7420
Ind. Eng. Chem. Res. 2005, 44, 7420-7427
Experiments and Numerical Simulations of Supercritical Fluid Extraction for Hippophae rhamnoides L Seed Oil Based on Artificial Neural Networks Jian-Zhong Yin,*,† Qin-Qin Xu,† Wei Wei,† and Ai-Qin Wang‡ School of Chemical Engineering, Dalian University of Technology, Dalian 116012, P. R. China, and Dalian Institute of Chemical Physics, Chinese Academy of Sciences, Dalian 116023, P. R. China
In this paper, a supercritical fluid extraction setup with an extraction volume of one liter was established with which Hippophae rhamnoides L seed oil was extracted using supercritical CO2. The experiments show that many factors have an impact on the oil yield, such as extraction pressure, temperature, and fluid flow rate, as well as seed particle size and filling quantity. For the extraction process of H. rhamnoides L seed oil, the recommended conditions were as follows: extraction pressure of 20-30 MPa, extraction temperature of 35-40 °C, supercritical CO2 flow rate of 0.15-0.3 m3/h, and extraction time of 4-5 h. Under such conditions, the oil obtained is very lucid and of good quality, and the yield can be greater than 90%. Gas chromatography analysis shows that the oil contains 12.3% saturated fatty acid and 87.7% unsaturated fatty acid. From the changes of oil yield with the extraction time, it can be concluded that the extraction process contains three stages: the fast extraction (line), transitional, and slow extraction stages. At the first stage, 75-80% of the oil has been extracted. On the basis of the experimental results, artificial neural network technology was applied to the simulation of the supercritical fluid extraction of vegetable oil. With a three-layer back-propagation network structure, the operation factors, such as pressure, temperature, and extraction time, were used as input variables for the network, and the oil yield was used as the output value. In the optimization of the topological structure of the net, six neurals from the middle hidden layer have been proven to be the optimum value, according to the minimum training and running time. With the normalization pretreatment of the initial input data, not only the convergence speed and accuracy has been improved greatly but also the problem of the derivative at zero has been solved. Therefore, the method is better than that of Fullana. On the basis of simplification of the extraction process, taking account of axial dispersion, a differential mass balance kinetic model has been proposed. With Matlab software as platform, an artificial neural networks-supercritical fluid extraction simulation system has been programmed. For the first time, the simulation for the supercritical fluid extraction process of H. rhamnoides L seed oil has been made, and the results show that the average absolute relative deviation is lower than 6%. 1. Introduction Hippophae rhamnoides L (HRL) is largely grown in part of northeastern and northwestern China. It is often used as shelterbelt to protect the environment. From the medical point of view, its sarcocarps and seeds usually have high pharmacodynamic values, especially the H. rhamnoides L seed oil (HRLSO), which has been used to cure the cancer and enhance the immune system. Therefore, there is a great demand for the HRLSO in the world market. In 1997, HRL was adopted in the pharmacopoeia of People’s Republic of China. The cost of 50 g HRLSO is about $25.00. Thus, if the product yield is 10 t (approximately 200 t of raw materials 100 L × 3 equipment is needed to process this amount of material), the total cost is about $5 million. Usually, the HRLSO was extracted using organic solvents (such as, hexane, acetone, ethanol, etc.) or mechanical methods. However, these two approaches * To whom correspondence should be addressed. E-mail:
[email protected]. Fax: +86 411 83634309. † Dalian University of Technology. ‡ Chinese Academy of Sciences.
