Analytica Chimica Acta 571 (2006) 167174

Combination of mid- and near-infrared spectroscopy for the

determination of the quality properties of beers
Fernando A. Ino n a,1 , Salvador Garrigues b , Miguel de la Guardia b,
b

a JENCK S.A., Av. Alvarez

Thomas, 228 Buenos Aires C1427CCP

Departamento de Qumica Analtica. Universidad de Valencia, Edicio Jer`onim Munoz C/Dr. Moliner, 50, 46100 Burjassot, Valencia, Spain

Received 24 December 2005; received in revised form 21 April 2006; accepted 26 April 2006
Available online 5 May 2006

Abstract
The combination of infrared (MIR) and near-infrared (NIR) spectroscopy has been employed for the determination of important quality parameters
of beers, such as original and real extract and alcohol content. A population of 43 samples obtained from the Spanish market and including different
types of beer, was evaluated. For each technique, spectra were obtained in triplicate. In the case of NIR a 1 mm pathlength quartz flow cell
was used, whereas attenuated total reflectance measurements were used in MIR. Cluster hierarchical analysis was employed to select calibration
and validation data sets. The calibration set was composed of 15 samples, thus leaving 28 for validation. A critical evaluation of the prediction
capability of multivariate methods established from the combination of NIR and MIR spectra was made. Partial least squares (PLS) and artificial
neural networks (ANN) were evaluated for the treatment of data obtained in each individual technique and the combination of both. Different
parameters of each methodology were optimized. A slightly better predictive performance was obtained for NIRMIR combined spectra, and in all
the cases ANN performs better than PLS, which may be interpreted from the existence of some non-linearity in the data. The root-mean-sqare-error
of prediction (RMSEP) values obtained for the combined NIRMIR spectra for the determination of real extract, original extract and ethanol were
0.076% w/w, 0.14% w/w and 0.091% v/v.
2006 Elsevier B.V. All rights reserved.
Keywords: Determination; Original and real extract; Alcohol content; Beers; Artificial neuronal networks; Partial least squares; Combination of spectroscopic
techniques; Near-infrared; Middle-infrared; Attenuated total reflectance

1. Introduction
Beer is one of the two most common alcoholic beverage
obtained by fermentation of cereals germinated in water in the
presence of yeast [1]. The main components of beers are water,
carbohydrates and ethanol, which are commonly used in the
brewing industry for quality control of the final product.
Original and real extract correspond to the amount of sugars
in the beers before and after producing the fermentation process,
respectively and are normally expressed in % w/w. Ethanol content is a key economic and organoleptic parameter affecting both,
the beer classification (in term of taxes) and its taste [1,2]. It is
usually expressed in % v/v.

Corresponding author. Tel.: +34 96 354 4838; fax: +34 96 354 4838.
E-mail address: miguel.delaguardia@uv.es (M. de la Guardia).
1 Present address: Laboratorio de An
alisis de Trazas, Departamento de
Qumica Inorganica, Analtica y Qumica Fsica, Universidad de Buenos Aires,
Pabellon 2, Ciudad Universitaria, (1428) Buenos Aires, Argentina.
0003-2670/$ see front matter 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.aca.2006.04.070

The official methods of the Analytical division of European

Brewery Convention for the determination of the aforementioned parameters are based on the distillation of beer and the
measurement of the distillate and remaining solution density
[3]. The density values of both solutions are introduced in semiempirical tables for obtaining the percentage of extracts and the
ethanol content. These methods involve a high sample handling
and are time consuming and therefore inefficient in terns of time
and cost. So, there is a need for developing fast alternative methods for routine quality control programs. Among those, the most
common strategy implemented in the industry is the use of autoanalyzers, which reduces the sample handling, but requires about
30 min for obtaining a triplicate measurement of each parameter.
In two previous contributions we have reviewed the determination of real extract, original extract and ethanol in beers by
means of infrared (MIR) [4] and near-infrared (NIR) [5] spectroscopy. In these works, we have also presented the methods
we have developed for the simultaneous determination of these
important parameters, as until that time most work conducted in

