See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/303300288

Determination of some physicochemical properties in Brazilian crude oil by

1H NMR spectroscopy associated to chemometric approach

Article  in  Fuel · October 2016

DOI: 10.1016/j.fuel.2016.05.049

CITATIONS READS

6 187

7 authors, including:

Lucas Duarte Paulo Roberto Filgueiras

Federal University of Juiz de Fora Universidade Federal do Espírito Santo
7 PUBLICATIONS   19 CITATIONS    96 PUBLICATIONS   247 CITATIONS   

SEE PROFILE SEE PROFILE

Lize Mirela Oliveira Eustáquio Vinicius Ribeiro de Castro

Petróleo Brasileiro S.A. Universidade Federal do Espírito Santo
11 PUBLICATIONS   72 CITATIONS    116 PUBLICATIONS   1,372 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Applications of MicrNIR View project

Projeto Emulsão polimerica View project

All content following this page was uploaded by Lucas Duarte on 08 January 2018.

The user has requested enhancement of the downloaded file.

 Fuel 181 (2016) 660–669

Contents lists available at ScienceDirect

Fuel
journal homepage: www.elsevier.com/locate/fuel

Full Length Article

Determination of some physicochemical properties in Brazilian crude oil

by 1H NMR spectroscopy associated to chemometric approach
Lucas M. Duarte a, Paulo R. Filgueiras b, Samantha R.C. Silva b, Julio C.M. Dias c, Lize M.S.L. Oliveira c,
Eustáquio V.R. Castro b, Marcone A.L. de Oliveira a,⇑
a
Analytical Chemistry and Chemometrics Group – GQAQ, Departament of Chemistry, Institute of Exact Sciences, Federal University of Juiz de Fora, University City,
CEP 36036-330 Juiz de Fora, MG, Brazil
b
Center of Competence in Petroleum Chemistry – NCQP, Laboratory of Research and Development of Methodologies for Analysis of Oils – LabPetro, Chemistry Department,
Federal University of Espírito Santo, Av. Fernando Ferrari, 514, CEP: 29075-910 Vitória, Espírito Santo, Brazil
c
CENPES/PETROBRAS, CEP 21941-598 Rio de Janeiro, RJ, Brazil

h i g h l i g h t s g r a p h i c a l a b s t r a c t

API gravity, carbon residue, wax

appearance temperature and basic
organic nitrogen in crude oil were
determined.
1

H NMR spectroscopy associated to
partial least squares regression to
determine such properties were used.

A previous study involving pattern
recognition to better known of
sample set was performed.

Principal component analysis was
used as pattern recognition method.

Accuracy, confidence intervals,
systematic and tendency errors were
evaluated in each models fitted.

a r t i c l e i n f o a b s t r a c t

Article history: An alternative method for establishing crude oil properties, using 1H nuclear magnetic resonance (1H
Received 16 February 2016 NMR) associated with partial least squares regression is proposed. It can be used for determination of
Received in revised form 6 May 2016 API gravity, carbon residue (CR), wax appearance temperature (WAT) and basic organic nitrogen
Accepted 9 May 2016
(BON). At a 95% confidence level, the main results obtained for API gravity, CR, WAT and BON models pro-
vide determination coefficients for prediction (R2p) and root mean square error for prediction (RMSEP)
equal to 0.945, 0.802, 0.857 and 0.789; and 0.8, 0.598% w/w, 3.8 °C, 0.009% w/w, respectively. The resid-
Keywords:
uals of each fitted model were evaluated taking into account systematic and random errors and all results
Crude oil
Physicochemical properties
were considered acceptable. The determination of these physicochemical properties in Brazilian crude oil
PCA was successfully achieved using a chemometric approach in association with 1H NMR.
PLS Ó 2016 Elsevier Ltd. All rights reserved.
1
H NMR

1. Introduction

Petroleum, meaning literally ‘‘rock oil”, is the term used to

describe variety of hydrocarbon-rich fluids that have accumulated
⇑ Corresponding author. in subterranean reservoirs. Crude oil varies dramatically in color,
E-mail address: marcone.oliveira@ufjf.edu.br (M.A.L. de Oliveira). odor and flow properties that reflect the diversity of its origin