have some disadvantages, such as residual solvent and poor quality products. Therefore, the use of supercritical fluid extraction (SCFE), especially carbon dioxide, instead of traditional technologies becomes more interesting for industry. Early in its research, the extraction of vegetable oil was considered to be a diffusion operation involving the transfer of solute from a complex, the porous planttissue matrix, to the supercritical fluid (SCF) solvent.1 Lately, a new point of view says that the mass transfer mechanism was treated with a desorption kinetics process to control the SCFE of some organics from the solid matrix.2 In the SCFE field, seed oil extraction from various types of plants has been studied and many factors influencing the extraction rate have been investigated.3-10 Recently, much attention has been given to modeling the SCFE processes of natural products from a variety of plants. Several models have been proposed to describe the extraction rate using the mass transfer coefficient in either the solvent phase or the solid phase.9,11,13-20 Among these models, one of the most widely used is based on the mass balances of differential extraction beds. Bulley and co-workers9 have
10.1021/ie049196s CCC: $30.25 © 2005 American Chemical Society Published on Web 08/12/2005
Ind. Eng. Chem. Res., Vol. 44, No. 19, 2005 7421
investigated the extraction process of oils from canola with SCF and presented a model based on mass balances. For simplicity, they neglected the influence of axial concentration distribution and suggested that the control factor was the dissolution rate in the fluid phase when the grain was big and, if the grain was small, that the overall mass transfer coefficient could be used to describe the transfer process. Roy11 found that the mass transfer process differed greatly as the solute concentration changed in the solid phase. In the first period of the extraction, the concentration of the solute in the solid phase is high so the mass transfer is controlled by the dissolution equilibrium in the fluid phase. Goto12 and Sovova13 have the same viewpoint, and they thought that the concentration of the solute in the solid matrix and the solubility of solute in the fluid phase were the key factors in the whole extraction process. The mass balance model can be expressed in a series of differential equations, and its conventional solutions should be obtained by means of numerical analysis. Roy11 regarded the difference between the solubility and concentration of the solid phase as the driving force of the mass transfer, and he got a numerical result using a dimensionless transform with Crank Nicholson’s rule. Perrut14 determined the relationship between the extraction time and the solute concentration in the fluid phase and the solid phase, respectively. In the extraction of essential oil from cloves, Reverchon15 obtained the analytical solution using the Laplace transform. For the mass transfer rate as the function of the fluid phase concentration and its solubility, Sovova13 used the dimensionless transform to simplify the mass balances model and found an exact analytical solution for the concentration in the solid phase. Recently, Reverchon16 proposed a method to simulate the SCFE process in which only one adjustable parameter, ki, was used and microscopic information about the seed structure was considered at the same time. Cocerro17 used the Laplace transform and the finite differences method to obtain the mathematical resolution of the desorption model equations for SCFE. Yin20 also used the Laplace transform method to treat the simulation of SCFE for HRLSO. In general, however, the above models require experimental and parameter fitting. Although a few parameters can be obtained from correlations, the intrinsic rate and fluid-solid equilibrium data should have a rather firm experimental basis (i.e., from adsorption isotherms, rate coefficients, equilibrium solubility, etc.). Few parameters are fitted using a set of experimental data, and generally, the model obtained is not capable of extrapolating experimental information with good accuracy, an undesirable situation from an engineering point of view. During the last 10 years, different research groups and companies have been working on the application of artificial neural networks (ANN) for modeling chemical engineering processes, especially those presenting highly nonlinear behavior.21 An ANN is an information processing paradigm that is inspired by the way biological nervous systems, such as the brain, process information. The key element of this paradigm is the novel structure of the information processing system. It is composed of a large number of highly interconnected processing elements (neurons) working in unison to solve specific problems. ANNs, like people, learn by example. An ANN is configured for a specific application, such as pattern recognition or data
Figure 1. Schematic diagram of experimental apparatus. B is the pump, D1 and D2 are check valves, E1 is the condenser, E2 is the heat exchanger, F is the filter, J1-J4 are the on/off valves, JB1-JB3 are the needle valves (always closed), JY1-JY4 are the needle valves, L is the wet test meter, R is the regulator, P is the pressure gauge, T is the temperature meter, T1 is the extractor, T2-T3 are the separators, V1 is the CO2 cylinder, and V2-V3 are the sample collectors.