168

F.A. Ino n et al. / Analytica Chimica Acta 571 (2006) 167174

this area only addressed the determination of only one parameter

at time.
One question that remains unanswered is if the combination
of the spectra acquired by NIR and MIR may improve the prediction capability of the multivariate models. There are many works
comparing one technique versus other, some combining both for
characterization or identification purposes, including those using
two-dimensional correlation analysis (i.e., [611]). Combined
NIR and MIR spectral data has been used together with multivariate curve resolution (MCR) for monitoring temperaturedependent transitions of proteins [12,13]). Also, Raman and
MIR have been used in combination for prediction of yarn
properties [14]. The authors concluded that concatenating the
Raman and infrared spectra does not enhance the PLS prediction performance, not even after wavelength selection. Apart
from these contributions no works have been found devoted
to the use of multitechnique combination for quantitative
purposes.
Regarding the chemometric treatment, to the best of our
knowledge no attempts to use neural networks for the simultaneous determination of quality properties of beer have been
conducted so far. Work published in the literature related to
neural networks and beers [1519] address the optimization or
prediction of industrial process rather quality control of the final
product or the characterization of beer samples according to their
mineral content by inductively coupled plasma atomic emission spectrometry (ICP-AES) [20]. No works has been found in
which the neural network was coupled to any vibrational spectroscopy for the determination of the aforementioned quality
properties. Therefore, this study was devoted to the evaluation
of a methodology based upon NIR and MIR measurements and
chemometric data treatment for the evaluation of beer parameters.
For increasing the application range of the multivariate
model, a heterogeneous sample population was chosen for
selecting the calibration and the validation datasets. The selection process of each sample set was carried out as reported
elsewhere [4,5].
As the combination of data from different techniques may
include the addition of non-linearity to the system we decided
to evaluate both, partial-least squares (PLS)-NIRMIR and artificial neural networks (ANN)-NIRMIR methodologies, for the
prediction of ethanol, real and original extracts. Models were
compared in terms of the root-mean-square-error of prediction
(RMSEP).
2. Experimental
For NIR measurement, it was used a Fourier transform NIR
Brucker MPA spectrometer controlled by Opus for Windows
software from Brucker Optik (Bremen, Germany), and with a
1.00 mm pathlength (50 l volume) quartz flow cell from Hellma
(Mullheim, Germany) mounted on a home-made adaptor.
For MIR measurement, it was employed a Nicolet Magna
Series 750 model FTIR spectrometer controlled by Omnic
for Windows software from Nicolet Instrument (Madison, WI,
USA) and equipped with Specaclamp IN-Compartment Contact

Sampler horizontal ATR from Graseby Specac (Orpington, UK)

with a 45 crystal ZnSe through top-plate.
Room temperature was monitored using a mercury thermometer with a precision of 0.5 C and it did not vary significantly during acquisition of the spectra of all the samples.
For sample preparation, a magnetic stirrer Select, (Singapur)
and a thermostatic bath Grant (Cambridge, UK) were used. A
Gilson Minipuls 2 from Gilson (Villiers-le-Bel, France) peristaltic pump was used to fill the flow cell through 0.5 mm i.d.
Teflon tubing.
3. Samples
A total of 43 beer samples, contained in sealed aluminium
cans, were obtained from the Spanish market. Duplicate cans of
each batch were always collected, one was used for NIR and MIR
measurements and the other was used to measure the properties
of interest by reference procedures. The characteristics of the
samples are indicated in Table 1, and a detailed description can
be found in [5].
Samples were degassed by stirring and then filtered before
filling the ATR cell or pumped through the flow cell of the NIR
instrument.
4. MIR analysis
Original samples were placed in the same temperature controlled room where the spectrometer was located before performing the analysis. Three sub samples were poured into the
ATR cell and the FTIR spectra were taken as follows. Sample
spectra were scanned between 4000 and 600 cm1 , by averaging 25 scans per spectrum with a nominal resolution of
4 cm1 (data spacing of 1.93 cm1 ) and using a mirror velocity
of 0.6329 cm s1 . The acquisition of each averaged spectrum
requires 35 s [4].
The background and blank spectra were acquired by filling
the ATR plate cell with Millipore Q-purified water (Bedford,
MA, USA) and using the same instrumental conditions than
those employed for samples. Background spectra were scanned
at intervals of seven samples and blank spectra were collected
after measurement of each sample, to ensure that memory effect
of IR spectra was negligible after cleaning the ATR crystal.
Both, sample and blank spectra were collected in absorbance
mode. The regions between 4000 and 3050 cm1 and between
850 and 600 cm1 were eliminated prior the calculations as it
was observed that variations in these regions cannot be ascribed
to variations in sample composition. Furthermore, the absorption
band around 2382 and 2314 cm1 , which is due to atmospheric
CO2 , was also cut off from raw spectra.
The spectra of the three sub samples of each beer were taken
by refilling the ATR cell once again.
5. NIR analysis
Samples were placed in the same temperature controlled
room where the spectrometer was located before to carry out
the analysis. During the acquisition of triplicate FT-NIR spectra

