Contents lists available at ScienceDirect

Industrial Crops & Products

journal homepage: www.elsevier.com/locate/indcrop

Application of an artificial neural network model for the supercritical fluidextraction of seed oil from Argemone mexicana (L.) seeds

Bhupendra Suryawanshi⁎, Bikash MohantyDepartment of Chemical Engineering, Indian Institute of Technology Roorkee, Uttarakhand, 247667, India

A R T I C L E I N F O

Keywords:Supercritical fluid extractionArgemone mexicana (L.) seedsArtificial neural networkFatty acidsOptimization

A B S T R A C T

In this work, a three-layer artificial neural network (ANN) was investigated to predict the cumulative extractionyield (CEY) of seed oil during the supercritical fluid extraction (SFE) of Argemone mexicana (L.) (A. Mexicana)seeds. The effect of five extraction parameters (i.e. temperature (°C), pressure (bar), particle size (mm), flow rate-CO2 (g/min) and co-solvent % (% of flow rate-CO2)) on the CEY of A. Mexicana seed oil was investigated andfound that all five extraction parameters have significant effect in the order (co-solvent % > pressure >particle size > flow rate-CO2 > temperature) on it. The ANN model was generated, using experimental dataand for this, a trainable, feed-forward-back-propagation (FFBP) network was used to predict the CEY of A.Mexicana seed oil with an acceptable level of accuracy. From the best performing ANN models, a singlemathematical equation was developed that can be used in predicting the CEY. In this regard, by changing, thenumber of neurons in the hidden layer and the algorithms, different networks were formed and compared withthe evaluation of networks accuracy in CEY. Finally, the six neurons in the hidden layer, using the Levenberg-Marquardt algorithm, found to be the most suitable network. The value of average-absolute-relative-deviation-percentage (AARD%) (3.33%), mean-square-error (MSE) (0.0038) and the coefficient of determination(R2=0.9838) showed that the ANN-FFBP [5-6-1] model is a better option for predicting the CEY. Furthermore,the fatty acids analysis was done by using gas chromatography (GC) which confirmed presence of leading fattyacids (C16:0, C16:1, C17:1, C18:0, C18:1n9c, C18:2n6c, and C20:5n3) in the extracted oil of A. Mexicana seeds.

1. Introduction

Separation and purification technology play a crucial role in che-mical processing. Supercritical fluid extraction (SFE) technology is themost attractive alternative to the conventional extraction technologiessuch as Soxhlet extraction, Cold press extraction, Steam distillation etc.,since, the solvent contamination and thermal degradation occurred inthe conventional separation technologies. In addition to this, conven-tional solvents (e.g. ethanol, methanol, chloroform, petroleum ether, n-hexane, toluene etc.) used in the conventional processes are unafford-able because of their high consumption of solvent, energy and the re-covery of the organic solvent from the final product. In addition, theseorganic solvents are also very harmful to environment and humans(Ghoreishi et al., 2009). SFE technology is getting more attention fromresearchers due to some unique properties of supercritical fluids (SFs)such as gas-like viscosities, liquid-like densities, low surface tension andhigh selectivity. Low polarity which is the only demerit of supercriticalcarbon dioxide (SC-CO2), that can be enhanced by adding some highpolar organic co-solvents which also increases its solubility power

(Ghoreishi and Heidari, 2013). SC-CO2 could be the most desirable andaffordable solvent in place of conventional solvents because the SC-CO2

is non-flammable, non-toxic, easily recoverable and inexpensive(Suryawanshi and Mohanty, 2018).

The physical phenomena of the SFE process has been modeled invarious aspects such as heat transfer phenomena based models(HTPBM’s), mass transfer phenomenon based models (MTPBM’s) andempirical models (EM’s) for example hot ball diffusion model (Bartleet al., 1990) under HTPBM category, shrinking core model (Roy et al.,1996; Goto et al., 1996) and desorption-dissolution-diffusion model(Lee et al., 1986; Reverchon et al., 1993; Sovova et al., 1994;Goodarznia and Eikani, 1998) under MTPBM’s category and empiricaldesorption models (Esquível et al., 1999; França de et al., 1999; Coceroand Garcia, 2001) under EM’s category.

Recently, SFE processes have appeared as an attracting field for theapplication of artificial neural network (ANN) primarily because of thehigh sensitivity of the mathematical models to small variations in theprocess variables. ANN has been introduced as a predictive tool foroptimizing the operating parameters during the SFE of different natural

https://doi.org/10.1016/j.indcrop.2018.06.057Received 6 March 2018; Received in revised form 9 June 2018; Accepted 14 June 2018

⁎ Corresponding author.E-mail addresses: surya_bhupen@yahoo.com (B. Suryawanshi), bmohanty.iitr@gmail.com (B. Mohanty).

Industrial Crops & Products 123 (2018) 64–74

0926-6690/ © 2018 Elsevier B.V. All rights reserved.

T

products by various researchers (Fullana et al., 2000; Kamali andMousavi, 2008; Shokri et al., 2011; Eslamimanesh et al., 2011;Lashkarbolooki et al., 2011; Khajeh et al., 2012; Ghoreishi and Heidari,2013; Lashkarbolooki et al., 2013; Pilkington et al., 2014). Recently,Ghoreishi and Heidari (2013) and Pilkington et al. (2014) have com-pared both RSM and ANN models during the SFE of Epigallocatechin-3-gallate from green tea and Artemisinin from Artemisia annua respec-tively and confirmed the accuracy of ANN. ANN models are receivingmore attention from researchers because of its reliable relationshipbetween inputs and outputs data’s of the complex and nonlinear sys-tems.

In this work, development of an ANN model to indicate how theCEY of A. Mexicana seed oil from its seeds might be optimized. An ANNmodel which consists of MLP (multi-layer perceptron) and FFBP (feedforward back propagation) network was developed to simulate the SFEof A. Mexicana seed oil, using SC-CO2 as the solvent. From the devel-oped ANN model, an equation was developed that is based on fiveoperating parameters including temperature (°C), pressure (bar), par-ticle size (mm), flow rate-CO2 (g/min) and co-solvent % that predict theCEY of A. Mexicana seed oil. The developed model shows its suitabilityfor predicting A. Mexicana seed oil recovery by analyzing their coeffi-cient of determination (R2), mean-square-error (MSE), and average-absolute-relative-deviation-percentage (AARD%) from experimentaldata. Furthermore, the chemical compositions of seed oil, obtainedfrom SFE process was analyzed and quantified using gas chromato-graphy (GC).

2. Experimental

2.1. Plant materials and chemicals

The seeds (10 kg) of A. Mexicana were collected from the authorizedsupplier of Ayurvedic raw materials (Herbal Automation,Hanumangarhi, Kankhal, Distt.- Haridwar State – Uttarakhand, India).The seeds were cleaned and dried under sunlight for at least 2–3 daysand then all seeds were shredded in a domestic grinder (Bajaj, India).The particle size distribution of the shredded seeds into their respectiveparticle sizes was done by certified test sieves & vibratory shaker(Endecotts Ltd., and Octagon 200, London, England). The shreddedseed particles according to their respective sizes have been collected inthree ranges as shown in Table S1 (as Supplementary material) and forconvenience, based on sieve openings, a ‘fraction of particles’ betweentwo successive sieves was assigned the average particle diameter.