http://dx.doi.org/10.1016/j.fuel.2016.05.049
0016-2361/Ó 2016 Elsevier Ltd. All rights reserved.
 L.M. Duarte et al. / Fuel 181 (2016) 660–669
[1]. Chemically, crude oils are formed in most of its constitution by
different hydrocarbons families – paraffinic, naphthenic or aro-
matic compounds – giving them different physicochemical charac-
teristics [2]. The wide range of compounds present in crude oil
implies a complex characterization, which is made by checking
for over twenty physicochemical parameters such as density, car-
bon residue, viscosity, refractive index, among others. This culmi-
nates in a long time evaluation, many times reaching six months
and resulting in higher economic costs.
The API gravity is an arbitrary scale for measuring the density of
petroleum and its derivatives and reference methods are given by
ASTM D4052 [3] (American Standards for Testing and Materials)
and ISO 12185 [4]. On the other hand, the amount of carbon resi-
due (CR) generated by an oil suggests an indication of the relative
tendency to oil coke formation and the ASTM D4530 is the official
methodology used to determination [5]. Regarding the wax
appearance temperature (WAT), such monitoring is of great impor-
tance because it assists in preventing blockage of the fluid by pre-
cipitation on the walls of pipelines, production plants and
transportation pipelines during the oil and gas production. The
WAT is an indicator of how well the oil will leak under cold condi-
tions. The WAT can be analyzed in crude oil by cross polarization
microscopy, viscosimetry, thermomicroscopy, fluidity point, and
differential scanning calorimetry (DSC) [6]. In crude oil is very
important to investigate and identify nitrogen compounds such
as basic organic nitrogen (BON), which could be related to the effi-
ciency decrease in the catalytic processes and contribute to the
instability of the products associated to crude oils during storage
step [7]. The BON is officially analyzed by UOP 269 method [8].
In order to save time and cost for the determination of the many
physicochemical parameters of oil, in recent years many studies
involving spectroscopy associated to chemometrics approach has
gained prominence in scientific literature due to its intrinsic
advantages such as non-destructive method, compatible efficiency,
quickness and low cost compared with conventional methods rec-
ommended by ASTM. Aiming the crude oil characterization, quan-
titative methods developed by multivariate calibration using
infrared data [9–12] are the most common found in literature
and some of these works are also shown as mid infrared (MIR)
[13,14] and near infrared (NIR) [15–17] according to the spectral
range used. Another kind of spectroscopy that has been gaining
prominence for crude oil analysis is the nuclear magnetic reso-
nance (NMR) due to the qualitative (in molecular level) and quan-
titative power for investigation of chemistry and physical
properties in complex mixtures [18–22]. Molina et al. [23] in
2007 analyzed a wide range of crudes using proton NMR (1H
NMR) associated to partial least squares (PLS) in order to estimate
cut product yields. Three years later they published other paper
with similar methodology, where vaccum residue of Colombia
crudes were used to predict saturated, aromatic, resin and
asphalthens (SARA) composition and some physicochemical prop-
erties [24]. Masili et al. [25] also used 1H NMR/PLS to predict den-
sity, UOP characterization factor, total acidity number, sulfur
content and true boiling point distillation yields, while Peinder
et al. [26] worked with IR, 1H NMR, 13C NMR and three sets of
merged spectra to predict the long residue properties. A recent
paper applied support vector machine to 1H NMR to predict some
distillation temperatures was published and their results were
compared to conventional PLS method [27]. In our knowledge
there is no previous works to predict BON by 1H NMR associated
to multivariate calibration.
Within this context, the present work aims explore 1H NMR
data in order to build and validate PLS models for determination
of API, CR, WAT and BON in a wide range of Brazilian crude oil from
light to heavy oils. Prior to construct the predictive models, a prin-
cipal component analysis study to the better knowledge of the
661
sample set and outlier identification was performed. The predict
values for all models have been shown within 95% confidence
interval. For the models validation, besides to take into account
the conventional multivariate parameters, the prediction residues
by statistical tests to systematic and tendentious errors were
evaluated.
1.1. Partial least squares
The pioneering work with PLS has been done in seventies by
Herman Wold in Econometrics area. However, using this approach
to chemical applications, initiated by Svante Wold and Harald Mar-