F.A. Ino n et al. / Analytica Chimica Acta 571 (2006) 167174

169

Table 1
Composition of samples employed in this study
Sample number

Classification

Original extract (% w/w)

Real extract (% w/w)

Alcohol (% v/v)

Alcohol (% w/w)

1
2
3
4
5a
6
7a
8a
9
10
11
12a
13
14
15
16
17
18a
19a
20
21
22a
23
24
25a
26a
27
28
29
30a
31
32a
33
34
35a
36
37a
38
39a
40
41a
42
43

100% malt
Especial
Normal
Normal
Normal
100% malt
Normal
Especial
Normal
Normal
With soda
Normal
Normal
German type
Normal
Normal
Normal
Alcohol free
With soda
Normal
Normal
Normal
Normal
Normal
100% malt
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal
Normal

12.12
12.95
10.32
10.6
11.35
11.51
10.7
13
10.9
10.33
8.72
10.46
10.32
12.47
10.49
10.41
10.78
4.25
9.11
10.44
10.4
10.6
10.82
10.7
12.07
10.47
10.84
10.54
10.52
10.67
10.77
10.34
10.7
10.57
10.69
11.09
10.51
10.89
10.73
10.64
10.81
10.87
11

4.01
4.57
3.44
3.54
3.63
3.76
3.63
4.81
3.7
3.47
2.9
3.45
3.45
4.15
3.53
3.41
3.58
2.17
3.02
3.43
3.5
3.57
3.52
3.4
4.45
3.75
3.53
3.48
3.64
3.44
3.62
3.6
3.51
3.36
3.42
3.78
3.47
3.76
3.72
3.63
3.72
3.8
3.96

5.34
5.55
4.4
4.6
5.03
5.06
4.62
5.43
4.76
4.47
3.71
4.53
4.47
5.47
4.53
4.55
4.7
1.31
3.94
4.57
4.5
4.67
4.76
4.76
5.53
4.38
4.77
4.59
4.49
4.71
4.66
4.39
4.74
4.67
4.74
4.81
4.42
4.79
4.73
4.62
4.74
4.83
4.62

4.19
4.35
3.52
3.62
3.95
3.99
3.62
4.23
3.74
3.51
2.92
3.59
3.51
4.29
3.56
3.57
3.69
2.17
3.11
3.59
3.55
3.64
3.74
3.74
4.34
3.44
3.75
3.61
3.53
3.71
3.66
3.44
3.73
3.66
3.72
3.78
3.47
3.76
3.72
3.63
3.72
3.8
3.62

Samples employed for calibration in all cases.

using the flow cell, the sample was continuously pumped through
the cell using a peristaltic pump (flow rate
=1.5 mL min1 ) and
refilled after each measurement.
Sample spectra were scanned between 800 and 2857 nm, by
averaging 25 scans per spectrum with a nominal resolution of
2 cm1 . The acquisition of each averaged spectrum required
35 s. The maximum standard deviation of the response (sR ) was
estimated to be 0.0005 a.u.
The background and blank spectra were acquired by filling
the cell with Millipore Q-purified water (Bedford, MA, USA)
and using the same instrumental conditions than those employed
for samples. Background spectra were scanned at intervals of
seven samples and blank spectra were acquired after measurement of each sample, to avoid cross-contamination.

