Artificial Neural Network Modelling Ginger Rhizome

Available online at www.globalilluminators.org

GlobalIlluminators

FULL PAPER PROCEEDING Multidisciplinary Studies

Full Paper Proceeding MIS G-2015, Vol. 1, 136-154 IS BN: 978-969-9948-31-2

MISG 2015

Artificial Neural Network Modelling Ginger Rhizome Extracted Using Rapid Expansion Supercritical Solution (RESS) Method N.A., Zainuddin.,1* ,Norhuda, I.2 , Adeib, I.S.3, A.N.Mustapa4 , S.H., Sarijo5 1,2,3 4

, Faculty of Chemical Engineering, University Teknologi MARA (UiTM),40450, Shah Alam, Selangor Darul Ehsan,Malaysia, 5 Faculty of Applied Science, University Teknologi MARA (UiTM),40450, Shah Alam, Selangor Darul Ehsan,Malaysia

Abstract In this study, a feed forward mu ltilayer back propagation with Levenberg-Marquardt training algorithm art ificial neural network (ANN) was developed to predict the particle size fro m ext raction of ginger rhizo me using supercritical carbon dio xide in Rap id Expansion Supercritical Solution (RESS). Sizes of solid o il particle formation analysis are carried using Scanning Electron Microscopy (SEM) and Image processing and analysis software, ImageJ. The ANN model account for the effect of ext raction temperature (40, 45, 50, 55, 60, 65 and 70°C) and pressure (3000, 4000, 5000, 6000 and 7000psi) on the size of particle. A two -layer ANN with two inputs variables (ext raction temperature and pressure) and one output (particle size) with 35 experimental data was used for the modelling purposed. Different network were trained and tested by changing the number of neuron in hidden layer. Using validation data set the network having the highest (nearest to value of one) regression coefficient (R) is 0.99721 and the lowest (nearest to value of zero) mean square error (MSE) is 0.00031 was selected as an optimu m ANN model. The suitable ANN model is found to be one hidden layer with 7 hidden neurons. © 2015 The Authors. Published by Global Illuminators. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/) Peer-review under responsibility of the Scientific & Review committee of MISG-2015.

Keywords: Artificial Neural Net work (ANN); Part icle Size; Ginger; RESS; Supercritical CO2

Introduction Ginger (Zingiber officinale, Rosc.) is a herb plant belonging to the Zingiberceae family. Ginger widely used in pharmaceuticals, cosmetics and as flavouring agents in foods and beverages industry(Puengphian & Sirichote, 2008). Ginger is known to contain high antioxidant that can provide additional defence against oxidation (Atashak, Peeri, Azarbayjani, & Stannard, 2014) and it is a natural dietary component, which has antioxidant and anticarcinogenic properties (Manju & Nalini, 2005).

*All correspondence related to this article should be directed to N.A., Zainuddin, Faculty of Chemical Engineering, University Teknologi MARA (UiTM),40450, Shah Alam, Selangor Darul Ehsan,Malaysia Email: [email protected] © 2015 The Authors. Published by Global Illuminators. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/) Peer-review under responsibility of the Scientific & Review committee of MISG-2015.

N.A., Zainuddin /MISG-2015/Full Paper Proceeding Vol. 1, 136-154

Over the past few years, demand for supercritical fluid extraction process for particle formation in industry was rapidly increased (Hakuta, Hayashi, & Arai, 2003). To date, there is a growing interest in studies on natural product to direct solid oil particle formation in the pharmaceuticals industries (James, 2001). Development of solid oil particle will increase efficiency of drug delivery, reduced drug dosage and improve bioavailability of the drug (Zhao et al., 2011). Rapid expansion of supercritical solutions (RESS) is a new and promising technology to produce small particles and is suitable to use for thermal degradable pharmaceutical compounds (Turk, 1999). Recently, RESS method have been used by researchers to produce fine particles for pharmaceutical drug substances (Sheth, 2008). RESS has successfully employed to produce small size contaminant free particles of heat sensitive materials such as aspirin and carbamazepine ((Huang, Sun, Chiew, & Kawi, 2005)&(Bolten & Türk, 2012)). Supercritical fluid can be referred as a substance conditions above it critical temperature and critical pressure, where properties of the substance particularly the mix-of gas like and liquid- like properties and it does not condense to form liquid or gas (Sapkale, Patil, Surwase, & Bhatbhage, 2010). Common solvent use as supercritical fluid is carbon dioxide (CO 2 ) due to it environmentally friendly properties, nontoxic, non- flammable, easy to recycle, inexpensive, easy to get, low critical temperature (31.05°C) and moderate critical pressure (7.38Mpa), and easy to remove from extraction product. To date, there is no reported data on particle size of solid oil particle from ginger powder from RESS technology. An understanding of the complex phenomena happening in the RESS process has not yet been reached till now and a lot of variables can affect the RESS process such as the temperature and pressure in the pre-expansion and post-expansion chamber, the nozzle diameter and geometry, and the solute concentration(Hirunsit, Huang, Srinophakun, Charoenchaitrakool, & Kawi, 2005). Determination of particle size is one of the critical parameter. Experimental studies on the solid oil particle formation from ginger rhizome using RESS technology over the whole range of temperature and pressure need a very expensive experimental investigation and time consuming(Shabanzadeh, Shameli, Ismail, & Mohagheghtabar, 2013). Predictive method for estimated was therefore recommended to predict particle size within the operating conditions specified. In this work, focus on application of Artificial Neural Network (ANN) was done for ginger solid oil particle size prediction. Artificial Neural Network (Ann) ANN model is the best technique for solving complex engineering problem. The basic advantage of ANN is that it does not need any mathematical model since an ANN learns from examples and recognizes patterns in a series of input and output data without any prior assumptions about their nature and interrelations (Shabanzadeh, Senu, Shameli, & Ismail, 2013). In last two decades ANN models have been used in different fields of engineering with a basic objective of achieving human like performance since it is a powerful tools to identify complex relationship between input and output data (Gnanasangeetha & Sarala Thambavani, 2014). The first significant paper on neural network model that inspired many other researchers in the neural network development have been proposed in 1943 by McCulloch and Pits , where the developed model had two inputs and a single output (Hu & Hwang, 2002). McCulloch and Pitts noted that a neuron would not activate if only one of the inputs was active. In 1949 Hebb formed the basis of „Hebbian learning‟ which is now considered as