Carbon dioxide (CO2) with 99.99% purity was purchased fromSigma gases ltd., India, in pressurized deep long tube cylinders. Ethanol(used as a co-solvent), sulphuric acid (98.50%), n-hexane, toluene, so-dium chloride, methanol, potassium bicarbonate and anhydrous sodiumsulfate (in powder form) were purchased from Merck Ltd., Mumbai,India. All other standard chemical reagents were of analytical and HPLCgrade. The water (ultrapure), obtained from Milli-Q system (Millipore,Bedford, MA, USA) was used in all solutions for analysis purposes.

2.2. SC-CO2 extraction

A schematic diagram of the experimental setup of the SC-CO2 ex-traction unit ((Suryawanshi and Mohanty, 2018) is shown in Fig. S1 (asSupplementary material). Liquid CO2 is stored in a long-standing cy-linder (a) and is passed through a micro filter (b) which filters the CO2

by entrapping the dust or other particles before getting to enter into thepiping system and maintain its purity. Then filtered CO2 is chilled downbelow 5 °C through a chiller (c). A high pressure reciprocating feedpump (d) is used to attain the critical pressure of the chilled CO2 andthen transfer it to a mixer (i) where, there is provision to add organicco-solvent (in our case ‘Ethanol’) which is pumped from a co-solventbottle (k) through a co-solvent pump (j). Then this modified CO2 passesthrough a preheater (e) where, preheating is done to attain the critical

temperature, before entering into the extractor (f). This extractor is astainless steel extraction cell of a volume of 1 L. A cylindrical basketmade of stainless steel with 17 cm height and 7.5 cm in diameter is usedto hold the material (50 g) to be extracted. To fill up the cylindricalbasket, glass beads (5 mm dia.) and glass wool were used as an inert andpacking material. Initially, the fixed bed consists of glass beads up to3.0 cm height then a layer of glass wool having a thickness of 0.5 cmthen again spread the layer of 3.0 cm of glass beads over the glass wool.A uniform distribution of SC-CO2 could be possible through this ar-rangement. Now, 50 g of seed, particles were placed above the previousarrangement followed by glass beads and glass wool with the samepattern as previously arranged. This packing of the fixed bed alsoprevents the flow of solid particles with SC-CO2. The system pressure ismonitored and controlled by an automatic back-pressure-regulator(BPR) (g). Precipitation of oil from the SC-CO2 occurs in cyclone se-parator (h) and then oil is collected from the bottom to a sampler flask(i). Ethanol (co-solvent) is removed from the extracted oil by vacuumeffect-rotary evaporator and then after achieving room temperature,the weight of oil is measured gravimetrically. The total, CEY was thendetermined from the sum of all oil samples extracted. The same pro-cedure was adopted for all experiments.

The extraction yield (ηTotal) was determined through Soxhlet ex-traction process by following the standard procedure given by AOAC(official methods of analysis) as the weight percentage of the quantityof oil (woil) obtained from dried seeds divided by the weight of A.Mexicana seeds (wseeds) and used as a reference for SFE later on.

=η wt ww

( . %) 100 *Totaloil

seeds (1)

2.3. Gas chromatographic analysis

The fatty acid analysis of SFE extracts of A. Mexicana seed oil wascarried out using a gas chromatograph-Thermo scientific of AgilentTechnology equipped with an FID and an HP-88 fused silica column(30m length× 0.25mm internal diameter× 0.25 μm film thickness).The injector and detector temperature were maintained constant at250 °C while the initial oven temperature was maintained at 120 °C for1min. Now, the oven temperature with a ramp rate of 5 °C/min wasincreased from 120 °C to 145 °C and maintained constant here at 145 °Cfor 2min. The temperature was again increased with a ramp rate of2 °C/min from 145 °C to 220 °C and maintained constant here at 220 °Cfor 2min. The sample (20 μL) was injected into the GC (without anyfurther dilution) using the split flow mode. The percentage of thecompounds was calculated by area normalization method. The fattyacids of oil were identified by comparing with the chromatogram ofstandard chemical (Supelco 37 component FAME mix, Supelco,Bellefonte, PA).

2.4. Data sets

The experimental data used for ANN modeling are presented inTable 1. This experimental data set was randomly divided into threesets such as a training set (32), validation (7) and testing set (7) re-spectively. The training data set (32) was used to adjust the parametersof the model for performing the computation of the network whiletesting data set (7) was used to calculate its estimation power and onother hand validation data set (7) was used to ensure robustness of thenetwork parameters. The validation is done to overcome from theproblem of overtraining probability because, if an ANN is learned toowell from the training data set, the rules might not fit as well for the restof the cases in the data set. To avoid this “overfitting” phenomenon, thetesting stage was used to control error; when it increased, the trainingwas stopped (Song et al., 2004).

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

65

2.5. Artificial neural network (ANN)

The ANN modeling is based on the working of the natural neuralnetworks of the human brains (Poularikas, 2002). In an ANN model,each neuron added the weighted inputs from different paths and thenapplies a transfer function to the addition; the resulting value is thendirected through outgoing paths to other neurons. A series of layers wasformed with neurons that are called multilayer perceptron (MLP)(Ghoreishi and Heidari, 2013). A multilayer perceptron (MLP) model isshown in Fig. 1 which is to be developed in MATLAB software version'Mathworks, Inc., 2013a'. It consists of an input layer which representsthe number of input parameters (i.e. temperature, pressure, particlesize, flow rate-CO2 and co-solvent %) as in our case, one hidden layerwith ‘n’ number of neurons which is to be decided by hit and trial ap-proach and one output layer, denoting CEY as the response. Eachneuron of the input layer, is connected to one or more neurons of thehidden layer that represents the nonlinear activation function and these

neurons are then connected to neurons of output layer through alearning algorithm function (Pilkington et al., 2014).

A neuron’s output was evaluated by applying a transfer function to a

Table 1Experimental data (training, validation and testing sets), observed and predicted extraction yield of seed oil of A. Mexicana by SFE.