Both, sample and blank, spectra were collected in absorbance

mode. The available spectral-analytical windows were from 800
to 1850 nm and from 2050 to 2400 nm. The spectral region from
1850 to 2100 nm was not available due to the high absorbance
of water in this region.
6. Data analysis
Spectra collected from the Opus and Omnic software were
exported in text format and analyzed using The Mathworks Inc.
(South Natick, MA, USA). Firstly, laboratory-written Matlab
functions were used for hierarchical cluster analysis to evaluate the similarity of samples in terms of their NIR spectra and to
assess the number of characteristic subsets in which the available

170

F.A. Ino n et al. / Analytica Chimica Acta 571 (2006) 167174

samples could be divided. Multivariate calibration calculations

were made, for PLS, with the MVC1 toolbox [21] and, for ANN,
through a home-made graphical user interface (GUI) for maximizing the potential of the neural network calibration routines
of the ChemoAC Toolbox [22].
The following figures for the models fit to the data and the
predictive power were used throughout the text. In all cases, the
scope was to evaluate the average deviation of the model from
the actual data.
PRESS, as the sum of squares prediction error, root-meansquare error of calibration (RMSEC), and root-mean-square
error of cross-validation (RMSECV) were used to evaluate the
coherence of the calibration models and to determine the prediction capabilities, of each model it was employed, the rootmean-square-error of prediction (RMSEP) established when the
model was applied to new data for which the reference values
are available.
In order to compare the spectroscopy methodology against
the reference ones, different quality indicators were also given:
(i) the absolute mean difference (dxy ) between predicted values
and reference data, (ii) the standard deviation of mean differences (sxy ), the quality coefficient (QC) and (iii) the pooled
standard error of prediction for validation samples (sreg ). As
stated by Massart et al., the QC is to be preferred over correlation
coefficient of the regression between predicted and reference
data not only because it gives a better idea of the spread of
the data points around the fitted straight line but also because it
gives some indication on the percentage error to be expected for
the estimated concentration [23].
To build and select PLS models, the selection of the optimum
number of factors, which minimizes the RMSECV, was based
on the criterion of Haaland and Thomas [24].
In the case of ANN, there are many factors to be considered
for building calibration models. These are (i) the requirement
or nor of a previous principal component analysis (PCA) for
reducing the number of variables, and in this case, the number
of extracted component; (ii) the topology of the network (number of input and output nodes), (iii) type of transfer function.
For each analyte, an extensive evaluation of the aforementioned
parameters was made. Regarding transfer function, only hyperbolic tangent and linear function were considered. In all the
cases, the number of hidden layers was fixed to one.
7. Results and discussion
7.1. Combined NIRMIR spectra of beer
Before combining spectra acquired by NIR and MIR, a series
of corrections were carried out. For correcting additive artefacts,
the average of absorbance values between 2007 and 2056 cm1
(for ATRMIR) and between 2220 and 2222 nm (for NIR) of
each individual spectrum was respectively subtracted. These corrected spectra were utilized for further calculations. In case of
ATRMIR, as the penetration depth is inversely proportional to
the wavenumber, the absorbance at a given wavenumber was
further corrected by multiplying the absorbance value by the
wavenumber value and rationing the result by the maximum

Fig. 1. Combined NIRMIR spectra of main types of beer samples. (a) 100%
malt beer; (b) normal beer and (c) beer without alcohol. Wavelengths and
wavenumbers were plotted in the same axis for clarity purposes.