2nd International Conference on “Role Of M ultidisciplinary Innovation For Sustainability And Growth Policy”(MISG- 2015)

137


an important part of ANN theory and in 1958, John von Neumann proposed the modern brain performance of neural network using digital computer for further investigation of ANN performance. In 1957, Frank Rosenblatt created neuron model in hardware (Hu & Hwang, 2002). In the same way to the neuron in brain, ANN is composed of two or more artificial neurons. ANN is a biologically inspired human brain working situations and potentially fulfils the vision of scientists to develop machines that can think like human beings (Sutariya, Groshev, Sadana, & Bhatia, 2013). ANN is a computer programs that mimic the way of human brain processes information (Agatonovic-Kustrin & Beresford, 2000). An ANN consists of processing elements, connections between them, training and learning algorithm and recall algorithms (Kasabov, 1998). The information comes into the body of an artificial neuron called inputs that are weighted by multiplying each input with weights (Ibrić, Djuriš, Parojčić, & Djurić, 2012). An input function calculates the summation of net input signal to neuron coming from all its input and pass through transfer function to produce an output. Activation signal function calculate the activation level of neuron as function of its summation input signal and of its previous state (Kasabov, 1998). An output signal is consider equal to the output of a neuron to its activation level (Krose & Smagt, 1996).

Materials And Methods Materials The Zingiber officinale, Rosc. rhizome used in this study was purchased from local market. Fresh ginger rhizomes were washed through with tap water to remove dirt. Skin of the ginger rhizome was peeled and then cut into cross section slices with 2 to 3 mm thick. An amount of 500 g ginger slice was weighed and undergo oven drying in oven model Memmert UFE 500 to remove moisture content to achieve 10 to 12% (Puengphian & Sirichote, 2008). Dried ginger sample then was ground using mechanical grinder Retsch model SM100 for 3 minutes through a plate of 1 mm pore size to obtain ginger powder. The sample was then stored in the refrigerator at 4°C until the further use.

Rapid Expansion Supercritical Carbon Dioxide (RESS) The studies of ginger particle extract from ground ginger sample at operating temperature (40, 45, 50, 55, 60, 65 and 70ºC), pressure (3000, 4000, 5000, 6000 and 7000psi) and 40 minutes of extraction time (Rodríguez- meizoso et al., 2012). Value for operating temperature, pressure for extraction of ginger rhizome selected based on the average of the operating condition on several others literatures that utilize RESS-CO2 from previous research done by many others researchers in particle formation using RESS method (Atila, Yıldız, & Çalımlı, 2010; Baseri & Lotfollahi, 2013; Bolten & Türk, 2012; Hezave, Aftab, & Esmaeilzadeh, 2010; Keshavarz et al., 2012; Pourasghar, Fatemi, Vatanara, & Rouholamini Najafabadi, 2012; Yildiz, Tuna, Döker, & Çalimli, 2007). RESS equipment in this study was modified from Supercritical Fluid Technologies Model SFT-100 equipment. Each study will be conducted using 8 g sample of ground ginger based on observation on the maximum mass of ground ginger that can be loading and fit in 25 mL extraction vessel. If mass of the sample


138


is higher than 8 g, it will cause some lose fitting at the top of the extraction vessel seal that will cause fault or leaking of CO 2 when the extraction process in progress. Extraction pressure and temperature was set at the desire value, and ginger sample was inserted in a cotton bag before put in the extraction vessel. The top of the extraction vessel was tightly sealed before closing the top cover. Dynamic valve was opened and restrictor valve was closed during the extraction time. Collection bottle was inserted in the expansion chamber. When extraction temperature achieved to the desired value, CO 2 pumped to feed high pressure liquid CO 2 (99 % purity provided by MOX Linde Gases Sdn Bhd) continuously in the extraction vessel at fixed solvent flowrate of 24 mL/min. Basically, liquid CO 2 will be converted to supercritical condition when pump into the extraction vessel (heated zone). To achieve the desired pressure set point, CO 2 pump will continuously actuated. After 40 minutes of extraction, the restrictor valve was quickly open to depressurize the supercritical solution for separation of solute from the solvent. Depressurize CO 2 at the ambient pressure will be converted into gas form and purged into the ambient. Extraction product will be collected in collection vials. To determine the extraction yield, weight of the collection vials will be measured before and after RESS process. The extraction product was weighed using the analytical balance model Mettler Toledo AB204-S. The procedure of extraction process was repeated in triplicate under desired operating conditions and the data were given as the average values.

Characterization Method and Instruments Scanning Electron Microscopy (SEM) Analysis on particle size was carried out using Scanning Electron Microscopy (SEM), model TM3000 Tabletop Microscope, brand Hitachi. For the analysis, samples were applied on double sticky carbon tape located on top of circular aluminium stub (Moreschi, Leal, Braga, & Meireles, 2006). SEM image from different region of the stub were captured.

