

and the Corresponding Areas of Integration
spectrum
portion (ppm) area hydrogen type
0.5-1.0 H1 CH3s and some naphtenic CH and CH2s
1.0-1.7 H2 CH2s and some CHs
1.7-1.9 H3 mostly CH and CH2 in positions
1.9-2.1 H4 RCH3 in olefins
2.1-2.4 H5 RCH3s in aromatic carbons
2.4-3.5 H6 RCH and CH2s in aromatic carbons
3.5-4.5 H7 bridging CH2s
4.5-6.0 H8 olefins
6.0-7.2 H9 mono-aromatics
7.2-8.3 H10 di-aromatics and some tri- and tetra-aromatics
8.3-8.9 H11 some tri- and tetra-aromatics
8.9-9.3 H12 some tetra-aromatics
1 shows the properties and classification of the different crudes
and a typical crude mixture that represents a real feed to the
Ecopetrol-Barrancabermeja refinery used in this work. The range
covered was from the very light one (crude 1) up to the heavy
aromatic one (crude 6). Atmospheric distillation was performed in
a 50 L unit [American Society for Testing and Materials (ASTM)
D-2892], while vacuum distillation was performed in a 20 L unit
(ASTM D-5236).
The fractionation and physicochemical characterization of the
crudes and its cuts were performed at the Research Center of the
Colombian National Petroleum Company (Instituto Colombiano del
Petroleo, Ecopetrol-ICP), whose laboratories are accredited with
the ISO 17025 standard in 2005. All of the physicochemical
measurements were performed according to the ASTM methods
available for each specific case.
The 1H NMR spectra of the crudes were obtained with Bruker
ACP 300 MHz spectrometers in 5 wt % solutions in CDCl3. This
concentration was found to be the best one to avoid concentration
dependence of the chemical shifts and, at the same time, provided
an excellent signal-to-noise ratio. The 30 pulses (Bruker zg30 pulse
sequence) were 9.6 s, and the delay time between scans was 4 s.
Measurements were performed with a delay time varying by 1, 2,
4, 6, and 10 s. The areas obtained at 10 s were compared with
those obtained at the other delay times. It was found that, except
for the ones obtained at 1 s delay time, all of the areas were the
same within experimental error. Then, most measurements were
performed at 2 s, and some were performed at 4 s (sweep width,
3937.0; dwell time, 254 s; offset frequency, 1814.8; number of
points, 16 384; line broadening, 0.1 Hz). Tetramethylsilane (TMS)
was used as reference, and 16 scans were averaged for each
spectrum employed in the correlations. The phase and baseline of
the resulting spectra was manually adjusted and corrected after using
a very large expansion of the y axis.
The 1H NMR spectrum of crudes is very complex and contains
a large number of signals because of the large distribution of
chemical shifts reflecting their great chemical complexity. The 1H
NMR spectra of the crudes are highly overlapped and crowded;
therefore, it is not possible to make an assignment of each of the
lines with respect to specific molecular structural types.2-10
However, it is possible to assign a portion of the spectrum to specific
structural groups, such as aromatics, paraffins, naphtenes, etc. The
spectrum was divided in 12 regions as shown in Table 2 and Figure
Figure 1. 1H NMR spectrum of crude 6 and the spectral partition
according to the different chemical shifts.
1. Each of these spectral regions encompasses well-defined atomic
groups, although some minor overlapping may still be present. The
integration was performed within each of these spectral segments,
minimizing the small shifts found from sample to sample. This was
achieved by carefully setting the TMS signal at 0 ppm. The resulting
integrals of the different segments of the spectrum were normalized,
and to have high accuracy, six integrations were averaged per each
spectral segment. Finally, the averaged integrals were correlated
with the physicochemical properties, product yields, and SimDis
curves. The total time required including preparation was typically
60 min per sample.
Correlation Method. The partial least-squares (PLS) regression
method is widely used for the prediction of properties of interest
of certain samples based on measurement, which are not directly
related to the properties in a known defined relationship.14 A typical
example of its application is the correlation of several important
coal properties to the fitted parameters of their 1H NMR spectra.15
This regression method is the least restrictive of the various
multivariate extensions of the multiple linear regression models.
This flexibility allows it to be employed in cases where the use of
multivariate methods is strongly limited, such as when there are
fewer observations than predictor variables.14
The PLS regression procedure has two steps:14 (1) training or
calibration and (2) a prediction run. In the calibration step, the PLS
method requires the response spectrum x (the 12 areas of the NMR
spectrum) and a property q (API, percent naphtas, etc.), for each
of the samples. The vectors x are then used as rows of a matrix X,
and the list of properties (q values) form a column vector. The
purpose of the training step is to obtain a quantitative relationship
between the response X (the NMR spectrum) and the property of
the sample q. Once this is achieved, the relationship is saved and
used to predict that property in a new sample using its NMR
spectrum (for further details, see the Appendix) During training,
the PLS regression equations were repeatedly employed using an
incremental number of factors or components.14,15 The more factors
that are considered, the more the structure in X is used to model
the relationship with the property q. Using only a few factors
excludes valuable information, and including too many factors
decreases the accuracy of the prediction. The best model then must
be chosen from the comparison between the prediction accuracy
and the number of PLS factors employed in the calculation.14 It
(12) Wauquier, J. P. Petroleum RefiningsCrude Oil Petroleum Products;
Editions Technip: Paris, France, 1995.
(13) Altgelt, K. H.; Boduszynski, M. M. Composition and Analysis of
HeaVy Petroleum Fractions; Dekker: New York, 1994.
(14) Otto, M. Chemometrics; Wiley-VCH: New York, 1999.
(15) Harmer, J. R.; Callcott, T. G.; Maeder, M.; Smith, B. E. Fuel 2001,
80, 1341-1349.
1676 Energy & Fuels, Vol. 21, No. 3, 2007 Molina V. et al.