wavenumber in the spectrum. After that, the mean spectrum

of each sample was calculated from acquired replicates. Augmented matrix were built allocating the MIR spectra at the right
of NIR ones.
A complete description of main spectral features in the IR can
be found elsewhere [4,5]. In the present study, the main bands
of interest are located in the range 22202354 nm for NIR and
between 850 and 1201 cm1 for MIR.
The aforementioned regions were selected due to two main
reasons: (i) to keep low the total number of wavelengths and
wavenumbers and (ii) because in our previous contributions the
best multivariate models were built from these spectral regions.
Thus augmented matrix was composed by 43 rows (samples)
and 316 columns (182 for NIR and 134 for MIR).
The hydroxyl groups (OH) of sugars absorb in MIR around
1052 cm1 corresponding to carbohydrates. NIR spectra of
beers are mainly related to the absorption of water in this spectral
region with some features of the other two main constituents:
ethanol and carbohydrates. As background signal was acquired
using water, the beer spectra shown are due to the absorption
of some beer constituent plus the reduction of the amount of
water. In the case of the formers, positive bands were observed,
whereas in the later case, negative peaks were found. From the
literature, the following general assignments can be made in the
region of interest: the bands located between 2200 and 2400 nm
are due to the first set of CH combination bands, and for OH, the
combination bands occur between 2050 and 2200 nm, region in
which a broad band is observed [25,26].
Fig. 1 shows the combined spectra of three different types
of beers. The 100% malt is the sample with the highest extract
value, and it can be seen that the band in the sugar region is
the highest one. On the other hand, the sample without alcohol
has the lowest extract value, and accordingly, it has the lowest
absorption in the aforementioned spectral region. In the selected
MIR region the main differences between the spectrum of the
beer without alcohol and that of a normal beer correspond to
the bands located at 1100, 1050 and 875 cm1 . These bands are

F.A. Ino n et al. / Analytica Chimica Acta 571 (2006) 167174

ascribed to the presence of ethanol. Similar differences can be

ascribed in the NIR spectra.

Table 2
Prediction capabilities of PLS-NIRMIR for real and original extract and ethanol
determination

7.2. Selection of the calibration set

Calibration and validation sets were based on the results
obtained from hierarchical cluster analysis [5]. In brief, a clustering method was used before chemometric data treatment in
order to evaluate possible classes among samples considered
and to select properly a representative calibration. For calculating spectral distance, Euclidian distance of the scores of the most
significant principal component (PC) was used (PC selected after
a principal component analysis (PCA) applied to NIR data in
the region from 2220 to 2370 nm). Ward method was used at
the linkage stage. The selection of the calibration and validation
datasets was made using the obtained dendrogram, selecting at
least one sample of each cluster for calibration. If the cluster
is comprised of more than one sample, the number of samples
selected for calibration was approximately the root square of the
total number of samples included in the cluster; the remaining
samples were integrated in the validation data set. So, the number of samples assigned to the validation was equal or higher
than the number of those employed for calibration. The samples
within a given cluster were selected randomly.
Following the aforementioned rules, calibration and validation sets comprised of 15 and 28 samples, respectively.
It must be noticed that a calibration set selection based on
a previous identification of the different class of samples, from
their spectra, offers guaranties about the consideration of samples of the some type than those to be analysed in the model
construction.
7.3. Determination of real extract, original extract and
ethanol using PLS and combined spectra
A calibration model was built in terms of combined NIR and
MIR data, considering the range of wavenumbers and wavelengths depicted in Fig. 1. No efforts were conducted to reduce
the wavelength/wavenumbers interval as their were already optimized for each individual spectroscopy.
Table 2 shows the calibration and validation results obtained
for the three properties analysed using the combined spectra. For
the real extract determination the optimum PLS method was
based on four extracted factors, obtaining a RMSEP value of
0.095% w/w and a QC value of 2.6%. In the case of original
extract, the optimum number of factors was 3, yielding a RMSEP
and QC values of 0.192% w/w and 1.8%, respectively. Finally,
for ethanol 4 factors were extracted, obtaining a RMSEP value
of 0.103% v/v and a QC value of 2.7%.
The aforementioned reported values can be compared to those
values obtained using only NIR or MIR data [4,5]. For MIR,
RMSEP and QC values for (i) real extract, (ii) original extract
and (iii) ethanol was: (i) 0.103% w/w and 2.9%, (ii) 0.2% w/w
and 1.9%, (iii) 0.12% v/v and 2.6%, respectively. In the case
of NIR, these values were: (i) 0.12% w/w and 3.0%, (ii) 0.18%
w/w and 1.5%, (iii) 0.10% v/v and 1.9%, respectively. As it can
be seen, results obtained when using the combined spectra tend