classification, through a learning process. Learning in biological systems involves adjustments to the synaptic connections that exist between the neurons. An ANN has some advantages such as generation characteristics, low computing time, etc. Recently, SCFE has become an interesting field for the application of ANN primarily because of the high sensitivity of the phenomenological models with regard to small changes in the process variables, which results in poor results. Fullana1 proposed a model based on ANN; it can simulate the SCFE process very well. However, they can not train their ANN system when there are zero points of input data. Therefore, near the original point of the coordinate, they still have to use the data fitting method. The aim of this work is to investigate the ANN simulation method for SCFE process and to improve the ANN model by changing the style of the initial sample data. First, the differential bed extraction runs were carried out under the condition of 15-30 MPa, 30-50 °C, and 10 g of seeds. Then, these experimental data were performed as original samples which were used to train the ANN. There are 176 data points in the total number of sample data. After the ANN system is trained using these data points, it will be tested byother experimental runs, not used as samples. Second, several integral bed runs were done in a manner similar to that used for the scale-up reactions with the same pressure and temperature conditions (but 120 g of seeds). Finally, the results of the ANN calculations were compared to those of the experimental runs to demonstrate the applicability of this method. 2. Experimental Section HRL seeds (produced in eastern area of Inner Mongolia, China) were milled and sieved to the appropriate particle size before use. They contain 5.01 wt % oil and 13.10 wt % water. The purity of carbon dioxide, supplied by Guangming Gas Plant, was better than 99.9%. Extraction measurements were carried out in a semibatch flow extraction apparatus (Figure 1). Supercritical carbon dioxide was used as the solvent. Liquid carbon dioxide from the supply cylinder (V) passes through a cold bath (E1, about -10 °C), and then, it is pumped with a two-plug pump (B) model 2JX-40/8 (Hangzhou,
7422
Ind. Eng. Chem. Res., Vol. 44, No. 19, 2005
P. R. China) and heated to the extraction temperature by a tubular heat exchanger (E2). The pressure is controlled with a back-pressure regulator (R, model H21X-100P, DW6, Yancheng, Jiangsu, P. R. China,). The experimental apparatus consisted, mainly, of a one liter internal volume extractor (T1) and two one liter separators (T2 and T3) operated in series. The analysis of the extracts was performed with gas chromatography (GC)-mass spectroscopy (MS), according to a tested procedure. The percentages of the various compounds were evaluated by areas on the GC trace without any correction factor. Two types of SCFE runs were done. In the first group, runs were performed using a differential bed with 50 cm i.d. and 3 cm long. In the second group, runs were performed using a 30 cm long bed. The extractor containing the raw material to be extracted was in a bath, thermostatically controlled by an electrical heating tape; the temperature inside the extractor was controlled by a digital controller (Yuyao, Zhejiang, P. R. China, model number TDA-8002) with an accuracy of (0.1 °C. The pressure at the exit of the extractor was measured using a pressure gauge with an accuracy of 0.2 MPa. After the stream of carbon dioxide, loaded with extract, left the extractor, it flowed through an on/off valve and a sequence of pressure expansion valves (needle valve, Yancheng, Jiangsu, P. R. China, model number WL21H-320P, DW6). The stream pressure was, in this way, reduced in three successive stages down to atmospheric pressure, and the oily extract was recovered in a glass collector. The water and volatile components were deposited in a second collector. The volume of carbon dioxide delivered was measured by a wet test meter (L, Changchun Meter Company, Jilin, P. R. China, model number LML-2) with an accuracy of (0.01 L. The pressure and temperature conditions were measured at the end of assembly. The estimated accuracy of the pressure measurement was (0.1 MPa, and temperature was measured with a mercury thermometer to within (0.1 °C. The extraction pressure, extraction temperature, particle size of the vegetable matrix, filling quantities, and solvent flow rate were identified as the parameters that mainly contribute to process optimization, and therefore, their effect on the yield of HRLSO has been investigated in detail. In this work, the oil yield is defined as the ratio of the mass of oil extracted to the mass of materials (water-free) filled into the extractor, and the unit of the oil yield is wt % or g of oil/100 g of seeds. 3. Mathematical Modeling Since the scale-up of the equipment and the evaluation of the cost of a process cannot be done without the mass transfer rate data in a convenient form, several methods have been proposed for modeling the supercritical extraction process. In this article, the modeling is based on the differential mass balance equation along the axial direction of the extractor bed.The operation conditions of pressure, temperature, and extraction time were treated as the input parameters, and the extraction yield was the output information of the ANN. The middle hidden layer neurons were used as a mapping functions for the input and output variables. More details for ANN can be found in the literature.1,21
3.1. Empirical Model. The modeling of the extraction process is based on the following hypotheses. (1) We assume that the behavior of all compounds extracted is similar and can be described by a single pseudo-component with respect to the mass transfer phenomena. (2) The concentration gradients in the fluid phase develop at larger scales than the particle size. (3) The solvent flow rate, with superficial velocity u, is uniformly distributed in all sections of the extractor. (4) The void fraction of the fluid, , is not affected by the reduction of the solid mass during extraction. By neglecting radial mass dispersion effects, we can write the the conservation equation of a solute in a packed bed as
F
∂C ∂C ∂2C + Fu0 - FDax 2 ) J ∂t ∂z ∂z
(1)
in which the axial mass dispersion is considered. In eq 1, J represents the rate of solute mass transfer from the outer surfaces of the seeds to the fluid per unit of bed volume. Assuming that the fluid travels in plug flow at the bed entrance and is free from solute, the initial and boundary conditions for the above problem are
C|z,t)0 ) 0 -FDax
(∂C∂z ) (∂C∂z )
z ) 0+,t
+ Fu0C|z)0+ ,t ) 0
z ) L,t
)0
(2a) (2b) (2c)
The mass transfer rate, J, will be written in terms of a product of the overall mass transfer coefficient, U, and the specific mass transfer area, A. So, UA will be the mass transfer rate per unit driving force, expressed as kg of solute/m3 bed s. The driving force is C* - C, hence J can be written as
