Sie sind auf Seite 1von 10

The Journal of Supercritical Fluids 130 (2017) 327–336

Contents lists available at ScienceDirect

The Journal of Supercritical Fluids


journa l homepa ge: www.elsevier.com/locate/supflu

Extraction kinetics and ANN simulation of supercritical fluid extraction of MARK


sage herbal dust
Branimir Pavlić, Oskar Bera, Senka Vidović, Lazar Ilić, Zoran Zekovi ć

University of Novi Sad, Faculty of Technology, Bulevar Cara Lazara 1, 21000 Novi Sad, Serbia

A R T I C L E I N F O A B S T R A C T

Keywords: The aim of this research was optimization of supercritical fluid extraction (SFE) of sage herbal dust obtained as
Sage herbal dust by-product from filter tea factory. Extraction kinetics modelling and artificial neural network (ANN) simulation
Supercritical fluid extraction (SFE) were used for that purpose. Experiments were performed within expanded Box-Behnken experimental design on
Kinetics modelling three levels and three variables. Influence of pressure (100–300 bar), temperature (40–60 °C) and CO 2 flow rate
Artificial neural network (ANN)
(0.2–0.4 kg/h) on total extraction yield was determined. In order to determine initial slope, extraction curves
Optimization
were fitted with five modified empirical models. Since Sovová model provided the best accordance with ex-
perimental data, initial slope obtained by this model was used as response variable for optimization with ANN
and multivariable models (linear, exponential, logarithmic I and logarithmic II). Optimized SFE parameters for
maximized initial slope were pressure of 283 bar, temperature of 60 °C and CO 2 flow rate of 0.4 kg/h.

1. Introduction selectivity by changing of the pressure and/or temperature and possi-


bility to obtain high-quality solvent-free extracts [11]. Extraction of
Sage (Salvia officinalis L.) represents an interesting medicinal plant bioactives from medicinal plants is defined by the mass transfer phe-
which is commonly used in the form of the filter tea. Besides that, sage nomena, physico-chemical interactions of solvent and solutes and
leaves (Salviae folium), essential oil (Salviae aetheroleum ) and its tincture properties of the plant material. SFE could be characterized as the
are official drugs according to European Pharmacopoeia [1]. Sage has diffusion-controlled process influenced by the solute’s molecular mass,
great commercial value due to its various biological effects such as: surface area and film thickness. Mathematical analysis of the SFE pro-
antidiabetic, antioxidant, gastroprotective, anti-inflammatory, spas- cess could be performed by application of approximated kinetic models.
molytic, antimicrobial and anticarcinogenic activities [2]. Its biological This could be particularly important for the scaling of the SFE processes
potential could be attributed to particularly high content of valuable from laboratory to pilot scale. Therefore, kinetic models used for SFE
bioactive compounds, polyphenols [3] and essential oil [4]. Recently, could be based on heat and mass transfer analogy, or differential mass
there have been reports about the possible utilization of herbal dust balance equations [12].
generated in filter tea factory for the recovery of various bioactive S. officinalis was extracted by pressurized CO 2 at following condi-
compounds such as polyphenols [5,6], chlorophylls and carotenoids tions: pressure of 9–128 bar and temperature of 25–50 °C and it could
[7]. Furthermore, sage herbal dust was successfully used as raw mate- be concluded that a part of essential oil was extracted immediately,
rial for extraction of polyphenols using novel extraction techniques while the other fraction was gradually extracted with cuticular waxes
such as ultrasound-assisted extraction, microwave-assisted extraction [13]. Glisic et al. [10] investigated isothermal (50 °C) SFE of sage at
[8] and subcritical water extraction [9]. To the best of our knowledge, 100–300 bar pressure range and observed constant increase of total
sage herbal dust has not been utilized for recovery of lipophilic com- extraction yield with the increase of pressure. However, essential oil
pounds, i.e. essential oil. compounds (monoterpenes and oxygenated monoterpenes) were se-
Sage contains 0.7–5.2% (v/m) essential oil [10], which yield highly lectively extracted at lower pressure (100 bar). Furthermore, Glisic
depends on applied extraction technique, as well as on plant geo- et al. [10] reported that content of diterpenes could be increased in
graphical origin. Even though sage essential oil has been commonly lipid extracts when ultrasonic pretreatment of sage was followed by SFE
obtained by hydrodistillation, in last decade supercritical fluid extrac- fractionation. On the other hand, Akalin et al. [14] applied response
tion (SFE) has been commonly used for that purpose. The reasons for surface methodology (RSM) for optimization of pressurized batch
that are well-known advantages of SFE such as modulation of solvent’s ethanol extraction at following experimental domain: temperature of


Corresponding author.
E-mail address: zzekovic@tf.uns.ac.rs (Z. Zeković).

http://dx.doi.org/10.1016/j.sup flu.2017.06.015
Received 9 May 2017; Received in revised form 22 June 2017; Accepted 23 June 2017
Available online 24 June 2017
0896-8446/ © 2017 Elsevier B.V. All rights reserved.
B. Pavlić et al. The Journal of Supercritical Fluids 130 (2017) 327–336