Parameters Response

Run Temperature (oC) Pressure (bar) Particle size (mm) Flow rate-CO2 (g/min) Co-solvent (% of flow rate-CO2) Observed yield Predicted yield

Training Set1 80 200 0.5 10 5 0.1289 0.13312 80 275 0.75 10 5 0.3470 0.36043 60 350 0.75 10 5 0.3430 0.34004 60 275 0.75 10 0 0.2500 0.25395 80 275 1 5 5 0.3000 0.29886 100 275 0.75 5 5 0.2620 0.30787 80 200 0.75 10 10 0.3025 0.30198 80 200 0.75 15 5 0.2225 0.20179 60 200 0.75 10 5 0.2800 0.256610 100 350 0.75 10 5 0.3700 0.370711 80 275 0.5 5 5 0.1899 0.190912 80 275 0.75 10 5 0.3743 0.360413 80 200 1 10 5 0.2060 0.204814 100 200 0.75 10 5 0.1474 0.153115 80 350 0.75 15 5 0.3120 0.313516 80 275 0.75 15 10 0.3598 0.360117 80 350 0.75 10 10 0.3986 0.368218 80 200 0.75 10 0 0.0660 0.061619 80 275 0.5 10 0 0.1450 0.150320 100 275 1 10 5 0.2575 0.259121 80 275 0.75 10 5 0.3610 0.360422 60 275 0.75 10 10 0.3295 0.331423 80 275 1 10 0 0.2000 0.201524 60 275 0.5 10 5 0.1845 0.176825 60 275 0.75 15 5 0.3200 0.309126 80 275 1 15 5 0.3385 0.336227 100 275 0.75 15 5 0.3620 0.346028 80 275 0.75 10 5 0.3664 0.360429 100 275 0.75 10 10 0.4240 0.408430 80 350 0.75 10 0 0.2800 0.280231 80 200 0.75 5 5 0.1695 0.169532 100 275 0.75 10 0 0.1710 0.1700

Validation set33 80 275 0.75 10 5 0.3720 0.360434 80 350 0.75 5 5 0.3800 0.337435 80 275 0.75 5 10 0.4031 0.411736 80 275 0.75 10 5 0.3756 0.360437 80 350 1 10 5 0.4050 0.368638 80 350 0.5 10 5 0.2250 0.223039 100 275 0.5 10 5 0.2626 0.2605

Testing set40 80 275 0.75 15 0 0.2450 0.253241 60 275 1 10 5 0.3425 0.343642 80 275 0.75 5 0 0.1595 0.159943 80 275 1 10 10 0.3800 0.386244 80 275 0.5 15 5 0.2025 0.204545 60 275 0.75 5 5 0.2900 0.345246 80 275 0.5 10 10 0.2377 0.2441

Fig. 1. Structure of the developed ANN with five inputs and one output.

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

66

weighted summation of its input and resulted an output, which will beused as input for other neurons, as indicated in Eq. (2).

∑= ⎛

⎝⎜ + ⎞

⎠⎟

=−

−

γ F W γ βjk ki

N

ijk i k jk1

( 1)

k 1

(2)

Where, γjk is the neuron ′j s output from ′k s layer and βjk is the biasweight from neuron j in layer k, Wijk is the weight connection and Fk isthe nonlinear activation transfer functions. The following steps werefollowed during ANN modeling of the system:

Step 1: Firstly, the available experimental data were partitioned intotraining, validation and testing subsets. The ANN model was developedusing ‘training subset’ data while its validation and examination ofperformance was done using ‘testing subset’ data.

Step 2: During, the training of the subset data, randomly selectedweight matrices were adjusted to minimize the back propagation error.

Step 3: Now, on the basis of hit and trial approach, the number ofhidden layers and the number of neurons in each layer were determinedto form suitable network architecture.

Step 4: The five input parameters (i.e. temperature, pressure, par-ticle size, flow rate-CO2 and co-solvent %) and one output parameter(i.e. CEY) were selected to establish the ANN system. Initially, thenumber of neurons in the hidden layer would be equal to one and thenincreases by one till optimum condition is achieved.

Step 5: Now, the types of ‘in-build transfer function’ for each stagemust be defined, using hit and trial approach. Different types of ‘in-builttransfers functions’ (i.e. radial, linear, hyperbolic tangent sigmoid,logarithmic sigmoid etc.) are available in the ‘nntool’ of MATLAB.

Step 6: The number of hidden layers were optimized. In our case,only one hidden layer was considered since (Cybenko, 1989) has re-ported that an ANN with only one hidden layer could approximate al-most any type of nonlinear relation. The selection of number of neuronsin the hidden layer is totally based on hit and trial basis. A few numberof neurons can produce a network with low precision while a highernumber of neurons can lead to over fitting and bad quality of inter-polation (Lashkarbolooki et al., 2013).

Step 7: In step 6, the number of, hidden layers and neurons in thehidden layers were optimized on the basis of some of the statisticalparameters (e.g. average-absolute-relative-deviation-percentage (AARD%), mean-square-error (MSE) and correlation-coefficient (R2)) whichare given below;

∑=−

×=

AARDn

Y YY

(%) 1 ([ ]

) 100i

niExp

iANN

iExp

1 (3)

∑= −=

MSEn

Y Y1 ( )i

n

iExp

iANN

1

2

(4)

=∑ − − ∑ −

∑ −= =

=

RY Y Y Y

Y Y( ) ( )

( )in

iExp

in

iExp

iANN

in

iExp

2 12

12

12 (5)

Where; n = Number of data points.YiANN = The ith predicted CEY

through ANN model.YiExp = The ith observed CEY through experiment.

Y = The average value of the experimental data of CEY.

2.6. Development of an ANN-based equation from weights and biases

The A. Mexicana seeds have been extracted through various con-ventional extraction techniques such as Soxhlet extraction method,using n-hexane (Mishra et al., 2009) and petroleum ether (Singh andSingh, 2010) as solvent, Percolation method (Bhattacharjee et al.,2010) using methanol as solvent and through a mechanical expeller(Pramanik et al., 2012). However, no studies have been reported forthe extraction of A. Mexicana seeds using SC-CO2 during the SFE.Therefore, a simple equation was developed to predict the extractionyield of A. Mexicana seed oil. Fig. 2 shows the typical arrangement ina single-hidden layer configuration of ANN model. The figure shows

that the model has five input variables/parameters (e.g., temperature,pressure, particle size, flow rate-CO2 and co-solvent % and theseparameters were coded here as ‘T’, ‘P’, ‘PS’, ‘FR’, and ‘CoS’ respec-tively) and one output variable (e.g. CEY). The ‘w’ and ‘b’ are theweights and biases respectively which were assigned randomly by thenetwork. In this work, different number of neurons was tested in theconfiguration of single hidden-layer model. In Fig. 2, the transferfunction associated with the hidden and output layers were TANSIG(shown with a curve in hidden layer) and PURELIN (shown with astraight line) respectively.

The development procedure of an ANN-based equation requires theexperimental extraction data (input and output). The steps are givenbelow;

Step 1: Normalization of the input experimental dataThe normalization of input experimental data (input layer) for ANN

was done using Eq. (6) (given below) to scale-up the inputs and outputso that they fall in the range of ‘+1’ to ‘–1’, corresponding to thehighest and lowest values respectively.

= − −−

+Y Y Y X XX X

Y( )*( )( )norm

actualmax min min

max minmin (6)

Where, Ynorm is the normalized value of Xactual. The values of Ymax andYmin are ‘–1’ and ‘+1’ respectively. Xactual, Xmin and Xmax are the actual,minimum and maximum values of the independent parameters of in-terest.

Step 2: Weight assignments at the hidden layerAfter normalizing the all input parameters, the assignments of

weights and biases to the normalized parameters were performed asshown in Fig. 2. These weights (‘w’) and biases (‘b’) were generatedautomatically by the network. The linear relationship (Eq. (7)) wasdeveloped for the above phenomenon.