tens, only came in the late seventies after the first publication of
Kowalski and his collaborators [28]. The PLS has its simplest form
used in many areas of chemistry and technology, i.e. in the form of
two predictive blocks. This is a method for relating two data matri-
ces, X and Y, using a linear model that goes beyond traditional
regressions, which also modeling X and Y structures; where X con-
tains the variables (instrumental responses) and Y the property of
interest to be modeled [29]. In the present work have been used
PLS1 and every time that Y is referred to matrix, in fact, Y will be
a column matrix (vector), because just one property will be esti-
mated by each model built. The matrix X, having n samples and
m variables and column matrix Y containing n samples with k
properties values, they are simultaneously decomposed into a
sum of h latent variables according to the following equation:
X
X ðnxmÞ ¼ T ðnxhÞ  P 0ðhxmÞ þ EðnxmÞ ¼ t i  p0i þ ei ð1Þ
h
X
Y ðnxkÞ ¼ U ðnxhÞ  Q 0ðhxkÞ þ F ðnxkÞ ¼ ui  q0i þ f i ð2Þ
h
where T and U, and P and Q refer to the scores and loadings from
the matrices X and Y respectively; E and F relate to residues matri-
ces. The product of T by P approaches the spectral data while U by Q
the reference values.
The regression vector b can be determined by the following
relationship:
 1
b ¼ W P 0ðhxmÞ W Q 0ðhxkÞ ð3Þ
where W is the matrix of weights of the PLS model. The regression
vector b considers the contribution of each variable to the PLS
model, i.e., the higher the b value, indicates how much the variable
for model calibration is important.
2. Material and methods
2.1. Reference methods
In the present work were considered 106 samples of Brazilian
crude oil. Samples from different oil fields – off-shore and on-
shore – located in sedimentary basin of Brazilian coast were used.
These oils showed different characteristics since the sample set
was comprised from light to heavy oils, i.e., with API values cover-
ing the range from 54 to 11.4. The analytical tests of physicochem-
ical properties were performed by standard methods. The API
gravity was determined according to ISO 12185 standard method
[4]. A small, typically less than 1 mL, portion of crude oil sample
was introduced into an oscillating sample tube (Anton Paar DMA
4500 digital density meter) using a dry and clean syringe. The
change in oscillation frequency caused by the change in the mass
of the tube is used to calibrate data and to relate with specific grav-
ity of the sample.
The CR was determined according to ASTM D 4530 (micro) stan-
dard method [5]. The CR is the tendency to form carbon deposits
662 L.M. Duarte et al. / Fuel 181 (2016) 660–669
under high temperature conditions in an inert atmosphere. A
weighed quantity of sample is placed in a glass vial and heated
to 500 °C under an inert (nitrogen) atmosphere in a controlled
manner for a specific time. The sample undergoes coking reactions,
and volatiles formed are swept away by the nitrogen. The
carbonaceous-type residue remaining is reported as a percent of
the original sample as carbon residue (micro).
The WAT was determined with curves of differential scanning
calorimetry (DSC). The analysis was performed under nitrogen
flow at a flow rate of 100 mL min1. Initially, the crude oil sample
was heated to 80 °C and then the DSC curve was recorded during
cooling at a rate of 1.0 °C min1 to 20 °C.
The BON was determined according to UOP 269 standard
method using an automatic titrator [8]. A perchloric acid solution
0.02 mol L1 was used as the titrant and the solvents used were
toluene and glacial acetic acid. The sample was heated on hot plate
and stirrer to ensure homogeneity of the system. The sample was
added 25 or 50 mL of toluene then added 75 mL of glacial acetic
acid. The resulting mixture was titrated with perchloric acid solu-
tion of 0.02 mol L1.
2.2. 1H NMR analysis
For the 1H NMR spectra acquisition, first samples with API grav-
ity lower than 30 were heated in an oven at 60 °C for approxi-
mately 20 min for greater homogenization and better fluidity.
After heating, the samples were prepared dissolving 5% w/w of
crude oil in dichloromethane-d2 (CD2Cl2). The CD2Cl2 was chosen
as solvent due its good solubility and to present signal at
5.32 ppm, i.e., it is a spectrum region with the lowest contribution
of main signals. Aliphatic and aromatic hydrocarbon from crude oil
samples are responsible for the main contribution for PLS models.
A soft homogenization is carried out so that no adhesion sample on
the bottle wall and consecutive loss, since working with a small
amount of sample, around 20–30 mg. This prepared solution was
transferred to the 5 mm NMR tube and the analysis were carried
out in the Varian spectrometer VRMNS 400 model, operating