ImageJ Analysis (ImageJ) Image processing and analysis software, ImageJ was used to perform particle size analysis (Bolten & Türk, 2012). 100 particles selected randomly from SEM image were used to calculate particle size of the ground ginger and particle obtained from RESS processes (Hezave & Esmaeilzadeh, 2011).

Data Set A set of input data with 35 samples was used as input matrix (35 x 2) with corresponding values of particle size obtained from RESS experiment used as target matrix (35 x 1) in the developed ANN model. There is no unique method for determining the size of the data to the appropriate network training (Witek-krowiak, Chojnacka, Podstawczyk, Dawiec, & Pokomeda, 2014). In this study, input data was divided randomly into three data sets which are 70 % for training (25 samples), 15 % for validation (5 sa mples) and 15% for training (5 samples)(Beale, Hagan, & Demuth, 2011). The training data was applied to 2nd International Conference on “Role Of M ultidisciplinary Innovation For Sustainability And Growth Policy”(MISG- 2015)

139


compute the network parameters while the validation data was applied to ensure robustness of the network parameters and the testing stage was used to control error; when it increased, the training stops. Inputs and output data are normalized in the range of 0–1 for the reduction of network error and higher homogeneous results (Ghoreishi & Heidari, 2012), (Kasabov, 1998) by using the following equation: Xnorm = (Xreal - Xmin )/( Xmax – Xmin ) ………………………………...Equation (3.1) Where:Xnorm = is a normalized value, Xreal = is a real value, Xmin = the minimum values for the variable X, Xmax = the maximum values for the variable X Ozgan, 2009, in his study on modelling of soil particle diameters normalized input and output data between 0 and 1 using the same equation as shown above (Ozgan, 2009). Table 0.1 shows the experimental data and normalization experimental data which were used in this ANN model.

Artificial Neural Network for Prediction of Particle Size Modelling for particle size prediction of solid oil particle from ginger rhizome in RESS process was established using MATLAB software (version 7.9.0.(R2009b)) neural network toolbox. A feed forward neural network and the back-propagation learning algorithm ANN was used in this study to developed an ANN model since it is the most commonly applied ANN layout (Sutariya et al., 2013; Witek-krowiak et al., 2014) . Feed- forward backpropagation neural network developed by Rumehart and McClelland in 1986 is the most widely applied ANN layout (Agatonovic-Kustrin & Beresford, 2000; Basheer & Hajmeer, 2000; Sutariya et al., 2013; YU, Onder Effe, & Kaynak, 2002) . Extraction temperature and extraction pressure were designated as input data, while particle size data from imageJ analysis was established as output data for further used subjected into ANN modelling for training, testing and validation. The predicted output data from ANN was optimized by referring to the Mean Square Error (MSE) value and regression value (R) values gained from developed model when comparison was performed between predicted and experimental value. Trial and

Table 0.1 :Original and normalization data of Particle Size

No.

Original Data Extraction Extraction Particle Pressure Temperature Size (Psi) (°C) (μm)

Training set 1 3000 2 4000 3 5000 4 6000

40 40 40 40

4.15 5.01 2.22 3.03

Normalization Data Extraction Extraction Pressure Temperature (Psi) (°C)

Particle Size (μm)

0.00 0.25 0.5 0.75

0.37 0.53 0.00 0.15

0.00 0.00 0.00 0.00


140


5 7000 40 3.46 1.00 0.00 0.24 6 3000 0.00 45 4.21 0.17 0.38 7 4000 0.25 45 5.95 0.17 0.71 8 5000 45 3.71 0.5 0.17 0.28 9 6000 0.75 45 3.64 0.17 0.27 10 7000 1.00 45 3.74 0.17 0.29 11 3000 50 4.50 0.00 0.33 0.44 12 4000 0.25 50 6.56 0.33 0.83 13 5000 0.50 50 5.33 0.33 0.59 14 6000 50 3.77 0.75 0.33 0.30 15 7000 1.00 50 4.45 0.33 0.43 16 3000 0.00 55 5.18 0.50 0.56 17 4000 55 7.46 0.25 0.50 1.00 18 5000 0.50 55 6.07 0.50 0.73 19 6000 0.75 55 3.59 0.50 0.26 20 7000 55 3.74 1.00 0.50 0.29 21 3000 0.00 60 2.99 0.67 0.15 22 4000 0.25 60 6.59 0.67 0.83 23 5000 60 4.89 0.50 0.67 0.51 24 6000 0.75 60 3.28 0.67 0.20 25 7000 1.00 60 3.44 0.67 0.23 Validation Set 26 3000 0.00 65 2.62 0.83 0.08 27 4000 0.25 65 6.1 0.83 0.74 28 5000 65 3.9 0.50 0.83 0.32 29 6000 0.75 65 2.91 0.83 0.13 30 7000 1.00 65 2.74 0.83 0.10 Test Set 31 3000 0.00 70 2.62 1.00 0.08 32 4000 0.25 70 5.2 1.00 0.57 33 5000 70 2.86 0.50 1.00 0.12 34 6000 0.75 70 2.63 1.00 0.08 35 7000 1.00 70 2.75 1.00 0.10 error method was applied in the network optimization process until the model shows the lowest value for MSE or nearest to zero are better and R value nearest to 1 for testing, validation and training data set, in order to determine the suitable number of neuron and hidden layer required for the ANN architecture in this study. FFigure 0.1 shows ANN schematic diagram used in this modelling study for particle size prediction. X1 and X2 represents two input parameters applied in this research as temperature and pressure that previously applied in the laboratory work. “D” represent target output which is particle size (μm) collected from experimental work (ImageJ analysis). “E” is a schematic diagram represent ANN model with one hidden layer with several of hidden layer and one output layer with one neuron. “O” represents the predicted particle size (μm) computed from network. “Wh ” represent weight connection in headen layer,and “Wo ” is


141


weight connection in output layer, “b” is bias. f(hn ) is transfer function for each neuron in hidden layer, in this case we used tansig and f(y) is transfer function in output layer, where purelinear was choose.