∑= + + +

+ + +=

A w T w P w PS w

FR w CoS b

( *( ) *( ) *( )

*( ) *( ) )

l i ni

j

l i T norm n l i P norm n l i PS norm n l i FR

norm n l i CoS norm n l i

, ,1

, , , , , , , ,

, , , (7)

Where, ‘wl i T, , ’ ‘wl i P, , ’ ‘wl i PS, , ’ ‘wl i FR, , ’ and ‘wl i CoS, , ’ represent the weightsassigned to the operating parameters temperature (‘T’), pressure (‘P’),particle size (‘PS’), flow rate-CO2 (‘FR’) and co-solvent % (‘CoS’) re-spectively. The number of neurons varies from ‘i’ to ‘j’ while the numberof hidden layer was denoted by ‘l’. (Tnorm)n, (Pnorm)n, (PSnorm)n,(FRnorm)n and (CoSnorm)n are the nth values of the normalized T, P, PS,FR, and CoS respectively. Al,i,n represents the nth sum of the weightednormalized variables.

Step 3: Transfer function for the hidden layerFollowing Fig. 2, the third step in developing the ANN-based model

equation requires the mathematical expression of the TANSIG transferfunction which was used in hidden layer. The expression for theTANSIG function is given as:

= =+ −

−TANSIG A BA

( ) 2(1 exp( 2* )

1l i n l i nl i n

, , , ,, , (8)

Eq. (8) is only applicable for the single hidden layer network.Step 4: Transfer function for the output layerStill following the Fig. 2, at the output layer, the PURELIN transfer

function was applied with the weight and biases which were alsogenerated by the network. They are then assigned to the previousvariable (Bl,i,n) as shown in Eq. (9).

∑= +=

a w B b( * )i

j

o i l i n o01

, , ,(9)

Where, ‘wo,i’ refers to the weight at the output layer (o) attributed toeach neuron (i); ‘Bl.i,n’ is the previously defined value of the TANSIGtransformed variable, associated with the sum of the nth normalizedinput variable at the last hidden layer (l); ‘bo’ is the bias at the outputlayer. ‘ao’ is the normalized final output.

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

67

Step 5: De-normalization of the normalized outputTo obtain the actual value of the output, there is a need to de-

normalize the output obtained in Eq. (9). At this step, the normalizedoutput (an) was de-normalized, using Eq. (6).

3. Result and discussion

3.1. Optimization of FFBP-ANN configuration

In this work, different operating conditions of temperature (60, 80,100 °C), pressure (200, 275, 350 bar), particle size (0.5, 0.75, 1.0 mm),flow rate-CO2 (5, 10, 15 g/min) and co-solvent % (0, 5, 10% of CO2

flow rate) were used in FFBP-ANN (feed-forward back-propagation ar-tificial neural network) model for the prediction of CEY (g oil/g seedmaterial). Firstly, 32 data points were used for training the ANN model.The correlation-capability (CC) of the developed ANN model wasevaluated on the basis of the total 46 experimental data points. Theobtained results, which are based on trial and error procedure lead tofind out 6 neurons as the optimal number of neurons in the hidden layeras shown in Fig. 3 as an optimized FFBP-ANN [5-6-1] model. The re-sults of the obtained AARD %, MSE and R2 values for the differentnumbers of the hidden neurons were given in Table 2. The network wastrained several times by applying randomly generated initial values ofthe network parameters (i.e. weight and bias coefficients) which couldaffect the values of AARD % and MSE and finally resulted in the bestvalues of AARD %, MSE and R2 as shown in Table 3, which able anyoneto reproduces every used data points in the present work. As mentionedearlier (Section 2.5), the network with least error measures (i.e. AARD% and MSE), and suitable regression coefficient (R2), was chosen as theoptimal network configuration. Finally, from the study, it can be un-derstood that a FFBP-ANN model with the configuration of 5× 6×1leads to the minimum level of error (i.e. AARD % & MSE) during thecorrelation of experimental data. Therefore, it can be concluded thatproposed ANN model with the number of neurons, 5, 6 and 1 in theinput, hidden and output layer respectively could be considered as thebest ANN architecture. The obtained optimized model was named asFFBP-ANN [5-6-1]. TANSIG and PURELIN were used as the activationfunctions in the hidden and output layer respectively. The ANN wastrained through over 1000 epochs with error back propagation (EBP)training. The correlation between the experimental and ANN predictedCEY for training, validation, testing and overall data sets are shown inFig. 4. The perfect fit (ANN model prediction equal to experimentaldata) was shown by solid line. A close proximity of the best linear fit to

the perfect fit has been observed, as shown in Fig. 4., which confirms agood correlation between the experimental CEY and the FFBP-ANN [5-6-1] model CEY.

3.2. Performance of ANN-FFBP models

The performance of the developed ANN-FFBP [5-6-1] model wasjudge by comparing its performance with some other tested ANN-FFBPmodels on the basis of some statistical parameters (e.g. AARD %, MSE,R2 and NSEC) as shown in Fig. 5. From the Fig. 5, it is clear that theANN-FFBP [5-6-1] model with single hidden layer performed well andresulted minimum AARD % (‘3.33%’ as shown in Fig. 5(a)), minimumMSE (‘0.0038’ as shown in Fig. 5(b)), maximum values of R2 (‘0.9835’as shown in Fig. 5(c)) and NSEC (‘0.9664’ as shown in Fig. 5(d)) amongall the tested ANN-FFBP models. The AARD% shows the spreading ofdata from the central point which is the lowest among all the testedANN-FFBP models for the configuration ANN-FFBP [5-6-1]. Similarpattern was also observed with MSE as shown in Fig. 5(b). However, incomparison, AARD is higher than MSE for all models. The highest valueof NSEC (‘0.9664’ as shown in Fig. 5(d)) indicates the highest efficiencyof the developed ANN-FFBP [5-6-1] model over the other tested models(ANN-FFBP [5-7-1] (‘0.9461’), ANN-FFBP [5-8-1] (‘0.9418’), ANN-FFBP[5-9-1] (‘0.9403’)). Similarly, the correlation coefficient (R2) was em-ployed to judge the performances of the models. The closer the corre-lation coefficient (R2) to one, the better the performance of the model.The Fig. 5(c) shows that the ANN-FFBN [5-6-1] continues to exhibit thebest performance.

Fig. 2. Single hidden layer based FFBPN to predict the CEY of the A. Mexicana seed oil in the SCFE process.

Fig. 3. Optimized FFBP-ANN network to predict the CEY of the A. Mexicana seed oil in the SCFE process.

Table 2Results of topological studies to find the optimal FFBP-ANN configuration.

Hidden neuron AARD % MSE R2

1 14.8620 0.0448 0.85002 9.6779 0.0350 0.94143 7.9362 0.0156 0.96584 10.5470 0.0106 0.92905 9.3650 0.0100 0.92186 3.3292 0.0038 0.98357 6.1287 0.0099 0.97598 7.0714 0.0241 0.97279 4.3730 0.0540 0.9728

The bold values indicate the condition, achieved by the optimal FFBP-ANNconfiguration. The purpose to make bold of this row is just to highlight it as anachieved optimal condition and nothing else.