280–320 °C, extraction time of 60–120 min and plant concentration in Table 1
ethanol of 8–12%. According to literature data, chemical profile of sage Experimental design with SFE parameters and obtained values of total extraction yield.
extracts obtained by SFE varies with applied extraction conditions.
Run Pressure Temperature [°C] CO2 flow rate Density [g/ Y [%]
However, the most abundant compounds in monoterpene fraction are [bar] [kg/h] cm3]
1,8-cineole, α-thujone and camphor, while sesquiterpene fraction is
usually consisted of viridiflorol [10,13]. Furthermore, diterpene poly- Box Behnken experimental design
1 100 60 0.3 0.290 1.81
phenols could be also present in these extracts [10,15]. Application of
2 300 50 0.2 0.871 5.66
batch ethanol-pressurized CO2 extraction at accelerated temperatures 3 200 40 0.4 0.840 6.58
(280–320 °C) could cause significant chemical alterations in sage oil, 4 200 60 0.2 0.724 5.07
which could result in large number of newly formed compounds with 5 200 40 0.2 0.840 5.14
relatively low content [14]. 6 300 60 0.3 0.830 7.82
7 100 50 0.4 0.384 3.57
The main goal of this research was to obtain different empirical
8 100 50 0.2 0.384 2.55
kinetic models which would be able to adequately describe SFE process 9 300 40 0.3 0.910 6.42
of sage herbal dust. Furthermore, another goal was to determine effects 10 200 60 0.4 0.724 7.48
of investigated SFE parameters (pressure, temperature and CO2 flow 11 200 50 0.3 0.784 5.56
12 100 40 0.3 0.629 2.75
rate) on initial slope of extraction curve and on model parameters. To
13 200 50 0.3 0.784 5.68
the best of our knowledge, artificial neural network (ANN) simulation 14 200 50 0.3 0.784 5.71
was not previously applied for mathematical modelling of SFE-sage 15 300 50 0.4 0.871 8.83
system. Therefore, fitting of extraction curves was followed by ANN and Additional experiments
multivariable modelling in order to determine optimal SFE conditions. 16 300 50 0.3 0.871 5.80
17 100 50 0.3 0.384 2.99
Initial slope of extraction curve was used as response variable in ANN
and multivariable optimization, since it represents solubility-controlled
extraction phase parameter. Table 1. Total extraction yield (Y) was measured after the each ex-
perimental run and result was expressed as grams of total extractable
2. Materials and methods compounds per 100 g of plant material (g/100 g), i.e. percentage (%).
The separator conditions were 15 bar and 25 °C.
2.1. Plant material
2.4. Modelling of the extraction curves
Sage (Salvia officinalis L.) originating from Montenegro was kindly
donated by local filter tea factory, Fructus DOO (Bačka Palanka, Experimentally obtained kinetic curves for each experimental run
Serbia). Aerial parts of sage herbal dust, obtained as by-product in filter (17 runs) were fitted to empirical models obtained from commonly
tea processing, was used as raw material in this research. Sample had used empirical equations used for modelling of SFE process. The first
mean particle size < 0.315 mm. model equation was obtained from the Brunner’s equation [17], which
represents a specific solution of Fick’s law:
2.2. Chemicals
Y = x0 (1 − e−kt ) (1)
Commercial carbon dioxide (Messer, Novi Sad, Serbia), purity > where Y represents the extraction yield (%), k is the rate constant
99.98%, was used for laboratory supercritical fluid extraction. All other (min−1), t is the extraction time (min) and x 0 is the initial content of the
chemicals were of analytical reagent grade. solute in the solid phase (g/g), which is usually obtained by the Soxhlet
extraction with suitable organic solvent. Adjustable parameter in this
2.3. Supercritical fluid extraction equation was k, while x 0 is constant. For mathematical modelling,
previous equation was modified with addition of one adjustable para-
The supercritical fluid extraction (SFE) experiments were performed meter (Y∞):
on laboratory scale high pressure extraction plant (HPEP, NOVA, Swiss,
Efferikon, Switzerland) described in detail by Pekić et al. [16]. The Model I : Y = Y∞ (1 − e− kt) (2)
main plant parts and properties, by manufacturer speci fication were: where Y∞ is the total extraction yield obtained for infinite extraction
gas cylinder with CO2, the diaphragm type compressor with pressure time and is specific for each experimental run, i.e. each set of process
range up to 1000 bar, extractor with heating jacket for heating medium parameters.
with internal volume 200 mL, maximum operating pressure of 700 bar, Similar model was proposed by Reverchon and Sesti Osseo [18] for
separator with heating jacket for heating medium (with internal volume SFE of basil oil:
200 mL, maximum operating pressure of 250 bar), pressure control
valve, temperature regulation system and regulation valves.
Sample (35.0 g) was placed in an extractor vessel and extraction
(
YN = 100 1 − e− ti
t
) (3)
process was carried out and extraction yield was measured after 15, 30, where YN is normalized extraction yield ((Y/Ymax)100) and ti is internal
45, 60, 90, 120, 180 and 240 min of extraction, in order to study the diffusion time. Approximating that t i for certain extraction system could
dynamics and kinetics of the process. First set of experiments consisted be considered as constant, addition of adjustable parameter (a) could
of Box-Behnken experimental design with three variables at three levels avoid use of ti [19]:
with three replicates at the central point (15 runs). Pressure (100, 200
−1
and 300 bar), temperature (40, 50 and 60 °C) and CO2 flow rate (0.2, a=
ti (4)
0.3 and 0.4 kg/h) were independent variables in the process, while all
other SFE parameters were held constant. Furthermore, experimental Adding Eq. (4) in Eq. (3) and a correction term (b), Reverchon and
design was expanded for two more runs, where different combinations Sesti Osseo model could be obtained:
of independent variables (pressure, temperature and CO2 flow rate)
YN = 100(1 − e(at +b) ) (5)
were applied in order to provide more detailed information about ex-
traction kinetics and in fluence of SFE parameters on total extraction Similarly to Eq. (2), modified Reverchon and Sesti Osseo model
yield. Expanded Box-Behnken experimental design is presented in could be obtained with addition of Y ∞ as adjustable parameter:

328
B. Pavlić et al. The Journal of Supercritical Fluids 130 (2017) 327–336

Model II : Y = Y∞(1 − e (at+ b)) (6) minimizes a combination of squared errors and weights and then de-
termines the correct combination. Weights and biases values of ANN
Third kinetic model could be derived by model proposed by
model were presented in Table S1 (Supplementary material). Bayesian
Esquivel et al. [20], given by the following equation:
regularization does not use data for validation and thus it is suitable for
t
me = x0 F ⎛ ⎞ small number of experimental data. ANN architecture was optimized by
⎝b + t ⎠ (7) varying the number of neurons in hidden layer and by changing initial
values of weights and biases.
where m e is the mass of extract (g), F is the mass of solid material (g), t
In order to determine the relative importance and influence of input
is the extraction time (min), x0 is the initial content of the solute in the
variables on initial slope, assessment process based on the connection
solid phase (g/g) and b is adjustable parameter (min). Again, this
weights partitioning (Yoon’s method) was applied. Also, the optimiza-
equation was modified with addition of one adjustable parameter
tion of the extraction procedure (maximizing the IS) was conducted in
[21,22] and used in following form:
MATLAB with built in functions fmincon and MultiStart with 15