with a magnetic field of 9.4 T using a 5 mm Broad Band probe
1
H/19F/X, at 25 °C; Tetramethylsilane (TMS) was used as internal
reference. Another instrumentals parameters were 399.73 MHz,
spectral window of 6410.3 Hz, acquisition time of 2.556 s, waiting
time of 1 s, pulse of 90° (11.7 lseg) and transient numbers equal to
64.
2.3. Chemometric approach
The prediction models were fitted to API gravity, CR, WAT and
BON properties and for each of them were used 106, 90, 92 and
91 samples, respectively. All models were built following the same
procedure. First, dataset was separated into calibration (70%) and
prediction (30%) using the algorithm developed by Kennard and
Stone [30]. Is worth emphasizing that prediction samples were
not consider to the first step which involved only the calibration
set. The preprocessing of data consists in the spectral alignment
using icoshift algorithm [31] followed by data transformation
using the Standard Normal Variation (SNV) to correct variations
in baseline [32]. The appropriate number of Latent Variables
(LVs) to optimize each model was determined using leave-one-
out cross validation method. The models were initially tested with
20 LVs and then optimized.
The Root Mean Square Error of Cross-Validation (RMSECV) and
Prediction were evaluated
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn   samples.
i¼1 yi  y i
b 2
RMSE ¼ ð4Þ
n equally spaced intervals and these aligned individually based on
where yi and ybi correspond to reference and prediction values
respectively and n is the number of samples of the dataset evalu-
ated. However, the easier way to understand the accuracy of the
developed methods is analyzing the RMSEP% because it provides a
direct view of the global percentage error, since there are several
physicochemical properties to analyze from different magnitudes.
The RMSEP% is given by:
RMSEP
RMSEP% ¼ 100 ð5Þ
yprediction
The coefficient of multiple determination for cross-validation
(R2cv) and prediction (R2p) also was evaluated,
P 2
yr  yp
R2 ¼ 1  Pi ð6Þ
 yÞ2
i ðy r
where yr and yp refers to the reference and prediction values
 is the mean vector containing predicted values
respectively, while y
by modeling.
The confidence interval for the values estimated by models was
calculated according to ASTM E1655-12 by equation
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
l ¼ y^i  t ða;glÞ SEC ð1 þ hi Þ ð7Þ
where t is the student’s t value at the 95% confidence interval
according to the degrees of freedom of the model and hi is the
statistic leverage defined by
þ
hi ¼ t 0i ðT 0 TÞ t i ð8Þ
where ti refers to a score vector of particular sample and T to score
matrix of calibration sample set. Combined with residual analysis
the leverage was also used to outlier investigation in the sample set.
Prediction residues were evaluated by statistical tests to sys-
tematic and tendentious errors. The systematic errors able to affect
the estimating always occur in the same way, that is, generates
results above or below the expected. Thus, in order to identifying
this type of error, the bias test was performed [33]. A properly
adjusted model must provide uniform residue as regards the mag-
nitude and are typically distribute around zero. Often, the presence
of trends in residual of calibration models can be identified
through the visual analysis of the graphic of residues against the
reference values. However, does not always visual analysis is
enough reliably to verify the presence or absence of trends. Thus,
a statistical test taking into account the evaluation of the residuals
distribution through non-parametric permutation was carried out
[34].
3. Results and discussions
3.1. Spectra alignment and outlier investigation
Fig. 1A shows clustering of 106 1H NMR spectra relating to dif-
ferent crude oils analyzed. Qualitatively the 1H NMR spectrum for
individual crude oil may be discussed according main regions,
which contains information taking into account principal types of
crude oil hydrocarbons. The hydrogen atoms in different chemical
environments can be interpreted as made by Masili et al. [25] and
Molina et al. [24].
It can be seen that spectra are qualitatively similar. Overlapping
signals across spectral range for all samples was found, except for
two doubtful signals of a particular sample appearing at 3.2 ppm
and other unfolded with two maximum at 2.81 and 2.79 ppm,
region where not was observed similar signals for the other
To perform the alignment, the spectral set was split into 50
 L.M. Duarte et al. / Fuel 181 (2016) 660–669
Fig. 1. 1H NMR spectral set composed of 106 crude oil samples.
an average spectrum used as a reference. After alignment the spec-
tral set can be seen in Fig. 1B and the visual effect alignment in
Fig. 2. As discussed in Section 2.3, the spectral matrix was also
