Overview of soft intelligent computing technique for supercritical fluid extraction

Optimization of Supercritical Fluid Extraction process with mathematical modeling is essential for industrial applications. The response surface methodology (RSM) has been proven to be a useful and effective statistical method for studying the relationships between measured responses and independent factors. Recently there are growing interest in applying smart system or artificial technique to model and simulate a chemical process and also to predict, compute, classify and optimize as well as for process control. This system works by generalizing the experimental result and the process behavior and finally predict and estimate the problem. This smart system is a major assistance in the development of process from laboratory to pilot or industrial. The main advantage of intelligent systems is that the predictions can be performed easily, fast, and accurate way, which physical models unable to do. This paper shares several works that have been utilizing intelligent systems for modeling and simulating the supercritical fluid extraction process.


INTRODUCTION
Traditional computational approaches and methods could only model and analyze simple systems. Complex systems such as in machining, biology, medicine and similar fields often remained unsolvable to conventional mathematical and analytical methods. Assumptions and simplification of mathematical model are sometimes not reliable and many conventional mathematical models have been both challenging and impractical. Soft computing however, deals with imprecision, uncertainty, partial truth and approximation to achieve tractability, robustness and more importantly low solution cost. Figure 1 shows the available components of soft computing. There are five types which are neural networks, fuzzy systems, evolutionary computation, ideas about probabilities and swarm intelligence. This study is will focus on an overview of previous research applying some of the components such as neural networks, fuzzy systems, genetic algorithm and its hybrid in supercritical fluid extraction process.
Conventional techniques are largely based on formal logical systems and rely heavily on computeraided numerical analysis. Soft computing techniques are intended to complement each other. Hard computing schemes strive for accurateness and full truth whereas soft computing techniques exploit the given tolerance of imprecision, partial truth, and uncertainty for a particular problem or data. In overall, soft computing technique is similar and imitating the genetic processes more closely than the traditional technique. It applies the ability to learn, observe and memorize in a situation of full of factors and important data.
Whereas, hybrid intelligence systems deal with the integration of two or more of the technologies. The combined use of technologies has resulted in effective problem solving in comparison with each technology used individually and exclusively. As illustrated in Figure 2, each of these technologies individually and in combination can be employed to solve problems. For example, when neural network combines with fuzzy systems, a neuro-fuzzy hybrid will develop and when neural network and evolutionary algorithm combines, a neuro-evolutionary will be developed.

PREVIOUS STUDY APPLYING SOFT INTELLIGENT COMPUTING TECHNIQUE
In order to develop a reliable model to represent the data, neural network have been applied to several supercritical fluid extraction processes in predicting yield of extraction [1][2][3][4][5][6] and solubility predictions [7][8][9]. Besides that, fuzzy systems also has been applied to determine the extractant quantities at desired temperatures and pressures [10].
There are several studies using a hybrid model combining ANN and FIS which is called as ANFIS to predict solubility [11] and mass of extract [12,13]. Another type of hybrid is by combining ANN with a conventional mathematical model [14]. This type of hybrid was mainly to solve the black box issues of not understanding the process of SFE. By ANN-Mathematical model hybrid, one can solve the issue [15].

119
Some studies even compare the results of application of ANN and ANFIS in order to identify the best system that can represent the data and reliable for optimization process [16][17][18][19].
Genetic algorithm also one of the promising tools in optimization of SFE [20][21][22][23][24]. It has been applied in determining constants in mathematical model in order to minimize error between model results and experimental data [25]. It can also be used in generating the non-linear binary interaction parameter of the conventional mathematical model such as in Peng-Robinson equation [26].

ARTIFICIAL NEURAL NETWORK (ANN) AND ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM (ANFIS)
Adaptive Neuro-Fuzzy Inference System (ANFIS) is considered as one type of hybrid model for optimization process since it combines the Artificial Neural Network (ANN) and Fuzzy Inference System (FIS) system. Table 1 describes the advantages and disadvantages applying ANN model independently. The 'black-box' nature of ANN models and unsatisfactory extrapolation by ANN models have led to the development of hybrid neural network models which combine ANNs with simple models. These are expected to perform better than ANNs in process identification tasks, since generalization and extrapolation are confined only to the uncertain parts of the process while the basic model is always consistent with first principles [28].
In a study on quercetin extraction from Rosa damascene Mill an experiment was conducted using a modified supercritical CO2 with ethanol as co-solvent [19]. The recovery of extraction was predicted using Adaptive neuro-fuzzy inference system (ANFIS) and Artificial neural network (ANN). The variables were translated into coded value of -2, -1, 0, +1 and +2. Table 2 shows the coded variables for this process. The coded variables are usually applied when RSM, ANN and ANFIS is going to be used as tools for optimization process. Generally, ANFIS architecture consists of 5 layers as shown in Figure 3. The architecture of ANFIS in Figure 3 maps the inputs through input membership functions (MF) and fuzzy rules. Similarly, output mapping is done through output membership functions along with its fuzzy rules. The number of membership function assigned to each input variable is chosen by trial and error. Figure 3 shows the general architecture of ANFIS for SFE system.
The hybrid of a neural network and fuzzy logic in ANFIS makes it have both, low-level learning and computational power of neural network and advantages of high-level human like thinking of fuzzy systems. Therefore, this neuro-fuzzy model able to overcome their individual disadvantages and can complement each other [29].