Model III : Y = Y∞ ⎛
t ⎞ random starting points in order to find global maximum.
⎝k + t⎠ (8)
2.5.2. Multivariable models
where Y is the total extraction yield (%), while Y∞ and k are the ad-
Besides ANN, multivariable models were also used to fit the de-
justable parameters.
pendence of IS on pressure, temperature and CO 2 flow rate. Four dif-
Fourth model was derived from kinetic model proposed by Kandiah
ferent multivariable models were tested: linear, exponential and two
and Spiro [23]:
logarithmic functions (Eqs. (12)–(15)). Dependent variable chosen to
Y = x0 { 1 − [ ƒ1 e(−k1 t ) + ƒ2 e( −k2 t) ]} (9) describe extraction kinetics was initial slope (IS) obtained from the V
kinetic model – Sovová model. IS was successfully used as kinetics
where f1 and f2 are the solute fractions extracted with the rate constant parameter in previous papers [25] and Sovová model exhibited the best
k1 and k 2, respectively. This model was modified by the addition of Y ∞ fit with experimental data.
as adjustable parameter in order to obtained final equation:
Linear model: IS = b1 + b2 p + b3 T + b4 q (12)
Model IV : Y = Y∞ {1 − [ƒ 1e ( −k1 t) + ƒ 2 e ( −k2 t) ]} (10)
Exponential model: IS = b1 + b 2 e b3 p + b4 e b5 T + b6 e b 7 q (13)
Modelling of the SFE process from various plant material could be
successfully performed using simplified mathematical models reported Logarithmic I model : IS = b1 + b2 ln(p) + b3 ln(T ) + b4 ln(q ) (14)
by Sovová [24]. These models were characterized by characteristic
Logarithmic II model : IS = b1 + b 2 ln(pTq ) (15)
times during the SFE process, i.e. characteristic time of mass transfer in
the fluid phase (t f), characteristic time of internal mass transfer (ti), where b 1-b 4 (Eq. (12)), b1-b 7 (Eq. (13)), b1-b 4 (Eq. (14)) and b1-b2 (Eq.
characteristic time of extraction equilibrium (teq) and mean residence (15)) are regression coeffi cients for linear, exponential, logarithmic I
time of the solvent in the extractor (t r). For modelling of the extraction and logarithmic II model, respectively. All fitting models were utilised
process which includes a plug flow in the extractor, Sovová proposed to determine the relative importance of variables on initial slope. Fit-
the following model: ting parameters in models were calculated using least square method
and experimental data points were normalized using feature scaling in
Model V: Y = Y∞ G t for t ≤ t1 =
t G
1 Km q˙ order bring all values into the range 0–1 for linear and exponential
t− t1
Y = Y∞ ⎛1 − (1 − G) e− ( ) ⎞ for t > t1 models or into the range 1–2 for logarithmic functions.
ti
⎝ ⎠ (11) Partial derivatives of multivariable models were used to calculate
and estimate the relative importance and influence of input variables on
where G, Km and ti are adjustable parameters, while q̇ is the specific CO2 initial slope:
flow rate (kg CO2/kg plant h). K m is the mass-related partition coeffi-
∂IS
cient which represents the equilibrium of the mass concentrations on
the particle surface and G is closely related to particle size and frag- Infx i =
mean ( ) ⋅100[%]; i = 1…3
∂x i

∑n= 1
n= i ∂ IS
mentation. mean ( ) ∂ xn
(16)

2.5. Influence analysis and optimization where Inf xi is relative influence of independent variables (pressure,
temperature and CO2 flow rate, respectively) and xi is their respective
Two different approaches were used in order to determine the in- coded value.
fluence of pressure, temperature and CO2 flow rate on extraction ki-
netics parameters: artificial neural network (ANN) and fitting of mul- 2.6. Statistical analysis
tivariable models.
The accordance between experimentally obtained extraction yields
2.5.1. Artificial neural network (ANN) and calculated values obtained from fitted model equations by five
All data analysis with ANN was performed using MATLAB software mathematical models was established by the sum of squared errors
(The Math Works Inc., License Number 1108951). The ANN models (SSer ) and coefficient of determination (R2).
with one hidden layer were designed using MATLAB Neural Network
Toolbox. The constructed ANN model had 3 inputs (pressure, tem- 3. Results and discussion
perature and CO2 flow rate) and one output (initial slope (IS) obtained
from the V kinetic model – Sovová model) (Fig. 1). SFE was applied in order to recover lipophilic bioactive compounds
Training of ANN was performed using the Bayesian regularization. from sage herbal dust. Extractions were performed at different set of
From all collected data, 85% has been used for training and 15% has SFE parameters and experimentally obtained values of total extraction
been considered for testing. Hyperbolic tangent sigmoid and linear yield (Y) are presented in Table 1. It could be observed that Y was in
function were used as transfer functions for input-hidden and output range from 1.81 to 8.83%, and the highest Y was obtained at pressure of
layer, respectively. Bayesian regularization updates the weight and bias 300 bar, temperature of 50 °C and CO2 flow rate of 0.4 kg/h. SFE of
values according to Levenberg-Marquardt optimization which terpenoid compounds from S. officinalis was comprehensively

329
B. Pavlić et al. The Journal of Supercritical Fluids 130 (2017) 327–336

Fig. 1. Schematic structure of the ANN architecture


used for optimization (8 hidden neurons).

researched in the last decade [10,26,27]. Babovic et al. [28] performed


SFE of sage at increased temperature (100 °C) and moderate pressure
(250–300 bar) in order to increase solvent selectivity towards diterpene
polyphenols (carnosic acid, carnosol and epirosmanol), however, Y was
decreased (1.53%). On the other hand, Glisic et al. [10] compared SFE
with conventional extraction techniques used for essential oil recovery
(Soxhlet extraction and hydrodistillation). Y of essential oil obtained by
hydrodistillation was 0.5%, while the highest Y obtained by SFE was
4.28%, which was achieved at the same set pressure (300 bar) and
temperature (50 °C), i.e. CO 2 density (0.871 g/cm3), as it was the case
in this work (Table 1). Fornari et al. [26] proposed two-stage fractio-
nation sage extracts obtained by SFE at 300 bar and 40 °C. First extract
was separated at 100 bar and 40 °C yielding 1.39%, while conditions in
Fig. 2. Influence of pressure on SFE kinetics at 50 °C and 0.3 kg CO2/h.
second separator were 10 bar and 40 °C providing 3.23% yield of ex-
tract rich in essential oil compounds.

3.1. Influence of SFE parameters on total extraction yield

Box-Behnken experimental design was expanded with two addi-


tional experiments (runs 16 and 17) in order to provide more thorough
analysis of SFE parameters (pressure, temperature and CO2 flow rate)
influence on Y (Table 1). Influence of pressure was determined at 100,
200 and 300 bar, while temperature (50 °C) and CO2 flow rate (0.3 kg/
h) were held constant and obtained kinetic curves were presented on
Fig. 2. It could be observed that Y increased from 2.99 to 5.80% with
increase of pressure from 100 to 300 bar for total extraction time
(240 min). Furthermore, Y increase was more pronounced for pressure
increase from 100 to 200 bar, while further increase of pressure Fig. 3. Influence of CO2 density on Y at 50 °C and 0.3 kg CO 2/h.