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

68

Table 3Weights and biases of the trainable FFBP-ANN model for the optimal architecture.

Neuron Hidden layer Output layer

Weights (wij)a Biases Weight Bias

T P PS FR CoS % bj (Wjk)b bk

1 −0.2913 0.1083 4.4960 0.4376 1.2266 0.8689 0.4038 −0.50762 0.1623 −1.1605 −2.1718 −1.0356 −1.3374 −3.4049 −0.34913 1.0228 −3.1123 1.6703 −0.5529 −1.3596 −2.1165 −0.58724 −2.0045 2.3014 −0.3659 −0.5558 −1.0969 −0.3537 −0.14585 0.7847 0.8025 −0.0588 −1.9313 −0.3303 0.2738 0.18276 −1.0731 −1.5492 1.7209 −0.4544 −0.2518 −2.8087 0.1479

a Weight connection from the input layer to hidden layer.b Weight connection from the hidden layer to output layer.

Fig. 4. The scatter plots that compare the experimental data (target) against the ANN predicted data for (a) training, (b) validation, (c) testing and (d) all data.

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

69

3.3. Development of an equation for optimized ANN-FFBP [5-6-1] model

As discussed in Subsection 2.6, the steps (1–5) for developing theANN-based equation model are to be followed. In this direction, thenormalization of the independent parameters (T, P, PS, FR and CoS)was done as follows;

= −T T0.05* 4norm (10)

= −P P0.0133* 3.667norm (11)

= −PS PS4* 3norm (12)

= −FR FR0.2* 2norm (13)

= −CoS CoS0.2* 1norm (14)

Where, Tnorm, Pnorm, PSnorm, FRnorm and CoSnorm are the normalized va-lues of independent parameters.

Weights and biases were assigned to the normalized parameters atthe hidden layer as shown in Fig. 2. The listed weights and biases inTable 3 were assigned for each of the input variables, based on the sixneurons in ANN-FFBP [5-6-1]. The values (A111 to A161) were calculatedfor each experimental conditions (run 1 to run 46), using Eq. (7). Stillfollowing the lead provided in the Subsection 2.6 and Fig. 2, the nextstep is the ‘TANSIG’ transfer function for ‘Blin’ at the hidden layer. Theexpression for the transfer function was expressed in Eq. (8) throughwhich the B111 to B161 were determined for each aggregate values of(A111 to A161). Next step is to determine the output using PURELIN

transfer function using Eq. (9). Now, the resulting expression was de-normalized, as described in the Subsection 2.6. The final expression forCEY, after de-normalization, was shown in Eq. (15).

= − − −

+ + +

Y B B B B

B B

0.072284* 0.062484* 0.105103* 0.026097*

0.0327087* 0.026476* 0.154136CEY 1 2 3 4

5 6 (15)

Where, Blin is given as Bi for each of six neurons in ANN-FFBP [5-6-1], asdefined in Eq. (8). Eq. (15), is the final expression of the ANN-basedequation which was used to see the effect of individual effect of para-meters on CEY.

3.4. Correlation analysis (CA)

The direction and strength of a linear relationship between any twovariables can be explained by some probability and statistic theory(Samani et al., 2007). In this regards, a coefficient (pearson-product-moment-correlation-coefficient (PPMCC) (rx,y)) was applied to thenature of the data to explain the degree of correlation of variables. ThePPMCC (rx,y) was obtained by dividing the co-variance of any of twovariables by the product of their standard deviations and it was used tomeasure the strength of the linear relationship which is defined as:

= =− −

r COV x yσ σ

E x μ y μσ σ

( , ) (( )( ))x y

x y

x y

x y,

(16)

Where; x,y, are variables. μx, μy are expected values of the variables andσx, σy are standard deviations of the variables (x,y). The PPMCC (‘rx,y’)

Fig. 5. A comparison of different ANN-FFBPN models on the basis of some statistical parameters (AARD%, MSE, R2 and NSEC).

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

70

can take a range from ‘+1’ to ‘–1’. A zero (‘0’) value of it indicates anon-association relationship between the two variables. A greater thanzero (‘0 < value’) value indicates a positive association, that is, as thevalue of one variable increase, so does the value of the other variablewhile a less than (‘0 > value’) value indicates a negative association;means as the value of one variable increases, the value of the othervariable decreases. The results obtained from the correlation analysis(CA) of the variable for the present SFE system were shown in Table 4.Since this system has become a multivariable system, therefore, thecorrelation coefficients produce a matrix in which an element (‘αij’)represents the relation between the ith and jth variable. It can be un-derstood from this matrix that co-solvent %, pressure, particle size, flowrate-CO2 and temperature have a significant effect in the order (CoS% > P > PS > FR > T) on CEY.

3.5. Sensitivity analysis (SA)

Sensitivity analysis (SA) was carried out to determine, how theuncertainty in the output of the developed FFBP-ANN model can berelated to different sources of uncertainty in the experimental datainput. In this study, the possible interaction of chosen parameters wasevaluated to determine their performance with respect to each other.

Therefore, five groups (i.e. one, two, three, four and five variables)were formed and investigated separately by the achieved optimal FFBP-ANN [5-6-1] model. The results of the analysis were shown in Table 5.From the Table 5, it is clear that CoS% with MSE=0.114 is the mosteffective parameter in the group of one variable. The other parameters(PS, FR, P, and T) have also shown the significant effect withMSE=0.121, 0.23, 0.233, and 0.263 respectively. As shown in Table 5,the value of MSE significantly decreased when CoS%, FR, P and PS wereused in interaction with the group of two variables. The lowest values ofMSE (0.071, 0.074 and 0.097) were determined during the interactionof PS with ‘P’, PS with ‘CoS%’ and PS with ‘FR’ respectively. Othercombinations (e.g. T+ FR, P+CoS, and FR+CoS) in the group of twovariables have also shown a strong interactive effect with MSE’s 0.185,0.133, 0.110 respectively. The minimum value of MSE for the group ofthree variables was found to be ‘0.035’ with the interaction ofP+ PS+CoS% while the lower value of MSE’s (0.054) for the group offour variables were determined with the interaction of(T+P+PS+CoS%). The value of MSE was decreased drasticallyfrom 0.054 to 0.0038when FR was involved in interaction with othervariables (T, P, PS, FR, and CoS%) in the group of five variables.

3.6. Comparison between experimental and ANN output

A comparison between experimental CEY and FFBP-ANN (5-6-1)model predicted CEY was shown in Fig. 6(a–e). From the Fig. 6(a–e), itis clear that an excellent agreement between experimental data andANN results was observed. The optimal conditions were 90 °C, 275 bar,0.81mm, 8.75 g/min and 10% for temperature, pressure, particle size,flow rate-CO2, and co-solvent % respectively. From the Fig. 6(a–e) itcan be understood clearly that temperature, pressure, particle-size, flowrate-CO2 and co-solvent %, have shown a significant effect on CEY of A.Mexicana seed oil.