transformed applying SNV preprocessing, resulting in the spectral
set adjusted shown Fig. 1C.
The alignment of the 1H NMR spectra was satisfactory using the
algorithm icoshift. However, it is important to highlight that the
chemical relevance is always more important than statistics-
mathematical considerations. Therefore, it can be noted that after
the alignment (Fig. 1B) the chemical shifts referring to signals of
interest are between 0 and 8 ppm. Hence a thorough assessment
regarding the sample set was carried out to verify the occurrence
of spectral outliers. In this context, principal component analysis
(PCA) was performed.
The first PCA was made in raw spectral set containing 106 sam-
ples (Fig. 3). In this treatment it could be noted a very large scatter-
ing of observations, with the structuring of a trend at the bottom
Fig. 3A (lower score values of PC1 and PC2) and another at the
top (higher score values for PC1 and PC2). However, the sample
number 36 circled stands out for being completely isolated and
distant from other observations. As a boundary condition, align-
ment and SNV transformation was carried out and then a new
modeling by PCA was made (Fig. 3B). Looking at Fig. 3B it is noted
that for smaller PC1 scores there is a greater pool of samples
beyond the percentage of explained variance has increased consid-
erably, as described in Table 1. Nevertheless, even after preprocess-
ing, sample number 36 remained isolated and further away from
the main group, which corroborate positively to the hypothesis
of the occurrence of a spectral outlier. This sample was removed,
a new PCA was performed and then a new spectra set is shown
in Fig. 4, which demonstrates that the sample 36 was sought.
Fig. 2. Visual effect of the alignment of 1H NMR spectra for crude oil samples: raw spectra (left); spectra aligned with icoshift: (right).
663
The doubtful signals disappeared in the spectral set, and the sam-
ple appeared isolated on PCA of Fig. 3 is no longer in Fig. 4.
The removal of this sample is very important because your per-
manence could result in problems in the modeling, decreasing the
predictive capability of the models. A reasonable explanation for
the presence of discrepant signals is the contamination with some
solvent during manipulation of the samples prior to instrumental
analysis for spectral acquisition.
3.2. PLS
To fit models were available 106, 90, 92 and 91 samples for the
properties API gravity, CR, WAT and BON, respectively. As a first
step, the LVs were selected graphically analyzing the profile of
RMSECV vs LVs (Fig. 5A, C, E and G) where optimized number of
LVs for each calibration model was 7, 4, 3 and 7, achieving RMSECV
of 1.4, 0.674% w/w, 5.5 °C and 0.013% w/w and R2cv of 0.978, 0.844,
0.757 and 0.936 for API gravity, CR, WAT and BON models according
to listed in Table 2. The optimization of LVs for CR and WAT models
is not performed by the global minimum of the curve (Fig. 5C and E),
i.e., smaller value of RMSECV, but also considered predictive step
where the test set was used and then evaluated the overall predic-
tive power of the model by RMSEP. Thus, 3 and 4 LVs for WAT and
CR models were selected as optimal by their lower RMSEP values
after use of the test set.
Using samples not involved in the calibration step (test set), it
was possible to assess the predictive capability of the models.
Fig. 5B, D, F and H shows the correlations between the reference
values and the predicted by the models for the test samples, lead-
ing to R2p values of 0.945, 0.802, 0.857 and 0.789 for the properties
API gravity, CR, WAT and BON respectively. The global accuracy of
the fitted models were evaluated through the absolute values of
RMSEP which were 0.8, 0.598% w/w, 3.8 °C and 0.009% w/w, how-
ever, RMSEP% was 3.0%, 14.8%, 11.2% and 8.0%, aiming that the
models better fitted showed better accuracy and error percentage
below 10%. All these parameters for prediction step respectively
correspond to API gravity, CR, WAT and BON models, which may
be screened in Table 2. For all models, the prediction values were
expressed as confidence intervals that contains the reference val-
ues for all test set samples (bar graphs in Fig. 8), where the bars
of reference values are always covered by the bars errors the con-
fidence interval.
Another issue addressed in this work was the post-modeling
analysis, which deals with the evaluation of residues, as discussed
in Section 2.3. The graphic analysis of Fig. 6A–D can lead to a sub-
jective assessment of residues, thus, these were submitted to the
aforementioned statistical tests. According to the bias test, there
is no evidence of systematic error for any model, since the tcalc
obtained was always less than critical, considering the confidence
interval of 95% (see Table 2).
664 L.M. Duarte et al. / Fuel 181 (2016) 660–669