(300 bar) did not provide significant Y increase. This is rather expected,
since CO 2 density at 50 °C and pressures of 100, 200 and 300 bar is

330
B. Pavlić et al. The Journal of Supercritical Fluids 130 (2017) 327–336

Fig. 5. It could be observed that the highest Y (8.83%) was achieved


when the highest CO2 flow rate (0.4 kg/h) was applied. On the other
hand, there was no significant difference in kinetic curves obtained at
0.2 and 0.3 kg/h (Fig. 5). CO2 flow rate influence mass transfer prop-
erties such as axial dispersion, convective mass transfer coefficient and
accumulation in the fluid phase, contrary to pressure and temperature
which affect both mass transfer and thermodynamic properties [11].
Extraction efficiency could be increased with reduction of mass transfer
resistance which could be achieved with high interstitial velocity,
which was achieved with 0.4 kg/h CO 2 flow rate. However, influence of
CO2 flow rate must be also evaluated from the economic point of view
since increased solvent consumption consequently increases opera-
tional costs. Furthermore, SFE kinetic curves could be divided in few
Fig. 4. Influence of temperature on SFE kinetics at 300 bar and 0.3 kg CO2 /h.
extraction periods controlled by different mass transfer mechanisms
controlled by convection, solubility and internal diffusion [31].
0.384, 0.784 and 0.871 g/cm3 , respectively. Therefore, significant in-
Therefore, influence of CO2 flow rate must be evaluated together with
crease of CO 2 density is proportional with the increase of Y (Fig. 3).
extraction time and extraction kinetics modelling would be essential for
Similar influence of pressure at isothermal conditions (50 °C) was re-
that purpose.
ported by Glisic et al. [10]. Furthermore, pressure has been the most
important SFE parameter infl uencing Y [26]. Increase of pressure leads
3.2. Extraction kinetics modelling
to increase of CO2 density increasing solvating power of CO2 and re-
ducing its selectivity towards volatile compounds. It is not always
The extraction kinetics of lipophilic compounds from sage herbal
suggested to perform SFE at the highest pressure, since it is possible to
dust at different set of pressure, temperature and CO2 flow rate
dilute target compounds content in extract. However, it must be con-
(Table 1) was observed after 15, 30, 45, 60, 90, 120, 180 and 240 min.
sidered that concomitant compounds would not necessarily decrease
Obtained experimental data were fitted to the empirical models de-
extracts quality and bioactivity.
scribed with Eqs. (2), (6), (8) , (10) and (11) and calculated parameters
Temperature effect on Y was investigated at 40, 50 and 60 °C, while
for five applied models are presented in Table 2. Accordance between
pressure and CO 2 flow rate were held constant (300 bar and 0.3 kg/h).
experimental data and applied models was described by sum of squared
Kinetic curves are presented at Fig. 4. The highest Y (7.48%) after
errors (SS er) and coefficient of determination (R 2) which were pre-
240 min of extraction time was at 60 °C, while the lowest Y (5.68%)
sented in Table S2 (Supplementary material). According to statistical
was observed at 50 °C under isobaric conditions, suggesting the com-
parameters, i.e. relatively low SSer and particularly high R 2, it could be
plexity of the temperature influence on supercritical CO2 and plant
observed that all applied models represented adequate approximation
material. Increased temperature causes decrease of CO2 density and
of experimental data (Table S2). Furthermore, model V provided the
further decrease of its solvating power. On the other hand, temperature
best fit with experimental data according to mean value of SS er and R2
affects vapour pressure of the solute causing increase of solubility,
(0.0675 and 0.9973) (Fig. 6f). Model I was described with Y ∞ and k as
particularly for volatile compounds. According to results, relatively
adjustable parameters. Furthermore, Y∞ was calculated parameter in
high Y was obtained at 40 °C, which decreased when extraction was
all investigated models. According to results from Table 2, it could be
performed at 50 °C (Fig. 4), due to decrease of CO2 density from 0.840
observed that investigated SFE parameters exhibited similar influence
to 0.784 g/cm3. Further increase of temperature provided the highest Y,
on Y ∞ as it was the case with experimentally observed Y. This is rather
suggesting prevailing influence on vapour pressure of sage terpenoids
expected since Y∞ represents asymptotic value of total extraction yield
over decrease of CO2 density at 60 °C (0.724 g/cm3 ). According to lit-
specific for each set of pressure, temperature and CO2 flow rate. On the
erature data, there is no information about temperature influence on Y
other hand, k from the model I was the highest at 200 bar
in similar experimental domain (40 –60 °C). Other SFE of sage were
(0.0165 min−1) and the lowest at 300 bar (0.0109 min−1) at fixed set
either focused on influence of pressure [10], particle size [29], or on
of temperature (50 °C) and CO2 flow rate (0.3 kg/h). Furthermore, k
batch ethanol-pressurized CO 2 extraction at temperature range
was the lowest at 50 °C (0.0109 min−1) and the highest at 60 °C
280–320 °C [14], which is not suitable for comparison with present
(0.0182 min−1) at constant pressure and CO2 flow rate, as it was the
data. However, similar influence of temperature in the same experi-
case with Y (Fig. 4). Experimental and fitted curve of model I for the
mental domain (40–60 °C and 300 bar) on SFE of lavender essential oil
sample 14 (central point; 200 bar, 50 °C and 0.3 kg CO2 /h) with char-
was reported by Nađalin et al. [30].
acteristic model parameters is presented on Fig. 6a. According to data
CO2 flow rate influence on Y was investigated at isothermal-isobaric
from Table 2, it could be observed that Y∞ was very similar for models I
conditions (50 °C and 300 bar) and kinetic curves are presented on
and II, however, influence of SFE parameters on a was rather complex
due to addition of correction term b. Addition of one adjustable para-
meter slightly increased agreement of experimental data and model II
(Table S2). However, it could be observed that fitting curves of models I
and II were very similar for sample 14 (Fig. 2a and b). In model III (Eq.
(8)), calculated Y∞ was generally the highest for experimental runs
(Table 2), however, trend of SFE parameters influence on Y ∞ was the
same as it was observed in other models. Influence of SFE parameters
on k from model III was completely opposite from the model I (Table 2),
which is expected since Y decreases with the increase of k (Eq. (8)).
According to statistical data (SSer and R2 ), model III (modified Esquivel
et al. model) exhibited the lowest fit with experimental results (Table
S2), contrary to SFE of winter savory [32] and coriander seeds [25] at
same equipment and similar set of SFE parameters, where Esquivel
et al. model exhibited better fit than one-stage diffusion model. Model
Fig. 5. Influence of CO2 flow rate on SFE kinetics at 300 bar and 50 °C. IV (modified Kandiah and Spiro model) could be separated on two