171

Factors
RMSECVa
RMSEPa
RRMSEP
d(xy) a
s(xy) a
QC
sc a
sR
sreg a

Real extract

Original extract

Ethanol

4
0.208
0.095
2.6%
0.029
0.092
2.60%
0.05
0.0005
0.04

3
0.211
0.192
1.8%
0.026
0.194
1.81%
0.05
0.0005
0.03

4
0.103
0.121
2.6%
0.039
0.117
2.69%
0.05
0.0005
0.03

Note: Spectral range used was that indicated in Fig. 1. sreg is the standard error
of prediction.
a Values given in % w/w for real and original extract and % v/v for ethanol.
For additional details see the text.

to be slightly better (or similar) than those obtained on using a

single technique.
Fig. 2 shows the net sensitivity vector associated to the PLS
method for real extract, original extract and ethanol determination. It can be appreciated that there is no direct match between
all vectors; all parameters are not directly associated to the same
NIRMIR spectral feature. Moreover, in the case of ethanol,
the model tends to be more focused in the MIR spectra that
on the NIR, whereas for the other properties the contribution
of MIR and NIR spectra to the net analyte signal seems to be
balanced.
Table 2 includes the figures of merit obtained under all aforementioned conditions. The standard error of prediction (that
includes the uncertainty from the model [27,28], considering the
standard deviation of reference concentration (sc ) and standard
deviation of the response (sR ) were 0.04% w/w 0.03% w/w, and
0.03% v/v, for original extract, real extract and ethanol respectively.

Fig. 2. Real and original extract and ethanol for net analyte sensitivity (sk )
vector for combined NIRMIR PLS model. Wavelengths and wavenumbers
were plotted in the same axis for clarity purposes.

172

F.A. Ino n et al. / Analytica Chimica Acta 571 (2006) 167174

7.4. Determination of real extract, original extract and

ethanol using neural networks
The neural network chosen for this work is a double-layer
feed-forward type with the error back-propagation learning rule.
An excellent review addressing all details of this ANN can be
found elsewhere [22]. The basis of ANN optimization is to select
the bias and weight values acting in the selected transfer-function
of each neuron for minimising the RMSEP values of a monitoring dataset.
The output nodes were fixed to one, and the optimization
of a ANN for each property has been made. Due to the high
number of sensors (wavelengths and wavenumbers) it is almost
impossible to construct and optimize the ANN from the original variables. As a matter of fact, evaluate the number of
neurons from 1 to 10 in the hidden-layer for selected MIR
range (8501200 cm1 ) took 27 h in a Pentium IV 3.0 Gz HT
computer. Therefore a compression technique, in particular a
pre-treatment based on PCA, has been adopted in order to reduce
the number of variables entering into the ANN.
As PCA is a linear compression method and ANN is preferably employed for non-linear systems, it is important to construct
a surface response curve including also high PCs when building an ANN from scores of extracted principal components
(PC). Optionally, the original data can be autoscaled before
performing PCA. We have tested the use of PCA with and
without autoscaled spectra, and in all the cases better predictive results have been achieved using raw spectra. It is important
to remember that it is not necessary to mean-center input variables before training since the biases act as offsets in the model.
ANN training is not based on variancecovariance maximization, and therefore it is not necessary to scale the different
variables to unit variance. The only constraint for ANNs is to
scale each input variable so that training starts within the active
range of the non-linear transfer functions. In this work, samples were range-scaled with a linear mapping called minmax
scaling [22], determined for the training set and applied to all
samples.
The response surface was explored up to 12 first extracted
principal components (PCs) from calibration dataset. Results
shown that just nine were needed for the model which requires
the highest number of PCs (see below). Regarding the explained
variance, the fist component explain the 99.86% of the variance,
the next one explain an additional 0.11%, while the contribution
of the remaining ones is bellow the 0.01%.
The optimum number of PCs and neurons for (i) real extract,
(ii) original extract and (iii) ethanol were: (i) 2 and 2, (ii) 9
and 4 and (iii) 5 and 7. In the latter case, similar RMSEP values
(0.090% v/v against 0.089% v/v) has been found with 3 PCs and
2 neurons. So, this bounder condition were preferred to reduce
de dimension of the model.
In each case the behaviour of non-lineal neurons has been
cheeked with the idea of replace as much non-linear neurons by
linear ones. Doing so, linear and non-linear behaviours of data
can be catched by specific neurons. This evaluation is carried
out plotting hidden nodes activation of calibration or validation
data at the end of training [22]. The activation of hidden nodes