Fig. 3. PCA for the spectral set without alignment (A) and after alignment (B).

Table 1 For the non-parametric permutation test, the linear and quadra-
Percentage of explained variance of PCs. tic functions were used in order to describe the residues of test set
Preprocessing PC1 (%) PC2 (%) PC3 (%) Total (%) for all models. API gravity, WAT and BON models showed no trend
Xraw 83.65 9.44 2.67 95.76
because the p-value was greater than or equal to 0.05, which
Xalign+SNV 95.40 3.51 0.67 99.58 allows inferring that within 95% confidence interval, the residues
Xalign+SNV-outlier 95.58 3.53 0.68 99.79 are randomly distributed around zero. These results can be seen

Fig. 4. 1H NMR spectral set (A) and PCA (B) after removal outlier.
 L.M. Duarte et al. / Fuel 181 (2016) 660–669 665

Fig. 5. LVs optimization to the calibration models for API gravity (A), CR (C), WAT (E) and BON (G) and correlation among reference and prediction values by PLS models for
API gravity (B), CR (D), WAT (F) and BON (H).
666 L.M. Duarte et al. / Fuel 181 (2016) 660–669

Table 2
Main statistical parameters obtained for the models built to API gravity, CR, WAT and BON.

Parameters API CR WAT BON

R2c 0.993 0.915 0.853 0.978
R2cv 0.978 0.844 0.757 0.936
R2p 0.945 0.802 0.857 0.789
RMSECV 1.4 0.674 5.5 0.013
RMSEP 0.8 0.598 3.8 0.009
RMSEP (%) 3.0 14.8 11.2 8.0
LVs 7 4 3 7
Esist tcal = 1.22 < ttab = 2.04 tcal = 0.59 < ttab = 2.05 tcal = 0.32 < ttab = 2.05 tcal = 0.89 < ttab = 2.05
Ealeat (b1) p-value = 0.16 > 0.05 p-value = 0.04 > 0.05 p-value = 0.05 > 0.05 p-value = 0.41 > 0.05
Ealeat (b2) p-value = 0.10 > 0.05 p-value = 0.07 > 0.05 p-value = 0.22 > 0.05 p-value = 0.25 > 0.05

Fig. 6. Residues of prediction models for API gravity (A), CR (B), WAT (C) and BON (D).

in the last two lines of Table 2 and in the histograms of Fig. 7. For
the CR prediction model the residues feature a slight linear trend,
which is confirmed by the p-value is equal to 0.04 (Table 2 and his-
togram of Fig. 7C). This result was mainly affected by two samples
with predicted values of 7.32 and 8.15% w/w respectively, which
can also be seen in the correlation graphic in Fig. 5D and residues
in Fig. 6B. However, it makes no sense to remove them because
although the residual is considerable, the Leverage for both is
below the limit, being 0.1037 and 0.0966 the values respectively,
while the limit is 0.1905.
Considering that the models aiming predict API gravity, CR,
WAT and BON were properly validated, these are ready to be
applied to a new sample set. In order to estimate such physico-
chemical properties for a new crude oil sample is just necessary
obtain their 1H NMR spectrum following the experimental process,
to applied pertinent preprocessing conditions (described in
Section 2.3) and then to use the information of the calibration step
such as loadings and regression coefficient (b).
4. Conclusions
In the present work PLS-1H NMR models to determination of
API gravity, carbon residue, wax appearance temperature and basic
organic nitrogen in Brazilian crude oil samples were successful
applied. The models were validate according to ASTM E 1655-12.
In order to achieve PLS models more robust and presenting higher
predictive capacity, PCA analysis to check outliers presence and
pattern recognize was used. On the other hand, it is important to
highlight that the alignment of the spectra before the modeling
is mandatory to 1H NMR when the chemometric approach is con-
sidered. The average accuracy and coefficient of determination for
 L.M. Duarte et al. / Fuel 181 (2016) 660–669 667