331
B. Pavlić et al. The Journal of Supercritical Fluids 130 (2017) 327–336

extraction curves presented on Fig. 6d. Fitted extraction curve for

ti [min]

139.66

115.78
143.71
117.82

137.51
118.47
157.49

57.20

66.27
77.20
80.50

76.55
93.26
65.97
72.24
88.26

78.49
model IV was very similar to I and II model curves (Fig. 6a, b and d).
Part 1 of the extraction curve represents the fraction extracted in so-
lubility-controlled extraction period, while part 2 represents the frac-
t1 [min]

tion extraction in diffusion controlled period. It could be observed that

45.00

46.41
36.92

31.32
47.39
32.82

35.21
21.77
51.86
45.00
14.99

52.26

54.57
48.39
19.77

45.00

17.77
Y∞ in this model was generally similar to Y∞ obtained in models I and
II, however, SFE parameters significantly influence the fractions ob-
tained in first and second phase of extraction, as well as kinetic coef-
0.14322

0.09012

0.07800

0.04740

0.06030
0.10836

0.10128
0.08454
0.08958

0.07134
0.06924
0.12042
0.05568
0.12300

0.08442

0.09846
0.06810
ficients for these periods (k 1 and k2 ) (Table 2). At fixed set of pressure
Km

and CO 2 flow rate, f1 was the highest at 50 °C (0.692%). Similarly, f 1


was also the highest at middle level of CO 2 flow rate at constant pres-
sure and temperature (Table 2). On the other hand, f 1 was 0.844, 0.500
0.5713

0.4531
0.5724
0.5602

0.3588
0.4044

0.3860

0.2640

0.3047

0.3998
0.2871
0.5943
0.4772
0.1755

0.6303

0.5113
0.4708

and 0.692% at pressure of 100, 200 and 300 bar, respectively, at fixed
G

temperature and CO2 flow rate. Model IV represented better approx-


imation of experimental data than models I and II (Table S2), as it was
Model V

Y∞ [%]

the case with SFE of basil at similar experimental domain [33]. The
5.66

5.78
5.89
7.82

3.52
2.12

5.55

3.72

3.23

6.54
9.25
5.94
6.86
5.26

8.37

2.88
7.20

lowest mean SSer (0.0675) and the highest R 2 (0.9973) were obtained
for the model V (Sovová model), therefore, this model exhibited the
best fit with the experimental data. The most important parameters in
]
−1

Sovová model are characteristic times of the extraction steps [24]. It


k2 [min

0.0037

0.0140

0.0175
0.0106
0.0160
0.0137
0.0134

0.0234

0.0177
0.0206

0.0206

0.0137
0.0166
0.0086

0.0470
0.0104

0.0852

could be observed that t 1 was < 60 min for all SFE runs, suggesting that
constant extraction rate in the first phase of extraction is solubility-
controlled process due to good transport properties of supercritical CO2.
f2 [%]

Due to direct effect on concentration gradient, CO 2 flow rate exhibited


0.502

0.680
0.500
0.468

0.156
0.663

0.500

0.093

0.244

0.310
0.697
0.500
0.500
0.500

0.387

0.379
0.320

greater influence on t1 comparing to pressure and temperature


(Table 2). Another characteristic time is t i , which represents the internal
diffusion-controlled extraction phase. Even though it is expected that ti
]
−1

should be flow-independent parameter [24], CO2 flow rate significantly


k 1 [min

0.0711

0.0045

0.0078
0.0244
0.0160
0.0137
0.0134

0.0148

0.0108
0.0097

0.0208

0.0292
0.0166
0.0396

0.0087
0.0104

0.0124

affected ti in the range from 0.2 to 0.4 kg/h, while temperature effect on
ti was negligible (Table 2). Fitted extraction curve of the model V and
run 14 with characteristic times could is presented on Fig. 6e. Mass-
f 1 [%]

0.303
0.500
0.500
0.500

0.613

0.621
0.681

0.498

0.320
0.500
0.532

0.844
0.337

0.500

0.907

0.758

0.692

related partition coefficient (Km) was between 0.0603 and 0.1432 and
was strongly influence by SFE parameters. Furthermore, all SFE para-
meters also exhibited high influence G, which is rather expected since
Model IV

Y ∞ [%]

this parameter is associated with particle fragmentation degree and


9.39
5.71
7.01
5.29

7.82

2.62
6.86

5.61

5.78
5.76
8.04

3.29
2.50

5.69

3.74

3.71

6.50

rupture of the cells occurs during SFE [30].

3.3. ANN simulation and multivariable modelling


k [min]

122.79

188.89

112.59
75.65
68.67
85.37
86.12

57.16

86.83
88.38

47.72

56.27
65.11
38.99

87.38
38.82

71.24

Recent optimization studies of SFE focus on application of novel


statistical techniques (RSM and ANN) in order to determine the best set
Model III

of process parameters (pressure, temperature, CO 2 flow rate, extraction


11.88
7.42
9.40
7.08
8.03
9.78

3.46
8.89

6.97
4.89
7.13
7.44

8.59
8.83

4.00
1.99

4.68
Y∞

time, particle size, cosolvent concentration, etc.), which would provide


the highest Y, or the highest yield of target compound(s). RSM has been
Calculated parameters of five empirical models applied for SFE kinetics modelling.

commonly applied procedure in most of the optimization studies [34],


− 0.0178

− 0.0072

− 0.0226
−0.0735

−0.0174
0.0337
0.0312
0.0048
0.0146
0.0244

0.0249
0.0139

0.0271
0.0184
0.0045
0.0251

0.0032

however, application of ANN for this purpose is gaining more and more
attention recently [35–40]. From economical point of view, the main
b

drawback of this approach is limitation for optimization of industrial


scale processes, since Y is usually the main response. Y is measured at
−0.0225
−0.0215
−0.0080
−0.0183
−0.0173
−0.0143
−0.0110
−0.0122
−0.0193
−0.0170
−0.0145
−0.0135
−0.0108
−0.0190
−0.0145
−0.0142
−0.0136

the end of the SFE process which lasts until the complete exhaustion of
the plant material, i.e. internal diffusion-controlled period. This is
a

hardly scalable on the industrial processes since it is not economically


feasible to perform extractions until diffusion-controlled period due to
Model II