Table 3. Fraction Yields of the Different Crudes

fraction range (C) crude 1 crude 2 crude 3 crude 4 crude 5a crude 6
Atmospheric Distillation Yield (Mass %)
gases C1-C4 2.4 2.9 0.3 0.2 0.4 0.4
naphta 1 15-60 5.8 4.8 0.6 0.7 1.0 4.5b
naphta 2 60-107 12.4 11.2 2.0 2.3 2.7
naphta 3 107-152 12.8 10.8 4.7 3.4 3.1
naphta 4 152-177 5.4 6.1 4.1 2.5 2.7
naphta 5 177-199 4.6 3.3 3.8 2.5 2.3
kerosene 1 199-221 3.6 3.0 3.1 2.6 2.2 7.0c
kerosene 2 221-249 6.3 5.7 5.5 3.9 3.8
diesel 1 249-288 8.6 10.7 8.6 6.0 6.6 3.7
diesel 2 288-327 8.3 9.1 8.7 7.6 7.6 6.2
diesel 3 327-371 7.6 8.4 8.6 8.3 8.2 3.2
atmospheric residue 371+ 22.2 24.0 50.0 60.0 59.4 75.0
Vacuum Distillation Yield (Mass %)
vacuum gas oil 1 371-399 3.1 3.2 3.1 2.6 1.8 6.2
vacuum gas oil 2 399-427 5.3 5.5 6.5 6.2 7.4 5.6
vacuum gas oil 3 427-482 7.2 7.0 10.4 11.4 10.6 9.4
vacuum gas oil 4 482-510 1.7 1.9 4.0 4.7 4.8 4.3
vacuum gas oil 5 510-550 1.8 1.9 1.8 3.0 2.6 2.3
vacuum residues 550+ 3.2 4.4 24.4 32.0 32.3 47.3
a Crude 5 is a typical blend of crudes used in the refinery of Ecopetrol-Barrancabermeja. b Naphta with a distillation range between 15 and 150 C.
c Kerosene cut between 150 and 250 C.