Fig. 7. Permutation histograms for linear (A, C, E and G) and quadratic functions (B, D, F and H) for to try to describe residues of the gravity API (A and B), CR (C and D), WAT (E
and F) and BON (G and H) models.
668 L.M. Duarte et al. / Fuel 181 (2016) 660–669
Fig. 8. Reference (REF) and prediction (PRED) values and confidence intervals for the gravity API, WAT, CR and BON.
prediction obtained to API gravity, carbon residue, wax appearance
temperature and basic organic nitrogen models were 0.8, 0.598%
w/w, 3.8 °C and 0.009% w/w; and 0.945, 0.802, 0.857 and 0.789
respectively. This methodology present as advantage the use of
small amount of sample and can predict four properties from a sin-
gle 1H NMR spectrum. Furthermore, this methodology signals pos-
itively to be used to any type of crude oil with similar
characteristics to Brazilian petroleum. Finally, these methods has
as target assist decision-making, considering to minimize cost
and analysis time, maximizing the optimization of the allocation,
processing and valuation of crude oil.
Acknowledgements
The authors wish to thank the Petrobras, LABPETRO-UFES, UFJF,
Conselho Nacional de Desenvolvimento Científico e Tecnológico
(CNPq-475055/2011-0 and 301689/2011-3) and Coordenação de
Aperfeiçoamento de Pessoal de Nível Superior (CAPES) for fellow-
ships and financial support.
References
[1] Speight JG, Wiley AJ. Handbook of petroleum product analysis; 2002.
[2] Farah MA. Petróleo e seus derivados. LTC; 2012.
[3] ASTMD4052. Standard test method for density, relative density, and API
gravity of liquids by digital density meter. Geneva int organ stand, West
Conshohocken, PA; 2011.
[4] ISO12185. Crude petroleum and petroleum products – determination of
density – oscillating U-tube method. Geneva Int Organ Stand; 1996.
[5] D4530-03 A. Standard test method for determination of carbon residue (Micro
Method) 1. ASTM Int 2003:1–7.
[6] Neto DAA, Gomes EAS, Neto B, Dantas TNC, Moura CPAM. Determination of
wax appearance temperature (Wat) in paraffin/solvent systems by
photoelectric signal and viscosimetry. Braz J Pet Gas 2009;3:149–57.
[7] Li B-Y, Liang Y-Z, Hu Y, Du Y-P, Song Y-Q, Cui H. Evaluation of gas
chromatography–mass spectrometry in conjunction with chemometric
resolution for identification of nitrogen compounds in crude oil. Talanta
2003;61:803–9. http://dx.doi.org/10.1016/S0039-9140(03)00383-7.
[8] 10 U. Nitrogen bases in hydrocarbons by titration. West Conshohocken, PA,
USA Am Soc Test Mater; 2010.
[9] Ruiz MD, Bustamante IT, Dago A, Hernández N, Núñez AC, Porro D. A
multivariate calibration approach for determination of petroleum
hydrocarbons in water by means of IR spectroscopy. J Chemom
2010;24:444–7. http://dx.doi.org/10.1002/cem.1284.
[10] Khanmohammadi M, Garmarudi AB, Garmarudi AB, De la Guardia M.
Characterization of petroleum-based products by infrared spectroscopy and
chemometrics. TrAC – Trends Anal Chem 2012;35:135–49. http://dx.doi.org/
10.1016/j.trac.2011.12.006.
[11] Müller ALH, Picoloto RS, Mello PDA, Ferrão MF, Dos Santos MDFP, Guimarães
RCL, et al. Total sulfur determination in residues of crude oil distillation using
FT-IR/ATR and variable selection methods. Spectrochim Acta – Part A Mol
Biomol Spectrosc 2012;89:82–7. http://dx.doi.org/10.1016/j.saa.2011.12.001.
[12] Filgueiras PR, Sad CMS, Loureiro AR, Santos MFP, Castro EVR, Dias JCM, et al.
Determination of API gravity, kinematic viscosity and water content in
petroleum by ATR-FTIR spectroscopy and multivariate calibration. Fuel
2014;116:123–30. http://dx.doi.org/10.1016/j.fuel.2013.07.122.
[13] Jingyan L, Xiaoli C, Songbai T, Wanzhen L. Research on determination of
nitrogen content in petroleum using mid-infrared spectroscopy. China Pet
Process Petrochem Technol 2011;13:1–7.
[14] Jingyan L, Xiaoli C, Songbai T. Research on determination of total acid number
of petroleum using mid-infrared attenuated total reflection spectroscopy.
Energy Fuels 2012;26:5633–7. http://dx.doi.org/10.1021/ef3002372.
[15] Laxalde J, Ruckebusch C, Devos O, Caillol N, Wahl F, Duponchel L.
Characterisation of heavy oils using near-infrared spectroscopy: optimisation
of pre-processing methods and variable selection. Anal Chim Acta