Y∞ [%]

high time consumption. Therefore, it was proposed in our previous


7.34
5.59
3.16
5.65
5.73
9.11
6.18
3.05
1.70
5.65
6.93
5.28
5.63
7.66
3.64
2.55
6.58

study that initial slope calculated from the kinetic curves could be
suitably used as response for optimization study [25], to make opti-
]

mization more applicable to large scale processes. Initial slope would


−1

0.0225
0.0160
0.0137
0.0134
0.0104
0.0182
0.0151
0.0134
0.0131
0.0232
0.0206
0.0074
0.0181
0.0165
0.0145
0.0109
0.0129
k [min

adequately represent solubility-controlled extraction phase, where ex-


traction rate is constant (Fig. 6). Therefore, initial slope obtained from
the Sovová model was used as response variable, since this model
showed the best fit with experimental data (Table S2, Supplementary
Model I

Y∞ [%]

material).
1.66
5.70
7.01
5.29
5.69
7.70
3.62
2.58
6.63
7.32
5.62
3.26
5.65
5.77
9.08
6.19
3.01

Results obtained by ANN could significantly vary with initial value


of parameters used for construction of ANN and fitting, as well as with
Table 2

Run

number of hidden neurons used for ANN construction. Therefore, the


10
11
12
13
14
15
16
17
1
2
3
4
5
6
7
8
9

number of neurons in hidden layer was varied from 1 to 20 and the

332
B. Pavlić et al. The Journal of Supercritical Fluids 130 (2017) 327–336

Fig. 6. Experimental data and fitting curves of sample 14 (200 bar, 50 °C and 0.3 kg CO 2/h) for five kinetic models.

training process of each network was repeated 10 times with random R2, ANN model exhibited the best fit with the experimental data, while
initial values of weights and biases in order to avoid mentioned influ- the logarithmic model II with the highest SSer (0.0038) and the lowest
ences on ANN results. The result of this procedure was creation of 200 R2 (0.747) exhibited the poorest fit. R2 value was in correlation with
ANNs in total. Neural networks with coefficient of determination (R2) number of regression used in each model, which is rather expected
higher than 0.5 were used for further analysis (198 of 200 ANNs). R2 since addition of regression coefficients in the model equation would
value showed rising trend with increase of the number of neurons in the increase the accordance with the experimentally obtained results.
hidden layer. Mean value of R2 was 0.768 for all created ANN, while the Therefore, accordance between experimental results and model equa-
best fitting was achieved with neural network with 8 hidden neurons tions in increasing order was: ANN, exponential, logarithmic I, liner and
(R 2 = 0.917, SSer = 0.0013). Therefore, this ANN was used for further logarithmic II, while SS er for these models was 0.0013, 0.0019, 0.0035,
optimization calculations in order to avoid “overfitting” with high 0.0036 and 0.0038, respectively.
number of hidden neurons. In order to avoid large number of regression Weight matrices obtained by ANN were used for calculation of re-
coefficients in ANN which were obtained by relatively small number of lative importance and its infl uence (Inf) on response variable. For ANN
experiments, experimental data were fi tted by alternative multivariable optimization, influence of each variable was calculated by equation
models (linear, exponential, logarithmic I and logarithmic II). proposed by Yoon et al. [41], while the same influence in multivariable
Regression plots for ANN and multivariable models with fitting quality models was calculated by partial derivatives of model presented in Eq.
parameters (SSer and R 2) are presented on Fig. 7. According to SS er and (16). Therefore, ANN simulation provided quantification of SFE

333
B. Pavlić et al. The Journal of Supercritical Fluids 130 (2017) 327–336

Fig. 7. Regression plot for ANN with best performance (8 hidden neurons) and multivariable models.

Fig. 8. Relative influence of SFE parameters obtained for all applied models.