3.6.1. Identification of most influencing parametersFrom the Fig. 6(a) it is clear that extraction yield increases up to

85 °C and then it starts to decrease due to its weak relationship betweentemperature and extraction yield due to ‘retrograde solubility inter-ference’ (Nei et al., 2008). Mukhopadhyay (2000) has explained theretrograde solubility interference by the relative influence of the den-sity effect and the volatility effect means an isobaric increase in tem-perature decreases density of the supercritical fluid solvent (in our caseit is SC-CO2) and hence decreases the solubility by the density effect. Onthe other hand, the same increase in temperature increases the volati-lity of the solute and hence increases the solubility by the volatilityeffect. The increase in extraction temperature from 60 to 85 °C increasesthe solubility due to solute vapor pressure enhancement and after 85 °Creduces the solubility due to the decrease in solvent density (Dökeret al., 2004). From Fig. 6(b), it is very clear that the CEY of seed oilincreased significantly with increasing pressure, this is due to the factthat increased pressure, enhance the SC-CO2 density which improvesthe solubility of free fatty acids (Nei et al., 2008). The further increasein the pressure reduces the SC-CO2 diffusivity giving rise to a reductionin yield. The negative effect of pressure (> 280 bar) on CEY of A.Mexicana seed oil could be explained through the physiochemicalproperties of the supercritical fluid. According to this, the excessiveincrement in pressure may cause fluid viscosity to increase which de-creases the supercritical fluid diffusivity resulting in a lower extractionyield (Wu et al., 1991; Kazan et al., 2014; Suryawanshi and Mohanty,2018). Some researchers (Lamb et al., 1981; Wu et al., 1991 and Savageet al., 1995) have also explained it through transition state theory.According to this theory, when pressure increases the activation volumedecreases leading to a decrease in the values of diffusivity, as it is acontributory factor for activation volume, leading to a decrease in ex-traction yield. From Fig. 6(c) it is clear that extraction yield increaseswith the particle sizes from 0.5 to 0.8mm and thereafter decreases from0.8 to 1.0mm. It may be due to the initial high pressure difference

Table 4Pearson product-moment correlation matrix of variables.

T P PS FR CoS % Y

ρ = 1.0000 0.0000 0.0000 0.0000 0.0000 0.0347 T0.0000 1.0000 0.0000 0.0000 0.0000 0.4975 P0.0000 0.0000 1.0000 0.0000 0.0000 0.3565 PS0.0000 0.0000 0.0000 1.0000 0.0000 0.0870 FR0.0000 0.0000 0.0000 0.0000 1.0000 0.5513 CoS %0.0347 0.4975 0.3565 0.0870 0.5513 1.0000 Y

Table 5Sensitivity analysis (SA) of the input variables for the ANN-FFBPN [5-6-1]model.

No. Combination MSE R2 Equation

Group of one variables1 T 0.263 0.0372 y=0.0003x+ 0.29032 P 0.233 0.5100 y=0.3497x+ 0.19993 PS 0.121 0.4500 y=1.0373x− 0.01084 FR 0.23 0.0800 y=0.6194x+ 0.10635 CoS % 0.114 0.5594 y=0.9544x+ 0.0163

Group of two variables6 T+P 0.379 0.5390 y=0.4228x+ 0.17237 T+PS 0.387 0.2701 y=0.9345x− 0.0018 T+FR 0.185 0.0288 y=−0.0881x+0.30989 T+CoS % 0.280 0.5345 y=1.0842x− 0.017310 P+PS 0.071 0.6860 y=1.2623x− 0.069711 P+ FR 0.219 0.5288 y=0.7106x+ 0.074712 P+CoS % 0.133 0.7518 y=1.097x− 0.026913 PS+ FR 0.097 0.3594 y=0.7242x+ 0.068114 PS+CoS % 0.074 0.7100 y=0.9307x+ 0.019915 FR+CoS % 0.110 0.4860 y=1.0782x− 0.0205

Group of three variables16 T+P+PS 0.061 0.7363 y=0.9319x+ 0.013817 T+P+FR 0.055 0.5283 y=0.7272x+ 0.06718 T+P+CoS % 0.269 0.7686 y=0.8375x+ 0.045719 P+PS+FR 0.054 0.6792 y=0.854x+ 0.033920 P+PS+CoS % 0.035 0.8342 y=0.8801x+ 0.044621 PS+ FR+CoS % 0.142 0.7049 y=0.8185x+ 0.0553

Group of four variables22 T+P+PS+FR 0.086 0.7053 y=0.8968x+ 0.02723 T+P+PS+CoS % 0.054 0.9219 y=0.8955x+ 0.041424 P+PS+FR+CoS % 0.077 0.9127 y=0.9879x+ 0.0136

Group of five variables25 T+P+PS+FR+CoS % 0.0038 0.9835 y=1.0193x− 0.0034

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

71

(∼200 bar) across the bed, which might have forced the small seedparticles to stick together, as a result of it compaction might have oc-curred leading to reduction in the interface area significantly and mighthave directed the SC-CO2 solvent to flow through some channels alongthe bed. At particle size 0.5mm, the compaction effect is substantial(leading to reduced surface area of solid-liquid contact) which reducesup to 0.8mm particles and thus the extraction yield improves fromparticle size 0.5 to 0.8 mm. However, when particle size further in-creases to 1mm the actual surface area of contact decreases and due to

comparatively large particle diameter the intra particle diffusion pathas well as resistance also increases leading to lower extraction yield.The extraction yield increases significantly with flow rate-CO2 (asshown in Fig. 6(d)) due to the fact that at increasing the flow rate, thethickness of the film layer around the solid particles reduced and masstransfer resistance surrounding the solid particle becomes small whileat higher flow rate-CO2 (above 10 g/min) the solvent may move tooquickly through the extraction bed and exit the extractor unsaturatedcausing a declination in extraction yield (Döker et al., 2004). The

Fig. 6. Comparison of experimental and ANN-FFBPN predicted CEY data, showing the effects of temperature, pressure, particle size, flow rate-CO2 and co-solvent %.

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

72

effects of three different amounts of co-solvent (0, 5 and 10% of CO2

flow rate) were investigated. From Fig. 6(e) it could be understood thatincrement in co-solvent amount enhanced the extraction yield sig-nificantly. The co-solvent increased the yield because it increases themolar density of the solvent (SC-CO2) which favors the solubilization ofmore polar substances from the seed particles (i.e. increases the equi-librium solubility of the solute in the phase) (Chassagnez-mendez et al.,2000; Schmitt and Reid, 1986).