Table 4. Standard Cut Product Yields of the Different Crudes

yield (mass %)
fractions range (C) crude 1 crude 2 crude 3 crude 4 crude 5 crude 6
naphta PIE-160 33.6 30.0 9.8 7.5 8.0 4.5
jet-A fuel 160-250 19.2 19.4 15.2 10.9 10.7 5.9
diesel fuel 250-343 19.7 20.8 20.0 16.7 17.0 13.3
vacuum gas oils 343-510 22.5 23.5 28.8 30.0 29.4 26.7
vacuum residue 510+ 5.0 6.3 26.2 35.0 34.9 49.6

was found that the use of five factors (C1-C5) in the first case
(physicochemical and product yield) provided the best prediction
accuracy, while in an second one (SimDis analysis), only three
factors and their products were found to give the best results. A
lower and higher number of factors generated a lower accuracy in
all cases. Initially, in all of these models using the PLS method, R2
was >0.99, employing the five factors obtained in the training step
(see Table 7). The high value of R2 led us to think that some
overfitting may be present. To avoid this problem, the statistical F
method was employed.16 To select the optimal model in each case,
we computed the predicted residual sum of squares (PRESS) for
cross-validated models with various numbers of factors h. The
model with the minimum PRESS is used as a benchmark, and the
number of factors associated with it was designated by h*. All
models with h < h* were compared against this benchmark. The
smallest model (fewest number of h factors), such that its PRESS
was not significantly greater than the PRESS for the model with
h* factors, was found this way. The selection started by computing
F(h) ) PRESS (model with h factors)/PRESS (model with h*
factors) for h ) 1, 2, ..., h*. Then, in the next step, the optimal
number of factors (the smallest h) was chosen, such that F(h) <
FR;m,m, where FR;m,m is the (1 - R) percentile of Snedecors F
distribution with m and m degrees of freedom (m ) number of
calibration samples). If the prediction errors have a 0 mean and
are mutually independent and normally distributed, Prob {F(h) >
Fa;m,m|h2 ) h*2} ) 2R. Here, h2 and h*2 are the prediction error
variance of the model with h and h* factors, respectively, and R is
a probability value to be selected. In our case, we used R ) 0.25
to have a reasonable balance between over- and under-fitting.16
To estimate prediction statistics, a leave-one-out cross-validation
(CV) regression method was employed.14 The best prediction model
is the one with the larger CV-q2. The resulting models for the
correlations are shown in the next section.
The correlation between physicochemical properties and NMR
spectra of crudes was determined using the univariate PLS method,
(16) Haaland, D.; Thomas, E. Anal. Chem. 1988, 60, 1193-1202.
while the yields of the different cuts were obtained from the
multivariate PLS method.14 Both methods are part of the PLS
module of Statistica 6.0.17
Results and Discussion
Table 3 shows the yields of the different fractions selected
in this work for the evaluation of each crude, and Table 4 shows
the yields of the most important refinery distillation products
at standard cut points, obtained from the crude oil simulated
distillation curve SimDis.12,13 These results have to be similar
to those obtained from the sum of the different cut yields of
Table 3 included in each product. Figure 2 shows the SimDis
curves of crude 2 and its fractions.
As seen in Figure 2, the overlap between the different cuts
of crude 2 is small; therefore, the yield and the properties
measured correspond to the hydrocarbon families present and
are representative of that distillation range. The distillation
curves obtained for all of the remaining crudes and its cuts were
very similar to that of crude 2 and are not shown.
Table 5 shows the physicochemical properties of the
crudes studied in this work. The American Petroleum Institute
(API) gravity values,11-13 KUOP factor,11-13 wax content, and
ion chromatography (IC) values11,13 confirm that crude 1 is the
most paraffinic one, while crude 6 is the most aromatic one.
This classification shows that S, N, V, and Ni content, insol-
ubles in nC7 and nC5, and the micro carbon residue (MCR)13
increase progressively when the paraffinic hydrocarbons de-
crease and the aromatic resins and asphaltenes increase, which
is corroborated with the increase in the density, mean aver-
age boiling point (MeABP) value,13 and viscosity. Crude 4
shows the lowest pour point and the largest acid number,
confirming that it is the more naphtenic of the studied
Colombian crudes.
1H NMR Spectra of Colombian Crude Oils Energy & Fuels, Vol. 21, No. 3, 2007 1677

Figure 2. SimDis curves of crude 2 and all of its fractions.