334
B. Pavlić et al. The Journal of Supercritical Fluids 130 (2017) 327–336

Table 3 (0.0675) and the highest R2 (0.9973). Initial slope obtained from this
Optimization of SFE conditions for maximal initial slope and comparison with experi- model was used as a response variable for ANN simulation multi-
mental results.
variable modelling, since it represents solubility-controlled extraction
Pressure Temperature [°C] CO2 flow Initial slope [%
phase which could be applicable for optimization of industrial scale
[bar] rate [kg/h] min−1 ] processes. Predicted SFE conditions which provided the highest value of
initial slope (0.13347% min−1) were pressure of 283 bar, temperature
Experimental 200 60 0.4 0.13347
of 60 °C and CO 2 flow rate of 0.4 kg/h, which was 3.83% higher com-
ANN model 283 60 0.4 0.13858
paring to the highest experimentally obtained initial slope observed in
sample 10. It could be observed that all SFE parameters exhibited po-
parameters influence on initial slope, comparing to influence analysis of sitive influence on initial slope, which was con firmed by both ANN and
SFE parameters on Y, which was discussed previously (Chapter 3.1). multivariable modelling. Furthermore, multivariable models exhibited
Relative infl uence of SFE parameters on initial slope obtained by ANN adequate fit with experimental data, which provided obtaining of
and multivariable models is presented on Fig. 8. It could be observed model equations with less regression coefficients comparing to ANN
that pressure was the most influential parameter with 35.18%, while model. Further perspectives of this work would be comparison of ex-
relative influence of temperature and CO2 flow rate was 23.84 and tracts sage extracts obtained by SFE with extracts obtained by con-
32.98%, respectively. Quantitative influence of SFE parameters was in ventional techniques (hydrodistillation and Soxhlet extraction with
accordance with literature data about ANN modelling of SFE of essen- organic solvents) in terms of chemical profile and yield of terpenoid
tial oil from Launaea acanthodes [40] and Coriandrum sativum [25]. compounds responsible for bioactive potential of this plant.
However, temperature exhibited negative influence on initial slope in
SFE of C. sativum (Zeković et al., 2017), while in this work all SFE Acknowledgements
parameters exhibited positive influence (Fig. 8). Positive influence of
pressure could be explained by increased solubility power of CO 2 at The authors would like to thank the Fructus (Bačka Palanka, Serbia)
elevated pressure, while increased CO2 flow rate directly affects the for kind donation of sage herbal dust sample.
concentration gradient and decreases mass transfer limitations. Since
temperature affects both solute vapour pressure and CO2 density, i.e. Appendix A. Supplementary data
solubility power, positive influence indicated that its in fluence on in-
crease of vapour pressure would be more expressed comparing to de- Supplementary data associated with this article can be found, in the
crease of solvent density. Therefore, it could be concluded that positive online version, at http://dx.doi.org/10.1016/j.sup flu.2017.06.015.
influence of temperature on extraction rate is directly connected with
increase of vapour pressure of volatile sage terpenoids. The same dis- References
tribution of SFE parameters influence was obtained by linear, ex-
ponential and logarithmic I model (Fig. 8), suggesting that adequate [1] Council of Europe, European Pharmacopoeia 7.0, 2.9.18 Prep. Inhal. Aerodyn.
Assess. Fine Part, (2010), pp. 274–285.
calculation of relative influence could be obtained with multivariable [2] M. Grdiša, M. Jug-Dujaković, M. Lončarić, K. Carović-Stanko, T. Ninčević, Z. Liber,
models with less regression coefficients. However, logarithmic II model I. Radosavljević, Z. Šatović, Dalmatian sage (Salvia officinalis L.): a review of bio-
showed limitation in determination of relative influence (Fig. 8), since chemical contents, medical properties and genetic diversity, Agric. Conspec. Sci. 80
(2015) 69–78.
all SFE parameters were described with one regression coeffi cient (b2) [3] Y. Lu, L. Yeap Foo, Polyphenolics of salvia—a review, Phytochemistry 59 (2002)
from Eq. (15). 117–140.
Initial slope was the response variable used for the ANN optimiza- [4] A. Occhipinti, A. Capuzzo, A. Arceusz, M.E. Maffei, Comparative analysis of α-and
β-thujone in the essential oil and supercritical CO 2 extract of sage (Salvia officinalis
tion within investigated experimental domain of SFE parameters
L.), J. Essent. Oil Res. 26 (2014) 85–90.
(pressure, temperature and CO2 flow rate). Determination of optimal [5] M. Milutinović, N. Radovanovi ć, M. Ćorović, S. Šiler-Marinković, M. Rajilić-
conditions was performed in MATLAB using ANN with 8 hidden neu- Stojanović, S. Dimitrijevi ć-Branković, Optimisation of microwave-assisted extrac-
rons which showed the best accordance with experimental data. tion parameters for antioxidants from waste Achillea millefolium dust, Ind. Crops
Prod. 77 (2015) 333–341.
Predicted value of initial slope at calculated optimal conditions as well [6] M. Ramić, S. Vidović, Z. Zeković, J. Vladić, A. Cvejin, B. Pavlić, Modeling and
as maximal experimentally observed value of initial slope, are pre- optimization of ultrasound-assisted extraction of polyphenolic compounds from
sented in Table 3. Predicted SFE conditions which provided the highest Aronia melanocarpa by-products from filter-tea factory, Ultrason. Sonochem. 23
(2015) 360–368.
value of initial slope (0.13858% min −1) were pressure of 283 bar, [7] B. Pavlić , A.V. ðurković, J. Vladi ć, A. Gavarić, Z. Zekovi ć, A. Tepić, S. Vidović ,
temperature of 60 °C and CO2 flow rate of 0.4 kg/h. Furthermore, it Extraction of minor compounds (chlorophylls and carotenoids) from yarrow–rose
could be noted that pressure was the only different variable in experi- hip mixtures by traditional versus green technique, J. Food Process Eng. 39 (2015)
418–424.
mentally observed and predicted optimized conditions. Finally, ANN [8] Z. Zeković, D. Pintać, T. Majkić, S. Vidović, N. Mimica-Dukić, N. Teslić, A. Versari,
simulation provided increase of initial slope for 3.83% comparing to the B. Pavlić, Utilization of sage by-products as raw material for antioxidants re-
highest experimentally obtained initial slope observed in sample 10. covery—ultrasound versus microwave-assisted extraction, Ind. Crop. Prod. 99
(2017) 49–59.
[9] B. Pavlić, S. Vidović , J. Vladić, R. Radosavljević, M. Cindri ć, Z. Zeković, Subcritical
water extraction of sage (Salvia officinalis L.) by-products—process optimization by
4. Conclusions response surface methodology, J. Supercrit. Fluid. 116 (2016) 36–45.
[10] S. Glisic, J. Ivanovic, M. Ristic, D. Skala, Extraction of sage (Salvia officinalis L.) by
supercritical CO2: kinetic data: chemical composition and selectivity of diterpenes,
SFE was successfully utilized as novel extraction technique for the J. Supercrit. Fluid. 52 (2010) 62–70.
recovery of lipophilic bioactive compounds from sage herbal dust ob- [11] M.M.R. de Melo, A.J.D. Silvestre, C.M. Silva, Supercritical fluid extraction of ve-
tained as by-product from filter tea factory. Experiments were per- getable matrices: applications, trends and future perspectives of a convincing green
technology, J. Supercrit. Fluids 92 (2014) 115–176.
formed using expanded Box-Behnken experimental design with pres-
[12] H. Sovová, Mathematical model for supercritical fluid extraction of natural products
sure (100, 200 and 300 bar), temperature (40, 50 and 60 °C) and CO2 and extraction curve evaluation, J. Supercrit. Fluids 33 (2005) 35–52.
flow rate (0.2, 0.3 and 0.4 kg/h) as independent variables. The highest [13] S.A. Aleksovski, H. Sovová, Supercritical CO 2 extraction of Salvia officinalis L, J.
yield (8.83%) was achieved at following conditions: 300 bar, 50 ° C and Supercrit. Fluids 40 (2007) 239–245.
[14] M.K. Akalın, K. Tekin, M. Akyüz, S. Karagöz, Sage oil extraction and optimization
0.4 kg CO 2/h. Kinetic curves were fitted with five commonly used by response surface methodology, Ind. Crop. Prod. 76 (2015) 829–835.
empirical models. According to fitting quality parameters (SSer and R2), [15] S.B. Glisic, M. Ristic, D.U. Skala, The combined extraction of sage (Salvia officinalis
all applied models fitted well with the experimental data, however, L.): ultrasound followed by supercritical CO 2 extraction, Ultrason. Sonochem. 18
(2011) 318–326.
Sovová model (V model) exhibited the best fi t due to the lowest SSer