J ) UA(C* - C)
(3)
In eq 3, C* is the mass fraction of solute oil in the fluid at saturation for the system temperature (solubility). A net trained with the differential data, extraction yield, pressure, temperature, and time, is available for neural computing. The vertical packed-bed extractor is divided into horizontal slices or layers, each containing approximately the same amount of seeds as that in the thin beds used in differential runs. If the amount of extracted oil from a layer is known, at given pressure and temperature, then the time elapsed after the start of extraction for any layer in the pile of layers will be calculated iteratively. The quantity of oil extracted is calculated from the equation k k-1 k ) qex + (Cki - Ci-1 )Q∆t qex j j
(4)
where j is the node number between layers i - 1 and i, k is the time instant, ∆t is the time elapsed between the instant k - 1 and k, and Q is the mass flow rate of fluid, which is assumed to be constant. By knowing the amount of seeds and the fluid flow rate in a differential bed, it is possible to infer the concentration of the solute in the fluid in contact with layer i. The expression for the hybrid neural concentration will be written as
Ind. Eng. Chem. Res., Vol. 44, No. 19, 2005 7423
CNN i
dq q ( dt ) ) ti
s
Q
(5)
where CNN is the oil concentration expressed by ANN i in a layer i, qs is the mass of seeds filled in a differential bed, q is the amount of oil extracted (kg of oil/kg of seeds), and dq/dt is the rate of oil extraction at time t. Finally, to integrate eq 1, it is necessary to have a value for the mass transfer rate given by eq 3. This will be so provided that fluid velocity is the same in the integral bed and in the differential extraction runs. The value used for UA for a layer i in the calculations is obtained from the following relationships NN CNN i Q ) Ji V
(UA)i )
Q/V (C*/CNN - 1) i
(6) (7)
is the oil concentration expressed by ANN where CNN i is the mass transfer rate in a layer i (kg/kg), JNN i expressed by ANN in a layer i (kg of solute/m3 s), and (UA)i is the mass transfer rate per unit driving force in a layer i (kg of solute/m3 of bed s). In the programs, the following partial calculations or modules were implemented. The axial dispersion coefficient was calculated from a correlation (Catchpole22). The solubility of oil in SCF was calculated with the del Valle23 method. The integration of eq 1 was made by using the method of lines in combination with the Crank-Nicolson algorithm. The fluid density was estimated using the Peng-Robinson equation of state (P-R EOS). Neural net computations were obtained from differential bed modeling. In this article, the ANN model was from Fullana.1 For this technology, a 3-layer backpropagation (BP) network structure is applied, and the operation factors such as pressure, temperature and extraction time are used as the input variables for the network, whereas the yield of oil extracted was treated as the output value for the network. In the optimization of the topological structure of the net, six neurals in the middle hidden layer have proven to be the optimum value, according to the minimum training time (the number of iterations) under the same simulation precision, Figure 2. 3.2. Normalization Pretreatment of the Initial Input Data. For the SCFE process, we should use the pressure, temperature, and time as the input parameters and the extraction yield as the output parameter for the ANN system when training the ANN system. Under our experimental conditions, the pressure, P, is between 15 and 30 MPa, the temperature, T, is between 30 and 50 °C, the time, t, is from 0 to 9000 s, and extraction yield, q, lies between 0 and 5 g. By comparing these four parameters, we found that there were infinite differences among them. This can make the network produce a huge difference for the joint weight in every mode. This input and output data could reduce the properties of the ANN system, make the ANN training very difficult, and at the same time, decrease the accuracy of the ANN prediction to SCFE process. On the other hand, the ANN system will not be trained very well when there are zero point input data. This is the reason that Fullana and co-workers1 called their method a “neural-regressive hybrid prediction system”. In this article, to overcome the above difficulties and make an
Figure 2. Training and learning curves for (a) 5 units and (b) 6 units.
improvement on Fullana’s method, we, for the first time, used the normalization method to modify the initial input and output data. The detailed formula is
xi - xi,min x/i ) 0.9 + 0.05 xi,max - xi,min
(8)
where xi,min is the minimum value of the ith input data, xi,max is the maximum value of the ith input data, and x/i is the new value after normalized treatment. It is obvious that all data can be mapped into [0.05, 0.95] after normalized treatment. The data not only have no huge differences but also have no zero point for ANN system training. Figure 2 shows the error (in terms of the average absolute relative deviations, AARD %) vs the number of iterations for back-propagation with 5 and 6 units in the hidden layer. From Figure 2, we can determine that the prediction of the accuracy and the overall convergent rate are very well. When training and simulating for the ANN system, the factor value of the weight function can be chosen from -0.5 to 0.5 randomly until the average total error between the real value and the expectation value should be a minimum. Similarly, the convergent speed and calculating time of the ANN system was strongly dependent on the learning factors. Usually, the learning factors can be any values from 0.8 to 1.0. In our work, the learning factors are given as 1.0 for easy in training the ANN (in fact, in many cases, the learning factor was selected to be 1.0).
7424
Ind. Eng. Chem. Res., Vol. 44, No. 19, 2005
Figure 3. Effects of temperature on the extraction efficiency for P ) 15 ([), 20 (9), 25 (0), and 30 MPa (O) with a particle size of 28 mesh (0.83 mm), a fluid flow rate of 0.2 m3/h, STP, and of filling fraction of 75 vol %.