of the data may indicate that the data set is mainly linear when
data are activated in their linear portion or indicates a non-linear
behaviour when data are activated in their strongly non-linear
portion. Thus, it is possible to obtain information on the degree
of non-linearity of a given data set.
Fig. 3 shows for real extract and ethanol the activation of
the nodes by validation samples. It can be seen that in three
(of four) of the activation nodes the activation occurs in the
linear portion, whereas in the case of ethanol it occurs in the
non-linear portion. Therefore, for each analyte the number of
linear and non-linear neurons has been evaluated in terms of
the RMSEP value. Table 3 summarizes different characteristics
of the optimum ANN model built for all the properties. For
determination of real extract the optimum ANN method was
based on two extracted factors and two neurons (one linear and
one non-linear). For original extract, nine factors were kept in
the optimized model and the topology of the network consisted
in three linear and one non-linear neuron in the hidden layer.
For ethanol determination, on the other hand, the model was
based on two non-linear neurons in the hidden layer and three
extracted factors.
Table 3 also includes the figures of merit obtained under
all the aforementioned conditions. Prediction capabilities using
ANN model for all the properties are better than those presented
in Table 2. For example, RMSEP values are 2.1%, 1.3% and
1.9% for original extract, real extract, and ethanol respectively.
The maximum percentage error to be expected for new determinations (QC) of these properties were 2.1%, 1.4% and 2.0%,
respectively, which are better than those found for the PLS treatment.
One question which remains open is how to estimate the
precision of the method from available data. The combined
spectrum of each sample was obtained by averaging triplicate
spectra of each technique. Therefore, the usual procedure for
obtaining triplicate combined spectra should be, in this case,
to acquire three times three spectra in each technique and then
combine them. An alternative approach could be to perform
the full combination individual spectrum, obtaining thus nine
combined spectra, and to predict the concentration from each
combined spectra using the ANN method developed. Following
this last procedure, the repeatability of the determination, estabTable 3
Prediction capabilities of ANN-NIRMIR for real and original extract and
ethanol determination
Dataset

Real extract

Original extract

Ethanol

Factors
Neurons in hidden layer
RMSECa
RMSEPa
RRMSEP
d(xy) a
s(xy) a
QC

2
2 (1L, 1H)
0.141
0.076
2.1%
0.007
0.077
2.12%

9
4 (1H, 3L)
0.059
0.145
1.3%
0.014
0.147
1.39%

3
2 (2H)
0.092
0.091
1.9%
0.004
0.092
2.01%

Note: Factors are the extracted principal components after performing PCA. L
and H mean linear and hyperbolic tangent transfer function.
a Values given in % w/w for real and original extract and % v/v for ethanol.
For additional details see the text.

F.A. Ino n et al. / Analytica Chimica Acta 571 (2006) 167174

173

Fig. 3. Activation nodes for (A) real extract and (B) ethanol when considering only hyperbolic tangent transfer functions.

lished from the pooled standard deviation of the nine predicted

concentrations, were, 0.01% w/w, 0.02% w/w, and 0.01% w/w,
for original extract, real extract and ethanol respectively. As can
be appreciated, the precision is of the order of the uncertainty of
reference values. The values calculated for QC in these conditions are almost equal to those obtained when individual spectra
are averaged before being combined.

combining or not the techniques, it can be argued that although

prediction errors are slightly lower in the combined method, the
improvement is not significant. In this aspect we are in agreement
with de Groot et al. [14]. The use of ANN instead of PLS can
be justified because the activation of the nodes lies in non-linear
portions.
Acknowledgments

8. Conclusions
The determination of real extract, original extract and ethanol
in beers can be successfully carried out through the combination
of MIR and NIR techniques. Calibration model can be built
using PLS or ANN, although better predictive capabilities are
obtained with the latter method. Regarding the advantages of

The financial support of the Direccio General dUniversitats

i Investigacio de la Generalitat Valenciana (Project GV04B/247,
Grupos 03-118 and invited professor FAI grant), Universitat
de Val`encia (Project UV-AE-20050203), University of Buenos
Aires (X013) and ANPCyT (PICT-2003 17932)) is acknowledge.