2011;705:227–34. http://dx.doi.org/10.1016/j.aca.2011.05.048.
[16] Balabin RM, Smirnov SV. Variable selection in near-infrared spectroscopy:
benchmarking of feature selection methods on biodiesel data. Anal Chim Acta
2011;692:63–72. http://dx.doi.org/10.1016/j.aca.2011.03.006.
[17] Abbas O, Rebufa C, Dupuy N, Permanyer A, Kister J. PLS regression on
spectroscopic data for the prediction of crude oil quality: API gravity and
aliphatic/aromatic ratio. Fuel 2012;98:5–14. http://dx.doi.org/10.1016/
j.fuel.2012.03.045.
[18] Alam TM, Alam MK. Chemometric analysis of NMR spectroscopy data: a
review. Annu Reports NMR Spectrosc 2004;54:41–80. http://dx.doi.org/
10.1016/S0066-4103(04)54002-4.
[19] Winning H, Larsen FH, Bro R, Engelsen SB. Quantitative analysis of NMR
spectra with chemometrics. J Magn Reson 2008;190:26–32. http://dx.doi.org/
10.1016/j.jmr.2007.10.005.
 L.M. Duarte et al. / Fuel 181 (2016) 660–669
[20] Silva SL, Silva AMS, Ribeiro JC, Martins FG, Da Silva FA, Silva CM.
Chromatographic and spectroscopic analysis of heavy crude oil mixtures
with emphasis in nuclear magnetic resonance spectroscopy: a review. Anal
Chim Acta 2011;707:18–37. http://dx.doi.org/10.1016/j.aca.2011.09.010.
[21] Filgueiras PR, Portela N de A, Silva SRC, Castro EVR, Oliveira LMSL, Dias JCM,
et al. Determination of saturates, aromatics and polars in crude oil by 13C NMR
and support vector regression with variable selection by genetic algorithm.
Energy Fuels. http://dx.doi.org/10.1021/acs.energyfuels.5b02377. 2016:acs.
energyfuels.5b02377
[22] Barbosa LL, Sad CMS, Morgan VG, Figueiras PR, Castro ERV. Application of low
field NMR as an alternative technique to quantification of total acid number
and sulphur content in petroleum from Brazilian reservoirs. Fuel
2016;176:146–52. http://dx.doi.org/10.1016/j.fuel.2016.02.085.
[23] Molina VD, Uribe UN, Murgich J. Partial least-squares (PLS) correlation
between refined product yields and physicochemical properties with the 1H
nuclear magnetic resonance (NMR) Spectra of Colombian Crude Oils
2007:1674–80.
[24] Molina VD, Uribe UN, Murgich J. Correlations between SARA fractions and
physicochemical properties with 1H NMR spectra of vacuum residues from
Colombian crude oils. Fuel 2010;89:185–92. http://dx.doi.org/10.1016/
j.fuel.2009.07.021.
[25] Masili A, Puligheddu S, Sassu L, Scano P, Lai A. Prediction of physical–chemical
properties of crude oils by 1H NMR analysis of neat samples and
chemometrics. Magn Reson Chem 2012;50:729–38. http://dx.doi.org/
10.1002/mrc.3872.
View publication stats
669
[26] de Peinder P, Visser T, Petrauskas DD, Salvatori F, Soulimani F, Weckhuysen
BM. Partial least squares modeling of combined infrared, 1H NMR and 13C NMR
spectra to predict long residue properties of crude oils. Vib Spectrosc
2009;51:205–12. http://dx.doi.org/10.1016/j.vibspec.2009.04.009.
[27] Filgueiras PR, Terra LA, Castro EVR, Oliveira LMSL, Dias JCM, Poppi RJ.
Prediction of the distillation temperatures of crude oils using 1H NMR and
support vector regression with estimated confidence intervals. Talanta
2015;142:197–205. http://dx.doi.org/10.1016/j.talanta.2015.04.046.
[28] Geladi P, Kowalski BR. Partial least-squares regression: a tutorial. Anal Chim
Acta 1986;185:1–17.
[29] Wold S, Sjostrom M, Eriksson L. PLS-regression: a basic tool of chemometrics.
Chemom Intell Lab Syst 2001;58:109–30.
[30] Kennard RW, Stone LA. Computer aided design of experiments. Technometrics
1969;11:137–48.
[31] Savorani F, Tomasi G, Engelsen SB. Icoshift: a versatile tool for the rapid
alignment of 1D NMR spectra. J Magn Reson 2010;202:190–202. http://dx.doi.
org/10.1016/j.jmr.2009.11.012.
[32] Barnes RJ, Dhanoa MS, Lister SJ. Standard normal variate transformation and
de-trending of near-infrared diffuse reflectance spectra. Appl Spectrosc
1989;43:772–7.
[33] ASTM E-05. Standard Practices for Infrared Multivariate Quantitative Analysis
1 2012:1–29.
[34] Filgueiras PR, Alves JCL, Sad CMS, Castro EVR, Dias JCM, Poppi RJ. Evaluation of
trends in residuals of multivariate calibration models by permutation test.
Chemom Intell Lab Syst 2014;133:33–41. http://dx.doi.org/10.1016/j.
chemolab.2014.02.002.