Input Layer

Hidden Layer

Output Layer

b f(h1 ) Wh1,1

Temperature

X1

f(h2 ) f(h3 ) f(h4 )

Pressure

X2

WO

f(h5 ) f(h6 )

O1 O2 O3 O4 O5 O6 O7

D

b

f(y)

O

E

Error propagate

f(h7 ) F Figure 0.1: Schematic diagram for multilayer feed forward back propagation ANN for solid oil ginger rhizome particle size prediction

Feed-forward back-propagation neural network developed by Rumehart and McClelland in 1986 is the most widely applied ANN layout (Agatonovic-Kustrin & Beresford, 2000; Basheer & Hajmeer, 2000; Sutariya et al., 2013; YU et al., 2002) . The back-propagation algorithm is the most popular supervised learning method because weights are adapted to minimise the error between the desired outputs and those calculated by the network. Jadid & Fairbairn, 1996 have reported about 80% of neural network application applied back-propagation learning algorithm since it considered to be most reliable and most applicable (Jadid & Fairbairn, 1996). Hence, feed-forward back-propagation neural network was developed in this study. Multilayer feed-forward back-propagation neural network means that the artificial neurons are organized in layers, and send their signals forward to produce solution and then the errors are propagated backwards through hidden layer toward input layer to modify the weight (Basheer & Hajmeer, 2000)(Basheer & Hajmeer, 2000). It consists of input layer, one or more hidden layer and one output layer (Bakhbakhi, 2011). Input layer receives experimental information and experimental parameter, and then send weight values to hidden layer and then the output layer produce the calculated value of the independent variable. Hidden layer is between input layer and output layer. In the hidden layer, input such as pressure and temperature was multiplied by weights, then the output is calculated at the output node using transfer activation function (Bakhbakhi, 2011). Then the output from hidden layer was sent to final layer which is the output layer that comprise a prediction of the target output of the model (Si-Moussa, Hanini, Derriche, Bouhedda, & Bouzidi, 2008) . 2nd International Conference on “Role Of M ultidisciplinary Innovation For Sustainability And Growth Policy”(MISG- 2015)

142


The size of a hidden layer is one of the most important considerations when solving actual problems using multilayer feed forward networks (Zurada, 1992). In this study, number of hidden layer was selected as one. The selection on the number of hidden layer was done based on study on prediction of particle size done by (Gnanasangeetha & Sarala Thambavani, 2014; Shabanzadeh, Senu, et al., 2013; Takayama, Fujikawa, Obata, & Morishita, 2003). Lashkarbolooki et al., 2011 also used one hidden layer in their study for prediction of solid solubility in supercritical CO2 . Sargolzaei, Asl, & Moghaddam, 2013 used one hidden layer in their research on prediction of oil yield using supercritical fluid extraction method . Witek-krowiak et al., 2014 ,Sun et al., 2003 and Basheer & Hajmeer, 2000 stating in their paper, one hidden layer are commonly applied and adequate to provide an accurate prediction of the performance of many processes. Training algorithm was optimized by observed the ideal number of hidden neuron in the hidden layer for the developed ANN model. There is no unique method for deter mining for determining the optimal number of hidden neuron (Witek-krowiak et al., 2014). Growing neural networks technique was used in this work, where training starts with a small number of hidden neuron and subject to the error calculated, the number of the hidden neuron may increase during the training procedure (Kasabov, 1998). So, the selection of the number of hidden neuron was achieved by trial and error method in agreement of obtained error function results respect to the training data set, for MSE is nearest to 0 (Faria, Alberto, Anjos, Nunes, & Maria, 2015; Fatehi, Raeissi, & Mowla, 2014; Izadifar & Abdolahi, 2006; Kuvendziev, Lisichkov, Zekovic, & Marinkovski, 2014; Shabanzadeh, Senu, et al., 2013). This is the most common approach for selection optimal number of the hidden neuron (Basheer & Hajmeer, 2000; Sun et al., 2003). In this study, the number of hidden neuron was set varied from 2 to 15. After finalizing number of hidden layer and number of hidden neuron, ANN is ready for training. Activation functions are used in ANN to produce continuous values rather than discrete ones (Youssef, Aly, & Zeidan, 2013). In this work, the activation functions assigned in hidden layer neurons are sigmoid functions (TANSIG) and the linear activation function (PURELIN) is used for the output layer neurons, since it is the most commonly used according to Kuvendziev et al., 2014. Gnanasangeetha et al., 2014 used the same activation function in their study for modelling of Zinc Oxide Nanoparticles. Application of ANN in prediction of particle size has been done and reported successfully by many numbers of researchers such as Shabanzadeh, Shameli, et al., 2013 had synthesized silver nanoparticles (Ag-NPs) using green method and develop back propagation ANN model for the prediction of particle size of Ag-NPs based on experimental data. Shabanzadeh et al., 2013 had done biosynthesized of silver nanoparticles (Ag-NPs) from silver nitrate aqueous solution using water extract from Curcuma longa (C. longa) tuber powder using Levenberg–Marquardt (LM) back-propagation neural network model.