174

F.A. Ino n et al. / Analytica Chimica Acta 571 (2006) 167174

References
[1] A.H. Varnam, J.P. Sutherland, Bebidas: Tecnologa, Qumica y Microbiologa, Editorial Acribia, S.A., Zaragoza, 1996.
[2] Council Directive 92/84/EEC: OJ L 316, 1992. 31.10.1992 Bull. 101992.
[3] European Brewery Convention. Analytica EBC, 4th ed., Brauerei und
Gertranke Rundschau, Zurich, 1987.
[4] F.A. Ino n, R. Llario, S. Garrigues, M. de la Guardia, Anal. Bioanal.
Chem. 382 (2005) 1549.
[5] R. Llario, F.A. Ino n, S. Garrigues, M. de la Guardia, Talanta 69 (2006)
469.
[6] N. Afonina, L. Bhattacharyya, J.P. Hennessey, Analyst 129 (2004) 1091.
[7] M.S. Lin, M. Al-Holy, H. Al-Qadiri, D.H. Kang, A.G. Cavinato, Y.Q.
Huang, B.A. Rasco, J. Agric. Food Chem. 52 (2004) 5769.
[8] L.E. Rodriguez-Saona, F.M. Khambaty, F.S. Fry, E.M. Calvey, J. Agric.
Food Chem. 49 (2001) 574.
[9] Y.L. Liu, R.K. Cho, K. Sakurai, T. Miura, Y. Ozaki, Appl. Spectrosc.
48 (1994) 1249.
[10] F.E. Barton, D.S. Himmelsbach, Appl. Spectrosc. 47 (1993) 1920.
[11] F.E. Barton, D.S. Himmelsbach, J.H. Duckworth, M.J. Smith, Appl.
Spectrosc. 46 (1992) 420.
[12] S. Navea, A. de Juan, R. Tauler, Anal. Chem. 75 (2003) 5592.
[13] L. Pillonel, W. Luginbuhl, D. Picque, E. Schaller, R. Tabacchi, J.O.
Bosset, Eur. Food Res. Technol. 216 (2003) 174.

[14] P.J. de Groot, H. Swierenga, G.J. Postma, W.J. Melssen, L.M.C. Buydens, Appl. Spectrosc. 57 (2003) 642.
[15] T. Becker, T. Enders, A. Delgado, Bioprocess. Biosyst. Eng. 24 (2002)
347.
[16] T. Kurz, M. Fellner, T. Becker, A. Delgado, J. Inst. Brew. 107 (2001)
241.
[17] I.C. Trelea, M. Titica, S. Landaud, E. Latrille, G. Corrieu, A. Cheruy,
Math. Comput. Simul. 56 (2001) 405.
[18] L.A. Garcia, F. Argueso, A. Garcia, M. Diaz, J. Indust. Microbiol. 15
(1995) 401.
[19] R. Simutis, I. Havlik, A. Lubbert, J. Biotechnol. 27 (1993) 203.
[20] A. Alcazar, F. Pablos, M.J. Martin, A.G. Gonzalez, Talanta 57 (2002)
45.
[21] A.C. Olivieri, H.C. Goicoechea, F.A. Inon, Chemom. Intell. Lab. Syst.
73 (2004) 189.
[22] F. Despagne, D.L. Massart, Analyst 123 (1998) 157R.
[23] D.L. Massart, B.G.M. Vandeginste, L.M.C. Buydens, S. de Jong, P.J.
Lewi, J. Verbeke, Handbook of Chemometrics and Qualimetrics (Part A
and B), Elsevier Science B.V, Amsterdam, 1997.
[24] D.M. Haaland, E.V. Thomas, Anal. Chem. 60 (1988) 1193.
[25] L.G. Weyer, Appl. Spectrosc. Rev. 21 (1985) 1.
[26] J.J. Workman, Appl. Spectrosc. Rev. 31 (1996) 251.
[27] K. Faber, B.R. Kowalski, J. Chemom. 11 (1997) 181.
[28] A.C. Olivieri, J. Chemom. 16 (2002) 207.