Table 5. Physicochemical Properties of Different Types of Colombian Crudes

properties crude 1 crude 2 crude 3 crude 4 crude 5 crude 6
density at 15 C (g/mL) 0.801 0.8103 0.8809 0.9106 0.9169 0.9836
gravity API (API) 45.1 43 29 23.8 22.7 12.3
acid number (mg of KOH/g) <0.05 <0.05 0.154 2.034 1.343 0.100
sulfur (mass %) 0.045 0.129 0.534 0.808 1.276 2.185
flash point (C) e15.0 e45 5 8 12 21
pour point (C) 9 6 -9 e33.0 -18 3
nitrogen (ppm) 179 264 2477 2520 3228 4856
wax content (mass %) 15.84 7.78 6.57 0.18 4.44 1.58
MCR (mass %) 0.23 0.77 5.69 5.18 6.69 15.85
insolubles in nC7 (mass %) 0 0.14 2.16 1.29 4.27 14.58
insolubles in nC5 (mass %) 0.03 0.32 5.75 3.21 7.1 18.72
vanadium (ppm) 0.151 3.56 12.84 36.01 93.71 421.80
nickel (ppm) 0.087 1.05 40.41 20.71 47.77 82.60
MeABPa (50% SimDis, C) 237.4 252.5 368.9 423.4 416 488.8
KUOP factor 12.14 12.12 11.91 11.84 11.72 11.30
correlation index 17.93 19.60 36.26 44.4 48.13 72.98
kin. viscosity at 40 C (cSt) 2.04 2.22 15.89 1.98 55.62 3510.60
simulated distilation (C)
IBPb -16.9 -14.7 68.3 58.1 38.2 69.6
5% 37.3 53.6 128.3 134.7 130.4 171.1
30% 146.8 160.8 272.5 316.8 314 380
50% 237.4 252.5 368.9 423.4 416 488.8
70% 328.3 341.9 474.6 533.7 524.7
90% 450.8 463.9
95% 510.1 536.4
a MeABP ) mean average boiling point. b IBP ) initial boiling point.

Equations 1-5 show the resulting polynomials for the Tables 6 and 7 show the values of R2 and CV-q2 found for
prediction of the crude distillation products obtained with the correlation of the standard cut product yields. The prediction
PLS correlation from 1H NMR data. for the yield of the jet A and vacuum residue is excellent (R2 >
0.98, and CV-q2 > 0.96) and similarly for naphta, diesel, and
percent naphtas ) -55.34 - 3.5532C1 + vacuum gas oil (R2 > 0.98, and CV-q2 > 0.90). All of the cuts
6.349 17C2 - 22.2954C4 (1) showed overlap; therefore, they have a higher correlation. This
was particularly noticeable for the heavier crudes, where an ad
percent jet A ) -21.5783 + 2.397 82C1 - hoc interpolation was necessary to obtain all of the gas oil cuts
0.0289 511C1C1 (2)
at the same final temperature.
percent diesel ) -21.1064 + 3.520 07C1 - Equations 6-21 show the resulting polynomials for the
0.074 790 3C1C1 (3) prediction of the physicochemical properties and SimDis
percent gas oil ) -29.0943 + 6.5898C1 - 0.184 131C1C1 curves obtained with the PLS correlation17 from 1H NMR
(4) data.
percent residue ) 80.8482 - 1.355 31C1 -
0.073 263 6C1C1 (5) (17) Statistica 6.0. StatSoft, Inc.: Tulsa, OK.
1678 Energy & Fuels, Vol. 21, No. 3, 2007 Molina V. et al.