Results And Discussion In this study, a multilayer feed- forward back propagation neural network was used. Extraction temperature and extraction pressure was set as input data and particle size from SEM and ImageJ analysis was set as output or target data. Before use as input and output data in ANN model, all the data was normalized. A set of input data with 35 samples was used as input matrix (35 x 2) with corresponding values of particle size obtained from RESS experiment used as target matrix (35 x 1) in the developed ANN model. Data was randomly divided into 3 main set which are training (25 samples), validation (5 samples) and test set (5 2nd International Conference on “Role Of M ultidisciplinary Innovation For Sustainability And Growth Policy”(MISG- 2015)

143


samples) for 70%, 15% and 15% respectively. Table 1 shows the experimental data and normalization experimental data which were used in this ANN model. The training data was applied to compute the network parameters while the validation data was applied to ensure robustness of the network parameters and the testing stage was used to control error; when it increased, the training stops. Figure 0.1 shows a graph of the MSE against the number of hidden neuron unit to choose the optimal topology decision for the network. As shown in Figure 0.1, the decision to use 7 hidden neuron units in hidden layer was taken based on the minimum value of MSE for training data set, 0.00031. If too many hidden neuron, it will lead to over fitting (Hu & Hwang, 2002). If too few hidden neuron it will cause the ANN model to be insufficiently flexible to represent the experimental signal (Almeida, 2001). As mention by (Hu & Hwang, 2002) in Handbook of Neural Network Signal Processing, it is difficult to determine the optimal amount of training. During the training and testing process, the training error decreases until over fitting phenomena occurs, when the error start to begin increase (Alshayea, 2011; Basheer & Hajmeer, 2000; Izadifar & Abdolahi, 2006; Li & Bridgwater, 2000). In this study, to prevent over fitting, the training step was stopped when error value on the validation data set was increased.

Table 0.1 shows characteristic of the developed model in this study.

Figure 0.1:Variation number of hidden neouron unit

Table 0.1: Characteristic of the developed ANN model Characteristic

Co mmentary

Algorith m

Feed forward back propagation

Minimized error function

MSE


144


Train ing algorith m

Levenberg-Marquardt

Learn ing type

Supervised

Hidden layer

Hyperbolic tangent transfer function

Output layer

Pure linear t ransfer function

Nu mber of neuron in input layer

2

Nu mber of h idden layer

1

Nu mber of neuron in hidden layer

7

Nu mber of neuron in output layer

1

Figure 0.2 shows the scatter regression (R) plots for the actual value from experimental data and predicted value for particle size from developed ANN model for training data set. The solid linear line represents the best fit between the net work output (predicted particle size from ANN model) and target output (particle size from experimental work). The best correlation coefficient of 0.99721 between the ANN models predicted value and the experimental value data obtained showed a good correlation for the prediction with MSE value of 0.00031. Shabanzadeh, et al., 2013 in their study use ANN model for prediction diameter of silver nanoparticles biosynthesized in curcuma longa extract, they get MSE value as 0.0155 and R value as 0.9996 (Shabanzadeh, Senu, et al., 2013). And in a study conducted by Shabanzadeh, et al., 2013 on prediction of silver nanoparticles using green method, they found the R value for training set as 0.98398 (Shabanzadeh, Shameli, et al., 2013) . It is prove that the value of MSE and R in this can be reflected as an exact value for the prediction of particle size. According to Sargolzaei et al., 2013, R value greater than 0.9 indicates a very satisfactory model performance, while R value in the range 0.8-0.9 signifies a good performance and value less than 0.8 indicates unsatisfactory model performance (Sargolzaei et al., 2013). The closer the R value is to 1, the better the model fits to the actual data (Shabanzadeh, Senu, et al., 2013). In this study, minimum value of MSE and the maximum of R were considered as the best neural network model. So the structure of the utilized NN has been configured with 2 inputs neurons, 7 hidden neurons in 1 hidden layer and one output neuron. Figure 0.3 shows comparison of the particle size between experimental value and predicted values from developed ANN model for training set. The developed ANN has the best correlation with the experimental results as displayed in Figure 0.4 where R2 value is 0.9499.


145


Figure 0.2 :Scatter regression plot of actual and ANN model predicted particle size value for training data set

Figure 0.3: Comparison of the particle size from experimental and predicted values from developed ANN model for training set

Figure 0.4: Correlation coefficient between normalized and predicted particle size from training set

The performance of the optimal ANN certify by using validation data test, where another new data set consists of 5 data points which have not been used during the training process. Figure 0.5 shows scatter plot for actual value and predicted value for particle size from validation data set. After analysing the data using ANN, R value for validation data set is 0.97449. This value demonstrates that the model has satisfactory predictive ability and good quality. (Shabanzadeh, Shameli, et al., 2013) and (Shabanzadeh, Senu, et al., 2013) in their study for prediction of silver nanoparticles found R value for validation data set as


146


0.97788 and 0.9920. As shown in Figure 0.6, the best validation performance is 0.013672 at epoch 11.

Figure 0.5: Scatter regression plot of actual Figure 0.6: Best validation perfomance and ANN model predicted particle size value for validation data set

Then, for testing fitness of developed ANN model, the scatter plot of experimental data versus the predicted particle size data is present in Figure 0.7. The testing data set shows, the result obtained from the developed ANN model are acceptable since the R value is 0.94767. Figure 0.8 shows comparison of the particle size from experimental and predicted values from developed ANN model for testing data set, and Figure 0.9 represent the correlation coefficient with the experimental results where R2 value is 0.9967.

Figure 0.7: Scatter regression plot of actual and ANN model predicted particle size value for testing data set


147


Figure 0.8: Comparison of the particle size from experimental and predicted values from developed ANN model for testing data set

Figure 0.9: Correlation coefficient between normalized and predicted particle size from testing data set

Figure 0.10 shows scatter regression plot of actual and ANN model predicted particle size value for all 35 data points use in this study, where the R value found in this model is 0.97477, that is can be considered as a very satisfactory model performance. Table 0.2 represent summary for MSE and R value for training, validation and testing data set that found in this study.