Figure 4. Effects of pressure on the extraction efficiency for T ) 30 (b), 35 (0), 40 (O), and 50 °C (9) with a particle size of 28 mesh (0.83 mm), a fluid flow rate of 0.2 m3/h, STP, and a filling fraction of 75 vol %.
4. Results and Discussion 4.1. Experimental. 4.1.1. Effect of Temperature. The effect of the extraction temperature on the oil yield is illustrated in Figure 3. It can be seen that the oil yield first increased as the temperature increased, attained a maximum value at 40 °C, and then decreased as the temperature further increased. This change of the oil yield with the temperature is the result of the coupling effect of temperature and pressure. On one hand, the increasing temperature resulted in the decrease of solvent density and thus a decrease of the solubility of the seed in SCF. On the other hand, the saturation pressure of the solute in SCF increased with the temperature increase, which improved the solubility. 4.1.2. Effect of Pressure. The effect of the extraction pressure on the oil yield is shown in Figure 4. The experiments were performed at four different temperatures and various pressures. The particle size is fixed at 20 mesh (about 0.83 mm), and the CO2 flow rate is fixed to 0.2 m3/h. It is clear that with the increase of pressure, the oil yield increased. It is well-known that with the increase of pressure, the density of SCF CO2 increased and the solubility of the solute increased. Furthermore, according to the P-R EOS, the variation
Figure 5. Effects of particle size on the extraction efficiency for P ) 30 MPa and T ) 35 °C for a fluid flow rate of 0.2 m3/h, STP and a filling fraction of 75 vol %.
Figure 6. Effects of filling quantity on the extraction efficiency for P ) 30 MPa, T ) 35 °C, and t ) 4 h with a flow rate of 0.2 m3/h, a particle size of 28 mesh (0.83 mm), and a fluid flow rate of 0.2 m3/h, STP.
of the density with pressure became significant between 5 and 25 MPa. For the practical application, we propose that 25 MPa should be employed. 4.1.3. Effect of Particle Size. Besides temperature and pressure, the particle size has a critical impact on the extraction efficiency. In the present work, we investigated the particle size effect at the pressure of 30 MPa and a temperature of 35 °C. As shown in Figure 5, the smaller the particle size, the higher oil yield was obtained. Especially for the particle size between 18 and 36 mesh (about 0.93-0.40 mm), the oil yield increased rapidly with the particle size decrease. Decreasing the particle size can improve the mass transfer, while the mass transfer resistance increases. Therefore, a suitable particle size should be employed. 4.1.4. Effect of Filling. Figure 6 shows the effect of filling on the oil yield. It can be seen that with a decrease of filling volume, the oil yield increased. This is because the smaller filling volume can enlarge the contact area between the fluid and the solid, thereby improving the mass transfer. When the filling fraction is about 45%, the oil yield after a 4 h extraction is about 4.2%. While when the filling fraction is 75%, the oil yield went down to 3.2%, after the same extraction period.
Ind. Eng. Chem. Res., Vol. 44, No. 19, 2005 7425
Figure 7. Effects of operating time on the extraction efficiency at T ) 35 °C for P ) 15 (0), 20 (b), 25 (O), and 30 MPa (9) with a particle size of 28 mesh (0.83 mm), a fluid flow rate of 0.2 m3/h, STP, and a filling fraction of 75 vol %. Table 1. Fatty Acids Composition of the Hippophae rhamnoides L. Seed Oil saturated fatty acids (12.293%)
unsaturated fatty acids (87.707%)
nutmeg
palmitic
stearic
oleic
linoleic
flaxen
0.14
9.7
2.2
23.1
36.6
26.2
However, from the economic point of view, a filling fraction of 75-80% is suitable. 4.1.5. Effect of Extraction Time. Figure 7 shows the yield vs time curves under the different operating conditions at the fixed flow rate of 0.2 m3/h. From Figure 7, it can be seen that the extraction process is composed of three stages: rapid extraction of free solute, transitional stage of surface and internal diffusion, and slow extraction based mainly on the internal diffusion.18-20 The time consumed in the first extraction stage depends both on the solute solubility in SCF CO2 and on the particle size. Under our experimental conditions, most parts of the seed oil was extracted in the first stage at 150-180 min. 4.1.6. Chemical Components Analysis. Extracted oil was analyzed by GC (HP-4890) with a column of PEG-20M (30 m length × 0.25 mm i.d.). The analysis was performed under the following conditions: the column temperature ramped from 70 to 280 °C, the detector temperature was maintained at 205 °C, and nitrogen was used as the carrier gas. Before analysis, the samples were pretreated by means of methyletherification and a 0.6 µl portion was injected into the GC. The chemical compositions of the extracted HRLSO, using the SCF process and the organic solvent extraction method separately, are listed in Table 1. The content of the saturated fatty acid is about 12.3% and that of the unsaturated fatty acid is about 87.7%. 4.2. Simulation. In this article, an ANN system was developed to simulate SCFE processes on the basis of Matlab 5.3 software (Figure 8.). With Matlab 5.3 software, which has an ANN Tool Box, we can make our own special program very conveniently. Therefore, the kinetic model of ANN was established on the basis of a differential mass balance of the packed bed. Using a 3-layer BP network with extraction pressure, temperature, and extraction time as the input variables and
Figure 8. Structural program diagram for the PDE (partial differential equation).