Table 6. Predicted and Measured Products (LOO)a and CV-q2

naphta jet-A fuel diesel vacuum gas oil vacuum residue
crude exp. calc. LOO exp. calc. LOO exp. calc. LOO exp. calc. LOO exp. calc. LOO
1 33.6 29.7 19.2 19.8 19.7 20.7 22.5 23.1 5.0 3.7
2 30.0 31.5 19.4 19.3 20.8 20.5 23.5 23.7 6.3 6.1
3 9.8 15.4 15.2 14.3 20.0 18.7 28.8 27.7 26.2 26.2
4 7.5 6.00 10.9 10.6 16.7 16.8 30.0 29.1 35.0 37.6
5 8.0 7.1 10.7 10.1 17.0 16.5 29.4 29.1 34.9 38.8
6 4.5 3.4 5.9 7.0 13.3 14.2 26.7 28.2 49.6 44.0
CV-q2 0.934 0.979 0.904 0.903 0.964
R2 99.98 99.81 96.76 99.75 98.82
a LOO ) leave one out.

Table 7. Number of Factors and Relationship with R2 and CV-q2

S (mass %) ) 7.367 + 0.459C1 + 0.277C2 + 0.226C3 - for the Different Fractions
0.376C4 (6) number of factors

N (ppm) ) 18 512.8 + 405.346C1 + 513.559C3 + select

fraction 1 2 3 4 5 (eqs 1-5)
213.942C4 (7)
R2 85.03 89.12 91.81 99.98 100.00 99.98
naphta
waxes (wt %) ) 8.283 + 2.232C1 + 1.393C2 + CV-q2 0.433 0.522 0.743 0.954 0.955 0.934
1.326C4 - 2.212C5 (8) jet A
R2 99.37 99.40 99.49 99.77 100.00 99.81
CV-q2 0.798 0.798 0.956 0.959 0.999 0.98
MCR (mass %) ) 80.295 + 2.972C1+ 2.505C3 + diesel
R2 86.15 87.68 90.59 97.83 100.00 96.76
CV-q2 0.67 0.693 0.774 0.848 0.954 0.904
1.670C5 (9) vacuum R2 49.43 66.70 75.28 96.03 100.00 99.75
insolubles in nC7 (mass %) ) 59.7892 + 3.688C1 + gas oil CV-q2 -5.632 -5.327 -0.908 -0.584 0.829 0.903
vacuum R2 98.56 98.77 99.16 99.95 100.00 98.82
1.035C2 + 1.290C3 + 0.026C5 (10) residue CV-q2 -0.675 -0.660 0.486 0.489 0.994 0.964
V (ppm) ) 3048.35 + 279.188C1 + 8.910 52C2 + Table 8. R2 and CV-q2 Values of the Physicochemical Properties
28.395C3 + 5.918 18C1C1 (11) and SimDis Curves