Figure 0.10: Scatter regression plot of actual and ANN model predicted particle size value for all data set


148


Table 0.2: The Mean Square Error (MSE) and Regression (R) values for the training, validation and testing MSE Value

R-Value

Train ing

0.00031

0.99721

Validation

0.01367

0.97449

Test

0.00892

0.94767

Table 0.3 shows the actual particle size in μm for experimental data from the entire experimental work, the predicted particle size from developed ANN model, the error between particle sizes from experimental data and developed ANN model, estimation of particle size using RS tool in MATLAB using pure quadratic model and estimation of particle size using equation from EXCEL. For the prediction of pure quadratic model,………………………….. Equation (0.1) was used. This equation was developed and the constant values were found from the prediction plot of RS tool in MATLAB: y=B0 +B1 X1 +B2 X2 +B3 X1 2 +B4 X22 ………………………….. Equation (0.1) Where: value of B0 , B1 , B2 , B3 , B4 is 0.309958, 1.053659, 0.427265, -1.17452 and -0.68898, respectively;X1 = Temperature (°C ) and X2 = Pressure (Psi) For the efficiency study of particle size using EXCEL, Equation (4.2) was applied. This equation was found from the plot of actual data versus predicted data. The R 2 value obtained from this relationship is 0.9625 as shown in Figure 0.11. Therefore it can be proven that, the actual and predicted data from developed ANN model, is in good agreement. y = 0.031x2 + 0.6916x + 0.7005 ………………………………...Equation (0.2) where x= particle size (μm)

Figure 0.11: Relationship between Actual particle size from experimental data and ANN developed predicted data .


149


Table 0.3: The estimated results obtained by proposed ANN model and data fitting

Particle Size (μm) No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35

Pressure (Psi) 3000 4000 5000 6000 7000 3000 4000 5000 6000 7000 3000 4000 5000 6000 7000 3000 4000 5000 6000 7000 3000 4000 5000 6000 7000 3000 4000 5000 6000 7000 3000 4000 5000 6000 7000

Temperature (°C) 40 40 40 40 40 45 45 45 45 45 50 50 50 50 50 55 55 55 55 55 60 60 60 60 60 65 65 65 65 65 70 70 70 70 70

Experimental Data 4.15 5.01 2.22 3.03 3.46 4.21 5.95 3.71 3.64 3.74 4.50 6.56 5.33 3.77 4.45 5.18 7.46 6.07 3.59 3.74 2.99 6.59 4.89 3.28 3.44 2.62 6.1 3.9 2.91 2.74 2.62 5.2 2.86 2.63 2.75

Predicted ANN model 4.14 4.98 2.22 3.01 3.39 4.25 6.36 3.52 3.63 3.99 4.50 7.30 5.28 3.78 4.18 4.30 7.36 5.68 3.45 3.98 3.55 6.63 4.91 3.38 3.44 2.94 5.82 3.87 2.83 2.79 2.65 5.22 2.85 2.79 2.68

Error 0.01 0.03 0.00 0.02 0.07 -0.04 -0.41 0.19 0.01 -0.25 0.00 -0.74 0.05 -0.01 0.27 0.88 0.10 0.39 0.14 -0.24 -0.56 -0.04 -0.02 -0.10 0.00 -0.32 0.28 0.03 0.08 -0.05 -0.03 -0.02 0.01 -0.16 0.07

Estimation of Particle Size (μm) Pure Quadratic in Equation MATLAB in Excel 4.10 3.84 4.18 4.94 2.39 4.06 3.08 3.49 2.47 3.46 4.16 4.60 5.91 4.94 4.82 3.69 3.63 4.25 3.72 3.23 5.00 4.44 6.57 5.33 5.27 5.21 4.64 3.75 4.39 3.62 5.11 5.07 5.40 7.59 6.04 5.28 3.58 4.71 3.69 3.72 3.05 4.78 6.60 5.11 5.00 4.82 3.30 4.43 3.45 3.41 4.19 2.73 6.07 4.52 3.87 4.40 3.84 2.98 2.83 2.82 2.73 3.21 3.54 5.14 2.93 3.43 2.73 2.86 2.60 2.84

Conclusion 2nd International Conference on “Role Of M ultidisciplinary Innovation For Sustainability And Growth Policy”(MISG- 2015)

150


In this research, an ANN model for prediction of particle size formation from extraction of powder ginger rhizome powder using RESS method was presented. The model accounts for the effects of extraction temperature and pressure on the particle size. Based on the results obtained, it can be concluded that the feed forward back propagation Levenberg– Marquardt model with one hidden layer and 7 hidden neurons is a good training and shows a good performance agreement in, predicting experimental data. Hence, it can be conclude that by using ANN models is a useful tool for saving time and cost for predicting the particle size.

Acknowledgement The author was grateful for financial support provided by Ministry of Higher Education, Malaysia and Universiti Teknologi MARA (UiTM) (600-RMI/ST/DANA 5/3/CG (17/2012)).