the extraction yield as the single response for the net, the network was trained with experimental data (there are a total of 176 data points). The results show that the trained network can simulate the extraction rate of the HRLSO. First, the differential bed extraction runs were carried out under the conditions of 15-30 MPa, 30-50 °C, and 10 g of seeds. Then, these experimental data were performed as original samples used to train the ANN. For the differential bed extraction, two types of calculations were carried out with a given neural network. In the first one, the samples given in Table 2 were read as input-output pairs. Then the pairs used for learning were coded, and the weights between the neural elements were evaluated randomly. From the first pair to the last one, their outputs were generated as weights were changed because of the difference in desired output. After the last pair had been done, the error was calculated for which the errors for all of the pairs tested were required. If the error was larger than the tolerance, the process was repeated until the calculation convergence. Finally, the weights were written for use in the normal operation of the network. The total number of sample data points is 176 (Table 2). After the ANN system was trained using these 176 data points, it was tested by another set of experimental runs, which had not been used as samples. Several integral bed runs were done in a manner similar to that of the scale-up runs with the same pressure and temperature conditions (but with 120 g of seeds). Finally, the results of the ANN calculations were compared to those of the experimental runs to demonstrate the applicability of this method (Table 3). On the basis of the integral bed extraction process analysis, the network can make a good prediction for the change of the fluid concentration with the bed position and extraction time. Using this system to train and predict our experimental data (for sample data), we find AARD values of less than 0.23% (for the training group) and 0.5% (for the test group). Figure 9 shows the relationship between the fluid concentration and the extraction bed layers, where the advancement of the mass transfer zone is seen for a bed of 30 cm in length. This simulates the integral bed using differential kinetic
7426
Ind. Eng. Chem. Res., Vol. 44, No. 19, 2005
Table 2. Experimental Data as Training Samples for the ANN Solventa yield (wt %) at time (min) no.
P (MPa)
T (°C)
0
10
20
30
40
50
60
70
80
90
100
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
15
30 35 40 50 30 35 40 50 30 35 40 50 30 35 40 50
0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
0.230 0.365 0.471 0.318 0.230 0.365 0.471 0.318 0.230 0.365 0.471 0.318 0.230 0.365 0.628 0.318
0.460 0.739 0.884 0.612 0.522 0.739 1.028 0.612 0.575 0.839 1.058 0.690 0.575 0.895 1.266 0.690
0.694 1.151 1.295 0.930 0.771 1.151 1.588 0.930 0.893 1.313 1.659 1.085 0.893 1.432 1.936 1.112
0.941 1.534 1.726 1.253 1.063 1.640 2.140 1.306 1.212 1.841 2.288 1.522 1.212 1.967 2.574 1.575
1.189 1.834 2.061 1.588 1.351 2.060 2.681 1.658 1.567 2.343 2.953 1.923 1.567 2.522 3.234 2.048
1.371 2.109 2.334 1.835 1.595 2.505 3.119 1.983 1.841 2.789 3.591 2.406 1.915 3.038 3.924 2.585
1.496 2.301 2.532 1.956 1.726 2.877 3.452 2.301 2.071 3.222 4.143 2.831 2.186 3.544 4.545 3.015
1.496 2.301 2.532 1.956 1.726 2.877 3.452 2.301 2.071 3.222 4.143 2.877 2.234 3.613 4.591 3.107
1.496 2.301 2.532 1.956 1.726 2.877 3.452 2.301 2.071 3.222 4.143 2.877 2.234 3.613 4.591 4.591
1.496 2.301 2.532 1.956 1.726 2.877 3.452 2.301 2.071 3.222 4.143 2.877 2.234 3.613 4.591 4.591
a
20
25
30
Flow rate: 0.2 m3/h. Particle size: 20 mesh (0.83 mm). Mass of seed: 10 g.