Ni (ppm) ) 322.633 + 12.1196C1 + 3.951 38C2 + distillation

property CV-q2 R2 curve SimDis CV-q2 R2
9.993 72C3 - 0.137 05C1C1 (12)
sulfur (wt %) 0.823 99.99 IBP 0.742 74.40
initial boiling point (IBP) (C) )67.6721 + 9.9033C1 + N (ppm) 0.855 99.97 5% 0.893 93.67
0.239 101C2 - 10.0075C3 (13) waxes (wt %) 0.986 99.74 10% 0.920 97.61
MCR (wt %) 0.846 99.93 20% 0.950 98.50
5% (C) ) 406.338 + 10.6858C1 - 2.088 66C2 + insolubles in nC7 (wt %) 0.882 99.97 30% 0.950 98.78
vanadium (ppm) 0.992 99.72 40% 0.949 99.05
4.440 92C3 (14) nickel (ppm) 0.917 99.73 50% 0.948 99.00
10% (C) ) 576.451 + 18.3643C1 + 4.921 91C2 + 60% 0.938 98.74
70% 0.951 98.22
6.648 23C3 (15)
20% (C) ) 584.961 + 22.9827C1 + 2.701 21C2 - with CV-q2 values > 0.82. The correlation found between the
3.550 01C3 (16) areas of the 1H NMR spectrum and the V and Ni content is
quite interesting and deserves further analysis. Protons near the
30% (C) ) 676.59 + 23.0699C1 - 0.515 013C2 -
paramagnetic metallic centers are influenced by their presence
2.7897C3 (17)
in compounds, such as the vanadyl porphyrines. The effect of
40% (C) ) 728.906 + 23.0109C1 - 2.043 65C2 - an unpaired electronic spin is so strong that the NMR lines of
3.398 23C3 (18) most of the protons of the molecule containing a metallic center
will not be observed. The influence of these centers change the
50% (C) ) 764.909 + 24.7934C1 - 1.282 21C2 -
chemical shifts and areas of the spectrum of more distant
6.059 69C3 (19)
protons, where the density of the unpaired spin tends to be
60% (C) ) 873.595 + 27.2275C1 + 0.905 615C2 - negligible or 0. The effect of these centers will also influence
4.1236C3 (20) protons belonging to the closest surrounding molecules, thus
extending their perturbation range beyond the molecules con-
70% (C) ) 873.322 + 28.1251C1 - 0.743 77C2 -
taining the metallic centers. This effect will influence quite
9.5666C3 (21) different types of protons, thus affecting the whole NMR
Also, in all of these PLS equations, R2 was good at employing spectrum. This mechanism may well provide the basis for
the F statistic to select the optimal number of factors obtained explaining the correlation found between the 1H NMR spec-
in the training step for the physicochemical properties. Table 8 trum of crudes and their V and Ni content by the PLS
shows the R2 and CV-q2 values of the physicochemical method.
properties and SimDis curves (eqs 6-21). The models obtained for the distillation curves of the crudes
The values of CV-q2 for physicochemical properties (V (SimDis) showed that the present analysis of the 1H NMR data
content and waxes) show that an excellent correlation can be can be used to construct the total boiling point (TBP) curve of
obtained using the 1H NMR spectra of crudes. The model for crudes with high accuracy. The values of both R2 and CV-q2
the Ni content was in good quality, and the others are acceptable (except for the IBP and 5%) show the degree of accuracy that
1H NMR Spectra of Colombian Crude Oils Energy & Fuels, Vol. 21, No. 3, 2007 1679

Figure 3. Experimental and predicted SimDis curves for crudes 1, 3, and 5.

may be achieved with the PLS method. Figure 3 shows the experimental and predicted SimDis values for three of the six crudes
studied in this work.
The correlations obtained between the measured and predicted values by the PLS method using eqs 13-21 are seen in Figure 3.
The same type of high-quality fitting was obtained with the PLS method with crudes 2, 4, and 6 and is not shown for the sake of
brevity.
Once the basic correlation using these well-known crudes is set, the prediction of the product yields and physicochemical properties
requires the determination of the 1H NMR spectrum of the crude or mixture under investigation. The total time for such determination
will be, in general, around 60 min per sample. Both off- and online application of this new methodology can be envisioned for
refinery feedback control and optimization.

Conclusion

The use of 1H NMR spectra provides a very fast and relatively inexpensive method to predict the physicochemical properties and
the refining product yields for a wide variety of crudes and their mixtures using the PLS regression method.

Acknowledgment. The authors thank to the Unidad de Investigacion of Ecopetrol-Instituto Colombiano del Petroleo for gen-
erous financial support (Convenio number 004 2003) and allowing for the publication of this paper. The authors also thank Drs.
Ricardo Fierro from Universidad Nacional de Colombia and Winston Quinonez from Universidad de Antioquia for the NMR spectra
and Jairo Rene Martnez from Universidad Industrial de Santander for acting as the University Coordinator of the Convenio with
Ecopetrol-ICP.