References Agatonovic-Kustrin, S., & Beresford, R. (2000). Basic concepts of artificial neural network ( ANN ) modeling and its application in pharmaceutical research. Journal of Pharmaceutical and Biomedical Analysis, 22, 717–727. Almeida, J. S. (2001). Predictive non- linear modeling of complex data by artificial neural networks. Current Opinion in Biotechnology, 13, 72–76. Al-shayea, Q. K. (2011). Artificial Neural Networks in Medical Diagnosis. International Journal of Computer Science, 8(2), 150–154. Atashak, S., Peeri, M., Azarbayjani, M. A., & Stannard, S. R. (2014). Effects of ginger (Zingiber officinale Roscoe) supplementation and resistance training on some blood oxidative stress markers in obese men. Journal of Exercise Science & Fitness, 12(1), 26–30. Atila, C., Yıldız, N., & Çalımlı, A. (2010). Particle size design of digitoxin in supercritical fluids. The Journal of Supercritical Fluids, 51(3), 404–411. Bakhbakhi, Y. (2011). Expert Systems with Applications Phase equilibria prediction of solid solute in supercritical carbon dioxide with and without a cosolvent : The use of artificial neural network. Expert Systems With Applications, 38(9), 11355–11362. Baseri, H., & Lotfollahi, M. N. (2013). Formation of gemfibrozil with narrow particle size distribution via rapid expansion of supercritical solution process (RESS). Powder Technology, 235, 677–684. Basheer, I. A., & Hajmeer, M. (2000). Artificial neural networks : fundamentals , computing , design , and application. Journal of Microbiological Methods, 43, 3–31. Beale, M. H., Hagan, M. ., & Demuth, H. . (2011). Neural Network Toolbox User’s Guide. Apple Hill Drive: The MathWork,Inc. Bolten, D., & Türk, M. (2012). Micronisation of carbamazepine through rapid expansion of supercritical solution (RESS). The Journal of Supercritical Fluids, 62, 32–40. Faria, S., Alberto, C., Anjos, R., Nunes, R., & Maria, R. (2015). Evaluation of extra virgin olive oil stability by artificial neural network. FOOD CHEMISTRY, 179, 35–43. Fatehi, M., Raeissi, S., & Mowla, D. (2014). The Journal of Supercritical Fluids An artificial neural network to calculate pure ionic liquid densities without the need for any experimental data. The Journal of Supercritical Fluids, 95, 60–67.


151


Ghoreishi, S. M., & Heidari, E. (2012). The Journal of Supercritical Fluids Extraction of epigallocatechin gallate from green tea via modified supercritical CO 2 : Experimental , modeling and optimization. The Journal of Supercritical Fluids, 72, 36–45. Gnanasangeetha, D., & Sarala Thambavani, D. (2014). ANN Approach for Modeling of Zinc oxide Nanoparticle Using Sensitivity Analysis. International Journal of Advanced Research in Science, Engineering and Technology, 1(4), 216–221. Hakuta, Y., Hayashi, H., & Arai, K. (2003). Fine particle formation using s upercritical fluids. Current Opinion in Solid State and Materials Science, 7(4-5), 341–351. Hezave, A. Z., Aftab, S., & Esmaeilzadeh, F. (2010). Micronization of ketoprofen by the rapid expansion of supercritical solution process. Journal of Aerosol Science, 41(8), 821–833. Hezave, A. Z., & Esmaeilzadeh, F. (2011). The effects of RESS parameters on the diclofenac particle size. Advanced Powder Technology, 22(5), 587–595. Hirunsit, P., Huang, Z., Srinophakun, T., Charoenchaitrakool, M., & Kawi, S. (2005). Particle formation of ibuprofen – supercritical CO 2 system from rapid expansion of supercritical solutions ( RESS ): A mathematical model. Powder Technology, 154, 83–94. Hu, Y. H., & Hwang, J. N. (2002). Handbook of Neural Network Signal Processing. New York: CRC Press LLC. Huang, Z., Sun, G.-B., Chiew, Y. C., & Kawi, S. (2005). Formation of ultrafine aspirin particles through rapid expansion of supercritical solutions (RESS). Powder Technology, 160(2), 127–134. Ibrić, S., Djuriš, J., Parojčić, J., & Djurić, Z. (2012). Artificial neural networks in evaluation and optimization of modified release solid dosage forms. Pharmaceutics, 4, 531–550. Izadifar, M., & Abdolahi, F. (2006). Comparison between neural network and mathematical modeling of supercritical CO 2 extraction of black pepper essential oil. The Journal of Supercritical Fluids, 38, 37–43. Jadid, M. N., & Fairbairn, D. R. (1996). Contributed Paper Neural-network Applications in Predicting Moment-curvature Parameters from Experimental Data. Engng Applic.Artif.Intell, 9(3), 309–319. James, K. (2001). PROCESSING PHARMACEUTICALS WITH SUPERCRITICAL. Kasabov, N. K. (1998). Foundations of Neural Networks , Fuzzy Systems , and Knowledge Engineering (second.). The MIT Press. Keshavarz, A., Karimi-Sabet, J., Fattahi, A., Golzary, A., Rafiee-Tehrani, M., & Dorkoosh, F. a. (2012). Preparation and characterization of raloxifene nanoparticles using Rapid Expansion of Supercritical Solution (RESS). The Journal of Supercritical Fluids, 63, 169–179. Krose, B., & Smagt, P. Van Der. (1996). An Introduction to Neural Networks. Kuvendziev, S., Lisichkov, K., Zekovic, Z., & Marinkovski, M. (2014). Artificial neural network modelling of supercritical fluid CO 2 extraction of polyunsaturated fatty acids from common carp ( Cyprinus carpio L .) viscera. Journal of Supercritical Fluids, 92, 242–248. Lashkarbolooki, M., Vaferi, B., & Rahimpour, M. . (2011). Comparison the capability of artificial neural network (ANN) and EOS for prediction of solid solubilities in supercritical carbon dioxide. F;uid Phase Equilibra, 308, 34–43. Li, Y. Y., & Bridgwater, J. (2000). Prediction of extrusion pressure using an artificial neural network. Powder Technology, 108, 65–73. Manju, V., & Nalini, N. (2005). Chemopreventive efficacy of ginger, a naturally occurring anticarcinogen during the initiation, post-initiation stages of 1,2 dimethylhydrazine-