MATHEMATICAL MODEL-NEURAL NETWORK HYBRID
Hybrid models can also be combining both physical laws and observed measurements and including all available knowledge of the process. This model can be arranged in series or in parallel. In the series approach, heuristic model estimates unmeasured process parameters of mathematical modeling such that the first principle constrains are satisfied. In the parallel approach, the hybrid model prediction is combination of the outputs of the mathematical and initial heuristic models thus residuals between the process and the mathematical model is compensated [27]. This type of hybrid can solve the black box issues of neural network where the knowledge of the process is lacking. When combining both, the model will become more reliable and can represents the data well. A parallel hybrid model was developed by researchers from Iran [13] for prediction of Epigallocatechin gallate (EGCG) recovery of supercritical extraction which combines conventional mathematical modeling based on the differential mass balance in the solid and mobile phases with ANFIS ( Figure 4). Using analytical model in this structure, the accuracy of ANFIS would be increased in nontraining domains. Additionally, ANFIS solves the problem of unknown analytical parameters. In other words, the hybrid model could simulate the extraction system with having any arbitrary value of adjustable parameters of mathematical modeling [27]. This model has no knowledge on the whole process. To overcome the black box problem of the RBF model, the model is correlated with the Peng-Robinson equation becomes a hybrid RBF-PR model ( Figure 5). In the Peng-Robinson equation of state, there is unknown interaction parameter k12, for a binary mixture, whereby the predicted solubility is sensitive. The k12 can be obtained from the physical property data for the mixture, but it requires trial and error process for obtaining it. Besides, the parameters were all temperature dependent which does not really fit the data compared to pressure. The proposed hybrid model by [30] can compensate the black box problem. It can fit the experiment data very well and also keep the physical meaning of the whole SFE process. The detailed description on previous research applying ANFIS and the hybrid of it for yield and solubility prediction is shown in Table 3.

GENETIC ALGORITHM (GA)
Genetic algorithms are a class of optimization programs that can handle complicated problems. They consist of a random-search method, that made efficient by using explorative search. The method resembles evolutionary development in nature, using "mutation" and "mating". In contrast to expert systems no heuristic knowledge is applied other than the ultimate goal [32]. In 2016, a group of researchers in Iran did an optimization of essential oil extraction from Launaea acanthodes Boiss using genetic algorithm [33]. In their study, they applied the hybrid ANN-GA methodology whereby, ANN model was used as the fitness function, which measures the quality of individuals in the population. The model was then used to obtain the optimal operation conditions, such as pressure, temperature, flow rate, and co solvent besides determining the maximum extraction yield of essential oil from L. acanthodes. The general procedure to apply GA for most system was as shown in Figure 6. Table 4 has the description of previous application of GA in SFE for various materials.

Essential oil
Optimization of extraction yield ANN model was used as the fitness function, which measures the quality of individuals in the population [33] Soybean meal Isoflavone Optimization of profitability The objective function of this study is profitability which is more comprehensive than the cumulative isoflavone production rate [24] Green tea Epigallocatechin gallate Optimization of extraction recovery The fitness function for this study is extraction recovery [25] Macela flowers (Achyrocline satureioides)

Bioactive compound
Determination of constant Adjustable parameters, K, a, and b, were determined by criteria that the sum of the squares of the difference between the experimental extraction yield and the predicted extraction yield by the model should be minimized [28] Solid compounds -Optimization of the radius of influence of FIS The Genetic Algorithm (GA), which is an optimization algorithm, was implemented to measure the optimum value of radius of influence of FIS parameter [13] Ferulaga angulate -Optimization of yield The ANN model coupled with genetic algorithm (GA) was used to generate the optimal value [34] Benzoic acid -Dynamic optimization of process A dynamic optimization model for supercritical fluid extraction is proposed by combining a transformation based genetic algorithm and the Peng-Robinson equation of state [35] Int J Adv Appl Sci

CONCLUSION
Overall, the application of smart system or artificial technique to model and simulate a chemical process has been done in SFE process together with RSM method. Smart system models generalize the experimental result and present process behavior and finally predict and estimate problem. The model can be applied individually or by combining two or more models called as hybrid to achieve the best estimation and prediction data. The appropriate model is beneficial in the development of new products by saving experimental time and cost. However, research and study on dynamic process for SFE using artificial technique is lacking and the area is recommended for further research.