Table 3. Comparison of the Experimental Data and the ANN Calculations for Testinga yield, wt % 15MPa, 35°C
20MPa, 35°C
25MPa, 35°C
30MPa, 35°C
time (min)
exptl
ANN
ARD (%)
exptl
ANN
ARD (%)
exptl
ANN
ARD (%)
exptl
ANN
ARD (%)
0 30 60 90 120 150 180 210 240 270 300
0.000 0.150 0.414 0.725 1.105 1.484 1.822 2.206 2.417 2.428 2.440
0.000 0.155 0.403 0.736 1.099 1.530 1.910 2.301 2.422 2.417 2.451
0.00 3.85 -2.78 1.59 -0.52 3.10 4.86 4.33 0.24 -0.47 0.47
0.000 0.334 0.713 1.059 1.438 1.853 2.206 2.494 2.781 2.783 2.784
0.000 0.322 0.679 1.093 1.404 1.956 2.313 2.417 2.854 2.877 2.877
0.00 -3.45 -4.84 3.26 -2.40 5.59 4.85 -3.09 2.61 3.39 3.35
0.000 0.357 0.817 1.254 1.726 2.267 2.685 3.165 3.548 3.549 3.550
0.000 0.368 0.806 1.323 1.761 2.359 2.808 3.337 3.452 3.487 3.682
0.00 3.23 -1.41 5.50 2.00 4.06 4.59 5.45 -2.69 -1.75 3.73
0.000 0.483 1.036 1.726 2.359 2.992 3.498 3.959 4.315 4.327 4.338
0.000 0.460 1.024 1.784 2.474 3.107 3.682 4.143 4.304 4.373 4.384
0.00 -4.76 -1.11 3.33 4.88 3.85 5.26 4.65 -0.27 1.06 1.06
a
Solvent flow rate: 0.35 m3/h. Particle size: 20 mesh (0.83 mm). Mass of seed: 120 g.
Figure 9. Concentration curves of beds simulated at different time intervals.
data, scaled up from laboratory values. As soon as the concentration wave reaches the bed exit, the solute starts to appear in the leaving stream resulting in the breakthrough curve. These phenomena are in accordance with the theoretical analysis result for the packed bed mass transfer processes. The results of the experiments and the ANN simulations are also compared in the paper, shown in Figure 10. From Figure 10, under the conditions of 35 °C, 15-30 MPa, and 0.2 m3/h, we calculated the oil yield as a function of the extraction time by ANN. It is indicated that the ANN system can predict the SCFE process very well, and the maximum error of AARD is less than 6%.
Figure 10. Comparison of simulation and experiments at T ) 35 °C for P ) 15 (]), 20 (0), 25 (4), and 30 MPa (×) with a particle size of 28 mesh (0.83 mm), a fluid flow rate of 0.2 m3/h, STP, and a filling fraction of 75 vol %.
5. Conclusions In this work, the artificial neural networks technology was applied to the simulation of the SCF extraction process of vegetable oil. For this technology, a 3-layer BP network structure is applied, and the operation factors, such as pressure, temperature, and extraction time, are used as input variables, whereas the oil yield
Ind. Eng. Chem. Res., Vol. 44, No. 19, 2005 7427
of the extraction is treated as the output value. In the optimization of the topological structure of the net, six neurals in the middle hidden layer have proven to be the optimum value. With the normalization pretreatment of the initial input data, not only the convergence speed and accuracy has been improved greatly but also the problem of a derivative at zero has been solved. Therefore, the method is an improvement on the model of Fullana.1 In addition, we investigated the impact of the output variables on the net properties and found that it is very convenient for the data correction when the extraction rate-time curve is set as training sample. An ANN-SCFE simulation system has been programmed. For the first time, the simulation for the SCFE process of HRLSO has been made, and the results show that the AARD is lower than 6%. Acknowledgment The authors gratefully thank the Natural Science Foundation of Liaoning Province, China for financial support (20031072). Nomenclature A ) specific mass transfer area, m-1 C ) solute concentration in fluid phase, kg/kg CNN ) oil concentration expressed by ANN in a layer i, kg/ i kg C* ) solubility, kg/kg C/C* ) Ratio of the oil concentration to the oil solubility Dax ) axial dispersion coefficient, m2/s J ) mass transfer rate, kg of solute/m3 s JNN ) mass transfer rate expressed by ANN in a layer i, i kg solute/(m3s) q ) mass of oil per unit weight of seeds, kg/kg qs ) mass of seeds in a differential run, kg Q ) solvent mass flow rate, kg /s t ) time, s u0 ) supercritical fluid velocity, m/s U ) overall mass transfer coefficient, kg/m2 s V ) volume of differential bed, m3 z ) axial coordinate, m Greek Symbols ) bed void fraction F ) fluid density, kg/m3
Literature Cited (1) Fullana, M.; Trabelsi, F.; Recasens, F. Use of Neural Net Computing for Statistical and Kinetic Modelling and Simulation of Supercritical Fluid Extractors. Chem. Eng. Sci. 2000, 55, 79. (2) Al-Jabari, M. Kinetic Models of Supercritical Fluid Extraction. J. Sep. Sci. 2002, 25, 8, 477. (3) Taniguki, M.; Tsuji, T.; Shibata, M.; Kobayashi, T. Extraction of Wheat Germ with Supercritical Carbon Dioxide. Agric. Biol. Chem. 1985, 49, 2367. (4) Christianson, D. D.; Friedrich, J. P.; List, G. R.; Warner, K.; Bagley, E. B.; Stringfellow, A. C.; Inglett, G. E. SFE of Dry Milled Corn Germ with Carbon Dioxide. J. Food Sci. 1984, 49, 229.