152


induced colon cancer. Clinica Chimica Acta; International Journal of Clinical Chemistry, 358(1-2), 60–7. Moreschi, S. R. M., Leal, J. C., Braga, M. E. M., & Meireles, M. A. A. (2006). GINGER AND TURMERIC STARCHES HYDROLYSIS USING SUBCRITICAL WATER + CO 2 : THE EFFECT OF THE SFE PRE-TREATMENT, 23(02), 235–242. Ozgan, E. (2009). Neural Network Based Modelling of Soil Particle Diameters Under Varying Quantity of Sodium Hexametaphosphate. Asian Journal of Chemistry, 21(4), 2632–2642. Pourasghar, M., Fatemi, S., Vatanara, A., & Rouholamini Najafabadi, A. (2012). Production of ultrafine drug particles through rapid expansion of supercritical solution; a statistical approach. Powder Technology, 225, 21–26. Puengphian, C., & Sirichote, A. (2008). [6]-gingerol content and bioactive properties of ginger (Zingiber officinale Roscoe) extracts from supercritical CO2 extraction. Asian Journal of Food and Agro-Industry, 1(01), 29–36. Rodríguez-meizoso, I., Castro-puyana, M., Börjesson, P., Mendiola, J. A., Turner, C., & Ibá, E. (2012). The Journal of Supercritical Fluids Life cycle assessment of green pilot-scale extraction processes to obtain potent antioxidants from rosemary leaves, 72, 205–212. Sapkale, G. N., Patil, S. M., Surwase, U. S., & Bhatbhage, P. K. (2010). A REVIEW SUPERCRITICAL FLUID EXTRACTION, 8(2), 729–743. Sargolzaei, J., Asl, M. H., & Moghaddam, A. H. (2013). Prediction of Super Critical Oil Extraction Yield Using Single and Combined Intelligent Systems. Global Perspective on Artificial Intelligence, 1(2), 15–22. Shabanzadeh, P., Senu, N., Shameli, K., & Ismail, F. (2013). APPLICATION OF ARTIFICIAL NEURAL NETWORK (ANN) FOR PREDICTION DIAMETER OF SILVER NANOPARTICLES BIOSYNTHESIZED IN CURCUMA LONGA EXTRACT. Journal of Nanomaterials and Biostructures, 8(3), 1133–1144. Shabanzadeh, P., Shameli, K., Ismail, F., & Mohagheghtabar, M. (2013). APPLICATION OF ARTIFICIAL NEURAL NETWORK ( ANN ) FOR THE PREDICTION OF SIZE OF SILVER NANOPARTICLES PREPARED BY GREEN. Digest Journal of Nanomaterials and Biostructures, 8(2), 541–549. Sheth, M. M. (2008). ULTRA-FINE PARTICLE FORMATION USING PRINCIPLE OF RAPID EXPANSION OF SUPERCRITICAL SOLUTIONS. Si-Moussa, C., Hanini, S., Derriche, R., Bouhedda, M., & Bouzidi, A. (2008). PREDICTION OF HIGH-PRESSURE VAPOR LIQUID EQUILIBRIUM OF SIX BINARY SYSTEMS , CARBON DIOXIDE WITH SIX ESTERS , USING AN ARTIFICIAL NEURAL NETWORK MODEL. Brazilian Journal of Chemical Engineering, 25(1), 183–199. Sun, Y., Peng, Y., Chen, Y., & Shukla, A. J. (2003). A pplication of artificial neural networks in the design of controlled release drug delivery systems. Advanced Drug Dlivery Reviews, 55, 1201–1215. Sutariya, V., Groshev, A., Sadana, P., & Bhatia, D. (2013). Artificial Neural Network in Drug Delivery and Pharmaceutical Research. The Open Bioinformatics Journal, 7, 49–62. Takayama, K., Fujikawa, M., Obata, Y., & Morishita, M. (2003). Neural network based optimization of drug formulations. Advanced Drug Dlivery Reviews, 55, 1217–1231. Turk, M. (1999). Formation of small organic particles by RESS : experimental and theoretical investigations. Journal of Supercritical Fluids, 15, 79–89. Witek-krowiak, A., Chojnacka, K., Podstawczyk, D., Dawiec, A., & Pokomeda, K. (2014). Bioresource Technology Application of response surface methodology and artificial


153


neural network methods in modelling and optimization of biosorption process. Bioresource Technology, 160, 150–160. Yildiz, N., Tuna, Ş., Döker, O., & Çalimli, A. (2007). Micronization of salicylic acid and taxol (paclitaxel) by rapid expansion of supercritical fluids (RESS). The Journal of Supercritical Fluids, 41(3), 440–451. Youssef, M. S., Aly, A. A., & Zeidan, E. B. (2013). Application of Artificial Neural Network Model in Calculation of Pressure Drop Values of Nanofluid. International Journal of Engineering and Technology, 3(5), 554–561. YU, X., Onder Effe, M., & Kaynak, O. (2002). A General Backpropagation Algorithm for Feedforward Neural Networks Learning. IEEE Transactions on Neural Networkd, 13(1), 251–254. Zhao, C., Wang, L., Zu, Y., Li, C., Liu, S., Yang, L., … Zu, B. (2011). Micronization of Ginkgo biloba extract using supercritical antisolvent process. Powder Technology, 209(1-3), 73–80. Zurada, J. M. (1992). Introduction to Artificial Neural Systems. West Publishing Company.


154