3.7. Gas chromatography (GC) analysis

The fatty acid analysis of SFE extracts of A. Mexicana seed oil wasdone using GC-FID of Thermo- scientific, Agilent Technology and re-sults were shown in Figs S2 & S3 (as supplementary material). Thefound fatty acids (as shown in Table 6) is the combination of saturatedfatty acids (SFAs) (Sl. nos. 1, 2, 4, 6, 8, 12, 15, 16, 20, 23) and un-saturated fatty acids (USFAs) (Sl. nos. 3, 5, 7, 9, 10, 11, 13, 14, 17, 18,19, 21, 22, 24). The results show that, four leading fatty acids (C16:0,C18:0, C18:1n9c and C18:2n6c) were present in the seed oil sample ofA. Mexicana, extracted, using SC-CO2 only while the seven leading fattyacids (C16:0, C16:1, C17:1, C18:0, C18:1n9c, C18:2n6c, and C20:5n3)were found in the seed oil sample extracted using ‘SC-CO2 plus co-solvent’. A comparison of fatty acids (FAMEs : fatty acids methyl esters)found in A. Mexicana seed oil sample (extracted by SFE with SC-CO2

only and SFE with SC-CO2 plus co-solvent) as shown in Table 6, re-vealed that the concentrations of leading fatty acids (C16:0, C16:1,C17:1, C18:0, C18:1n9c and C20:5n3) are increased during the ex-traction by SFE (SC-CO2 + co-solvent) in comparison to SFE (with SC-CO2 only) due to the enhancement in polarity of SC-CO2 because of theaddition of co-solvent (ethanol) %. The presence of these leading fattyacids has also been confirmed by Azam et al., 2005; Kumar and Sharma,2011. The extract, using a combination of ‘SC-CO2 plus co-solvent’ also

reported a larger number of fatty acids (22 (Table 6) out of 37 (Std.FAME mixture)) while SFE (with SC-CO2 only) reported less number offatty acids (20 out of 37). Though, the SFE (with SC-CO2 only) reportedless number of fatty acids but the percentage of unsaturated fatty acids(USFA=82.35%) is higher than the USFA (75.54%) found in the seedoil extracted by SFE (with SC-CO2 plus co-solvent). The higher con-centration of unsaturated FAMEs could be responsible for the lowerviscosity of trans-esterified oil which is one of the most desirable qua-lities for biofuel production. In the present case, SUPELCO 37 compo-nent FAME mixture (Std. solution for 37 fatty acids) was used to cali-brate the GC.

4. Conclusion

In this study, the SFE of A. Mexicana seeds using SC-CO2 was in-vestigated. The suggested FFBP-ANN [5-6-1] model with five numbersof input neurons, six numbers of hidden neurons and one number ofoutput neuron was found to be suitable for this study. From the findingof results and discussions, it can be seen that all five extraction para-meters have significant effect, in the order (CoS% > P > PS >FR > T) on extraction yield of A. Mexicana seed oil. A good agreementbetween the predicted FFBP-ANN results and experimental results wasseen for the chosen subsets (i.e. training, validation and testing). Fromthe results, it has been proved that FFBP-ANN model developed in thiswork could be an effective and efficient tool to predict the CEY of seedoil of A. Mexicana seeds by SFE using SC-CO2. The leading constituentof the fatty acids were Palmitic acid (C16:0), Stearic acid (C18:0), Oleicacid (C18:1n9c), Linoleic acid (C18:2n6c), Palmitoleic acid (C16:1) andArachidic acid methyl ester (C20:0) which confirm its suitability for theproduction of bio-fuel

Table 6Fatty acid compositions (%w/w) of oil extracted by SFE (with SC-CO2 only (run 04)) and (with SC-CO2 plus co-solvent (run 03)).

A Mexicana seed oil

Sl.no.

Fatty Acids SFE extracted withSC-CO2 only

SFE extracted withSC-CO2 plus co-solvent

Components % %n-hexane Make-up solvent Make-up solvent

1 Lauric acid methyl ester (C12:0) 0.12 0.132 Penta-decanoic acid methyl ester (C15:0) NF 0.253 cis-10-Penta-decenoic acid methyl ester (C15:1) NF 0.084 Palmitic acid methyl ester (C16:0) 13.46 17.845 Palmitoleic acid methyl ester (C16:1) 0.41 1.406 Heptadecanoic acid methyl ester (C17:0) 0.07 0.137 cis-10-Heptadecenoic methyl ester (C17:1) 0.02 1.328 Stearic acid methyl ester (C18:0) 3.03 4.609 Oleic acid methyl ester (C18:1n9c) 26.99 31.0810 Linolelaidic acid methyl ester (C18:2n6t) 0.15 0.2511 Linoleic acid methyl ester (C18:2n6c) 53.11 39.4412 Arachidic acid methyl ester (C20:0) 0.42 0.1213 Linolenic acid methyl ester (C18:3n3) 0.12 0.2214 cis-11-Eicosenoic acid methyl ester (C20:1) 0.12 0.0915 Heneicosanoic acid methyl ester (C21:0) 0.10 0.1316 Behenic acid methyl ester (C22:0) 0.07 0.9717 cis-11,14,17-Eicosatrienoic acid methyl ester (C20:3n3) 0.04 NF18 Erucic acid methyl ester (C22:1n9) 0.04 NF19 Arachidonic acid methyl ester (C20:4n6) NF 0.1520 Tricosanic acid methyl ester (C23:0) 0.38 0.2121 cis-13,16-Docasadience acid methyl ester (C22:2n6) 0.42 0.1922 cis-5,8,11,14,17-Eicosapentaenoic acid methyl ester (C20:5n3) 0.84 1.0823 Lignoceric acid methyl ester (C24:0) NF 0.0824 cis-4,7,10,13,16,19-Docosahexaenic acid methyl ester (C22:6n3) 0.07 0.23

USFA % 82.35 75.54SFA % 17.65 24.46USFA/SFA 4.657 3.088

NF: Not found.

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

73

Acknowledgement

The authors would like to thank for the financial support providedby Indian Institute of Technology Roorkee, Uttarakhand (India) to ful-fill this research study.

Appendix A. Supplementary data

Supplementary material related to this article can be found, in theonline version, at doi:https://doi.org/10.1016/j.indcrop.2018.06.057.

References

Azam, M.M., Waris, A., Nahar, N.M., 2005. Prospects and potential of fatty acid methylesters of some non-traditional seed oils for use as biodiesel in India. BiomassBioenergy 29, 293–302.

Bartle, K.D., Clifford, A.A., Hawthorne, S.B., Langenfeld, J.J., Miller, D.J., Robinsona, R.,1990. A model for dynamic extraction using a supercritical fluid. J. Supercrit. Fluids3, 143–149.

Bhattacharjee, I., Chatterjee, S.K., Chandra, G., 2010. Isolation and identification of an-tibacterial components in seed extracts of Argemone mexicana L. (Papaveraceae).Asian Pac. J. Trop. Med. 3, 547–551.

Chassagnez-mendez, A.L., Machado, N.T., Araujo, M.E., Maia, J.G., Meireles, M.A.A.,2000. Supercritical CO2 extraction of Curcumins and essential oil from the Rhizomesof turmeric (Curcuma longa L.). Ind. Eng. Chem. Res. 39, 4729–4733.

Cocero, M., Garcia, J., 2001. Mathematical model of supercritical extraction applied to oilseed extraction by CO2+saturated alcohol - I. Desorption model. J. Supercrit. Fluids20, 229–243.

Cybenko, G., 1989. Approximation by superpositions of a sigmoidal function. Math.Control Signals Syst. 2, 303–314.

Döker, O., Salgın, U., Mehmeto, Ü., Çalımlı, A., Sanal, U., 2004. Modeling of extraction ofβ-carotene from apricot bagasse using supercritical CO2 in packed bed extractor. J.Supercrit. Fluids 28, 11–19.