Appendix

The procedure for a particular property (percent naphta) will be outlined here to show the different steps involved in the correlation
between the areas of the NMR signals and properties. The notation employed is such that the area NMR1 (H1) of crude 1 is 1H1,
while area 2 for the same crude is 1H2, etc. and for crude 2 and H1 area it will be 2H1, etc. These values formed matrix X, while

[ ] []
the different properties generated matrix q as shown in eq A1.

1H1 1H2 1H3 1H4 1H5 1H6 1H7 1H8 1H9 1H10 1H11 1H12 33.6
2H1 2H2 ... 2H12 30.0
3H1 9.8
X) q) (A1)
... 7.5
... 8.0
6H1... ... 6H12 4.5

The application of PLS produces the weights (wt) for the different areas. In the present case, the results are shown in
Table A1.
1680 Energy & Fuels, Vol. 21, No. 3, 2007 Molina V. et al.

Table A1. Weight of Variables in Matrix X

H1 H2 H3 H4 H5 H6
W1 -0.323 153 -0.327 058 -0.231 773 -0.081 934 -0.282 105 -0.249 314
W2 -0.004 961 -0.001 773 -0.312 771 -0.550 686 -0.233 628 -0.053 086
W3 0.039 645 -0.075 135 -0.395 929 0.216 320 -0.148 720 -0.405 809
W4 0.107 277 0.098 138 0.391 564 -0.067 503 -0.152 031 0.187 807
W5 -0.374 811 -0.283 593 -0.162 309 -0.012 336 -0.090 791 0.717 032

H7 H8 H9 H10 H11 H12

W1 -0.350 734 -0.335 602 0.154 011 -0.348 289 0.340 543 0.299 659
W2 0.133 401 -0.011 630 -0.603 306 0.385 472 -0.089 759 -0.054 178
W3 0.052 660 0.096 729 0.226 367 0.494 334 -0.482 267 0.249 223
W4 -0.021 144 0.061 814 -0.270 972 -0.129 403 -0.190 296 0.794 950
W5 0.039 552 0.030 362 0.218 477 0.244 953 -0.271 511 0.215 126

With Table A1 and matrix X, the factors (Ci) were calculated as followed:

C1 of crude 1 ) (-0.323 153 1H1) + (-0.327 058 1H2) + (-0.231 773 1H3) + ...
(0.299 659 1H12*) ) 24.181 637 14
C3 of crude 4 ) (0.039 645 4H1) + (-0.075 135 4H2) + (-0.395 929 4H3) + ...
+ (0.249 223 4H12) ) -6.216 845 619...
The values for the Cis for each crude are shown in Table A2.
Table A2. Cis Factors
C1 C2 C3 C4 C5 percent naphtas
crude 1 24.181 637 14 -5.324 013 79 -13.303 423 45 13.955 121 76 -5.268 423 249 33.6
crude 2 23.823 380 33 -7.882 114 39 -11.475 055 96 15.662 478 95 -5.192 700 979 30.0
crude 3 20.291 228 92 -3.902 492 67 -11.250 672 53 14.354 718 37 -4.554 571 469 9.8
crude 4 17.234 296 56 -3.725 528 32 -6.216 845 619 15.025 922 67 0.289 819 827 7.5
mixture 16.768 961 67 -3.435 588 48 -6.772 319 875 14.010 889 33 -2.030 028 194 8.0
crude 6 13.718 716 74 -4.110 847 78 -5.546 934 884 13.589 306 92 -1.646 006 628 4.5

C1-C5 are taken as independent variables and the percentage of naphtas is dependent for a MLR regression. Several models with
different Cis were considered. To each of them, leave one out (LOO), CV-q2, and the F factor analysis were applied. The chosen
model had to comply with the F factor criteria and show the largest R2, and its CV-q2 value had to be the closest to 1. The resulting
equation for the percentage of naphtas is shown in eq 1. All of the other models (eqs 2-21) were obtained by a similar procedure.
EF060335N