335
B. Pavlić et al. The Journal of Supercritical Fluids 130 (2017) 327–336

[16] B. Peki ć, Z. Zeković, L. Petrovi ć, A. Tolić, Behavior of (–)-α-Bisabolol and (–)-α- [29] O.J. Catchpole, J.B. Grey, B.M. Smallfield, Near-critical extraction of sage, celery
Bisabololoxides A and B in chamomile flower extraction with supercritical carbon and coriander seed, J. Supercrit. Fluids 9 (1996) 273–279.
dioxide, Separ. Sci. Technol. 30 (1995) 3567–3576. [30] V. Nađ alin, Ž. Lepojević, M. Ristić , J. Vladić, B. Nikolovski, D. Adamović,
[17] G. Brunner, Gas Extraction: An Introduction to Fundamentals of Supercritical Fluids Investigation of cultivated lavender (Lavandula officinalis L.) extraction and its ex-
and the Application to Separation Processes, Ed. Darmstadl, Ger. (1994). tracts, Chem. Ind. Chem. Eng. Q. 20 (2014) 71–86.
[18] E. Reverchon, L. Sesti Osseo, Modeling the supercritical extraction of basil oil, [31] R.P.F.F. da Silva, T.A.P. Rocha-Santos, A.C. Duarte, Supercritical fluid extraction of
Proceedings of the 3rd International Symposium on Supercritical Fluids, Perrut and bioactive compounds, TrAC Trends Anal. Chem. 76 (2016) 40–51.
Brunner, Strasbourg, France, 1994 p.189. [32] J. Vladić, Z. Zeković, S. Jokić, S. Svilović, S.V. Kovačević, Winter savory: super-
[19] Z. Zekovi ć, Ž. Lepojević, A. Tolić , Modeling of the thyme-supercritical carbon di- critical carbon dioxide extraction and mathematical modeling of extraction process,
oxide extraction system. I. The influence of carbon dioxide flow rate and grinding J. Supercrit. Fluids 117 (2016) 89–97.
degree of thyme, Sep. Sci. Technol. 36 (2001) 3459–3472. [33] Z. Zeković , S. Filip, S. Vidović , S. Jokić, S. Svilović , Mathematical modeling of
[20] M.M. Esquıvel, M.G. Bernardo-Gil, M.B. King, Mathematical models for super- Ocimum basilicum L. supercritical CO2 extraction, Chem. Eng. Technol. 37 (2014).
critical extraction of olive husk oil, J. Supercrit. Fluids 16 (1999) 43–58. [34] K.M.M. Sharif, M.M.M. Rahman, J. Azmir, A. Mohamed, M.H.A.H. a Jahurul,
[21] S.N. Naik, H. Lentz, R.C. Maheshwari, Extraction of perfumes and flavours from F. Sahena, I.S.M.S.M. Zaidul, Experimental design of supercritical fluid extraction –
plant materials with liquid carbon dioxide under liquid-vapor equilibrium condi- a review, J.Food Eng. 124 (2014) 105–116.
tions, Fluid Phase Equilib. 49 (1989) 115–126. [35] A. Shokri, T. Hatami, M. Khamforoush, Near critical carbon dioxide extraction of
[22] I. Papamichail, V. Louli, K. Magoulas, Supercritical fluid extraction of celery seed anise (Pimpinella anisum L.) seed: mathematical and artificial neural network
oil, J. Supercrit. Fluids 18 (2000) 213–226. modeling, J. Supercrit. Fluids 58 (2011) 49–57.
[23] M. Kandiah, M. Spiro, Extraction of ginger rhizome: kinetic studies with super- [36] G. Zahedi, A. Azarpour, Optimization of supercritical carbon dioxide extraction of
critical carbon dioxide, Int. J. Food Sci. Technol. 25 (1990) 328–338. Passiflora seed oil, J. Supercrit. Fluids 58 (2011) 40–48.
[24] H. Sovová, Steps of supercritical fluid extraction of natural products and their [37] M. Khajeh, M.G. Moghaddam, M. Shakeri, Application of artificial neural network
characteristic times, J. Supercrit. Fluids 66 (2012) 73–79. in predicting the extraction yield of essential oils of Diplotaenia cachrydifolia by
[25] Z. Zeković, O. Bera, S. Đurović, B. Pavlić, Supercritical fluid extraction of coriander supercritical fluid extraction, J. Supercrit. Fluids 69 (2012) 91–96.
seeds: kinetics modelling and ANN optimization, J. Supercrit. Fluids 125 (2017) [38] J. Azmir, I.S.M. Zaidul, K.M. Sharif, M.S. Uddin, M.H.A. Jahurul, S. Jinap, P. Hajeb,
88–95. A. Mohamed, Supercritical carbon dioxide extraction of highly unsaturated oil from
[26] T. Fornari, A. Ruiz-Rodriguez, G. Vicente, E. Vázquez, M.R. García-Risco, Phaleria macrocarpa seed, Food Res. Int. 65 (2014) 394–400.
G. Reglero, Kinetic study of the supercritical CO 2 extraction of different plants from [39] S. Kuvendziev, K. Lisichkov, Z. Zeković, M. Marinkovski, Artificial neural network
Lamiaceae family, J. Supercrit. Fluids 64 (2012) 1–8. modelling of supercritical fluid CO2 extraction of polyunsaturated fatty acids from
[27] N. Gañán, E.A. Brignole, Supercritical carbon dioxide fractionation of T. minuta and common carp ( Cyprinus carpio L.) viscera, J. Supercrit. Fluids 92 (2014).
S. officinalis essential oils: experiments and process analysis, J. Supercrit. Fluids 78 [40] G. Sodei fian, S.A. Sajadian, N. Saadati Ardestani, Optimization of essential oil ex-
(2013) 12–20. traction from Launaea acanthodes Boiss: utilization of supercritical carbon dioxide
[28] N. Babovic, S. Djilas, M. Jadranin, V. Vajs, J. Ivanovic, S. Petrovic, I. Zizovic, and cosolvent, J. Supercrit. Fluids 116 (2016) 46–56.
Supercritical carbon dioxide extraction of antioxidant fractions from selected [41] Y. Yoon, G. Swales, T.M. Margavio, A comparison of discriminant analysis versus
Lamiaceae herbs and their antioxidant capacity, Innov. Food Sci. Emerg. Technol. artificial neural networks, J. Oper. Res Soc. 44 (1993) 51–60.
11 (2010) 98–107.

336

Das könnte Ihnen auch gefallen