Eslamimanesh, A., Gharagheizi, F., Mohammadi, A.H., Richon, D., 2011. Artificial neuralnetwork modeling of solubility of supercritical carbon dioxide in 24 commonly usedionic liquids. Chem. Eng. Sci. 66, 3039–3044.

Esquível, M.M., Bernardo-Gil, M.G., King, M.B., 1999. Mathematical models for super-critical extraction of olive husk oil. J. Supercrit. Fluids 16, 43–58.

França de, L.F., Reber, G., Meireles, M.A., Machado, N.T., Brunner, G., 1999. Supercriticalextraction of carotenoids and lipids from buriti (Mauritia flexuosa), a fruit from theAmazon region. J. Supercrit. Fluids 14, 247–256.

Fullana, M., Trabelsi, F., Recasens, F., 2000. Use of neural net computing for statisticaland kinetic modelling and simulation of supercritical fluid extractors. Chem. Eng. Sci.55, 79–95.

Ghoreishi, S.M., Heidari, E., 2013. Extraction of Epigallocatechin-3-gallate from green teavia supercritical fluid technology: neural network modeling and response surfaceoptimization. J. Supercrit. Fluids 74, 128–136.

Ghoreishi, S.M., Shahrestani, Rahman G., Ghaziaskar, H.S., 2009. Experimental andmodeling investigation of supercritical extraction of mannitol from olive leaves.Chem. Eng. Technol. 32, 45–54.

Goodarznia, I., Eikani, M.H., 1998. Supercritical carbon dioxide extraction of essentialoils: modeling and simulation. Chem. Eng. Sci. 53, 1387–1395.

Goto, M., Roy, B.C., Hirose, T., 1996. Shrinking-core leaching model for supercritical-fluid extraction. J. Supercrit. Fluids 128–133.

Kamali, M.J., Mousavi, M., 2008. Analytic, neural network, and hybrid modeling of su-percritical extraction of α-pinene. J. Supercrit. Fluids 47, 168–173.

Kazan, A., Koyu, H., Cemre, I., Yesil-celiktas, O., 2014. Supercritical fluid extraction ofPrunus persica leaves and utilization possibilities as a source of phenolic compounds.J. Supercrit. Fluids 92, 55–59.

Khajeh, M., Moghaddam, M.G., Shakeri, M., 2012. Application of artificial neural networkin predicting the extraction yield of essential oils of Diplotaenia cachrydifolia bysupercritical fluid extraction. J. Supercrit. Fluids 69, 91–96.

Kumar, A., Sharma, S., 2011. Potential non-edible oil resources as biodiesel feedstock: anIndian perspective. Renew. Sustain. Energy Rev. 15, 1791–1800.

Lamb, W.J., Hoffman, G.A., Jonas, J., 1981. Self-diffusion in compressed supercriticalwater. J. Chem. Phys. 74, 6875–6880.

Lashkarbolooki, M., Vaferi, B., Rahimpour, M.R., 2011. Comparison the capability ofartificial neural network (ANN) and EOS for prediction of solid solubilities in su-percritical carbon dioxide. Fluid Phase Equilib. 308, 35–43.

Lashkarbolooki, M., Shafipour, Z.S., Hezave, A.Z., 2013. Trainable cascade-forward back-propagation network modeling of spearmint oil extraction in a packed bed using SC-CO2. J. Supercrit. Fluids 73, 108–115.

Lee, A.K.K., Bulley, N.R., Fattori, M., Meisen, A., 1986. Modelling of supercritical carbondioxide extraction of canola oilseed in fixed beds. J. Am. Oil Chem. Soc. 63, 921–925.

Mishra, V., Saxena, D.K., Das, M., 2009. Effect of argemone oil and argemone alkaloid,sanguinarine on Sertoli-germ cell coculture. Toxicol. Lett. 186, 104–110.

Mukhopadhyay, M., 2000. Natural Extracts Using Supercritical Carbon Dioxide. CRCPress, Washington D.C.

Nei, H.Z.N., Fatemi, S., Mehrnia, M.R., Salimi, A., 2008. Mathematical modeling andstudy of mass transfer parameters in supercritical fluid extraction of fatty acids fromTrout powder. Biochem. Eng. J. 40, 72–78.

Pilkington, J.L., Preston, C., Gomes, R.L., 2014. Comparison of response surface metho-dology (RSM) and artificial neural networks (ANN) towards efficient extraction ofartemisinin from Artemisia annua. Ind. Crops Prod. 58, 15–24.

Poularikas, 2002. Handbook of Neural Network Signal Processing.Pramanik, P., Das, P., Kim, P.J., 2012. Preparation of biofuel from argemone seed oil by

an alternative cost-effective technique. Fuel 91, 81–86.Reverchon, E., Donsi, G., Osseo, L.S., 1993. Modeling of supercritical fluid extraction from

herbaceous matrices. Ind. Eng. Chem. Res. 32, 2721–2726.Roy, B.C., Goto, M., Hirose, T., 1996. Extraction of ginger oil with supercritical carbon

dioxide: experiments and modeling. Ind. Eng. Chem. Res. 35, 607–612.Samani, N., Gohari-Moghadam, M., Safavi, A.A., 2007. A simple neural network model for

the determination of aquifer parameters. J. Hydrol. 340, 1–11.Savage, P.E., Gopalan, S., Mizan, T.I., Martino, C.J., Brock, E.E., 1995. Reactions at su-

percritical conditions: application and fundamentals. AIChE J. 41, 1723–1777.Schmitt, W.J., Reid, R.C., 1986. The use of entrainers in modifying the solubility of

phenanthrene and benzoic acid in supercritical carbon dioxide and ethane. FluidPhase Equilib. 32, 77–99.

Shokri, A., Hatami, T., Khamforoush, M., 2011. Near critical carbon dioxide extraction ofAnise (Pimpinella Anisum L.) seed: mathematical and artificial neural networkmodeling. J. Supercrit. Fluids 58, 49–57.

Singh, D., Singh, S.P., 2010. Low cost production of ester from non edible oil of Argemonemexicana. Biomass Bioenergy 34, 545–549.

Song, X., Mitnitski, A., Macknight, C., Rockwood, K., 2004. Assessment of individual riskof death using self-report data: an artificial neural network compared with a frailtyindex. J. Am. Geriatr. Soc. 52, 1180–1184.

Sovova, H., Kucera, J., Jez, J., 1994. Rate of the vegetable oil extraction with supercriticalCO2-II. Extraction of grape oil. Chem. Eng. Sci. 49, 415–420.

Suryawanshi, B., Mohanty, B., 2018. Modeling and optimization: supercritical CO2 ex-traction of Pongamia pinnata (L.) seed oil. J. Environ. Chem. Eng. 6, 2660–2673.

Wu, B.C., Klein, M.T., Sandler, S.I., 1991. Solvent effects on reactions in supercriticalfluids. Ind. Eng. Chem. Res. 30, 822–828.

B. Suryawanshi, B. Mohanty Industrial Crops & Products 123 (2018) 64–74

74