(5) List, G. R.; Friedrich, J. P.; Pominsky, J. Characterization and Processing of Cottonseed Oil Obtained by Extraction with Supercritical Carbon Dioxide. J. Am. Oil Chem. Soc. 1984, 61, 1847. (6) Snyder, J. M.; Friedrich, J. P.; Christianson, D. D. Effect of Moisture and Particle Size on the Extractability of Oils from Seed with Supercritical CO2. J. Am. Oil Chem. Soc. 1994, 61, 1851. (7) Eggers, R.; Sievers, U.; Stein, W. High-Pressure Extraction of Oil Seed. J. Am. Oil Chem. Soc. 1985, 62, 1222. (8) Favati, F.; King, J. W.; Mazzanti, M. Supercritical Carbon Dioxide Extraction of Evening Primerose Oil. J. Am. Oil Chem. Soc. 1991, 68, 422. (9) Bulley, N. R.; Fattori, M.; Meisen, A.; Moyls, L. Supercritical Fluid Extraction of Vegetable Oil Seeds. J. Am. Oil Chem. Soc. 1984, 61, 1362. (10) Yin, J. Z.; Xun, X. W.; Ding, X. W.; Liang, H. H. Modeling of Supercritical Fluid Extraction from Hippophae Rhamnoides L. Seeds. Separ. Sci. Technol. 2003, 38, 4041. (11) Roy, B. C.; Goto, M.; Hirose, T. Supercritical Fluid Extraction Oils from Ginger Root: Experiments and Modeling. Ind. Eng. Chem. Res. 1996, 35, 607. (12) Goto, M.; Sato, M.; Hirose, T. Extraction of Peppermint Oil by Supercritical Carbon Dioxide. J. Chem. Eng. Japan 1993, 26, 401. (13) Sovova, H. Rate of Vegetable Oil Extraction with Supercritical CO2sI. Modelling of Extraction Curve. Chem. Eng. Sci. 1994, 49, 409. (14) Perrut, M.; Clavier, J. Y.; Poletto, M.; Reverchon, E. Mathematical Modeling of Sunflower Seed Extraction by Supercritical CO2. Ind. Eng. Chem. Res. 1997, 36, 430. (15) Reverchon, E.; Daghero, J.; Marrone, C.; Mattea, M.; Poletto, M. Supercritical Fractional of Fennel Seed Oil and Essential Oil: Experiments and Mathematical Modeling. Ind. Eng. Chem. Res. 1999, 38, 3069. (16) Reverchon, E.; Marrone, C. Modeling and Simulation of the Supercritical CO2 Extraction of Vegetable Oils. J. Supercrit. Fluids 2001, 19, 161. (17) Cocero, M. J.; Garcia, J. Mathematical Model of Supercritical Extraction Applied to Oil Seed Extraction by CO2 + Saturated AlcoholsI. Desorption Model. J. Supercrit. Fluids 2001, 20, 229. (18) Reverchon, E. Mathematical Modeling of Supercritical Extraction of Sage Oil. AIChE J. 1996, 42, 1765. (19) Sovova, H.; Kucera, J.; Jez, J. Rate of the Vegetable Oil Extraction with Supercritical CO2sII. Extraction of Grape Oil. Chem. Eng. Sci. 1994, 49, 415. (20) Yin, J. Z.; Wang, A. Q.; Wei, W.; Liu, Y.; Shi, W. H. Analysis of the Operation Conditions for Supercritical Fluid Extraction of Seed Oil. Sep. Purif. Technol. 2005, 43, 163. (21) Bhat, N. V.; McAvoy, T. J. Use of Neural Nets for Dynamic Modeling and Control of Chemical Process Systems. Comput. Chem. Eng. 1988, 14, 573. (22) Catchpole, O. J.; Andrews, E. W.; Toikka, G. N.; Wilkinson, G. T. In Proceedings of the 3rd International Symposium on Supercritical Fluids; Perrut, M.; Brunner, G. Eds; International Society for the Advancement of Supercritcal Fluids: Nancy Cedex, France, 1994; 47. (23) del Valle, J. M.; Aguilera, J. M. An Improved Equation for Predicting the Solubility of Vegetable Oils in Supercritical CO2. Ind. Eng. Chem. Res. 1988, 27, 1551.
Received for review August 31, 2004 Revised manuscript received June 8, 2005 Accepted June 27, 2005 IE049196S
