Full-grown samples of freshwater carp (Cyprinus carpio L.) were procured from a local lake in the central part of North Macedonia. Obtained fish specimens were scaled and eviscerated subsequently. Extracted viscera material was ground using a commercial mincing machine with a 4 mm sieve to produce sufficiently homogenised natural matrixes. These have an increased mass-transfer contact surface area required to complete subsequent freeze-drying and separation processes successfully. Prepared viscera matrixes were measured in mass (approximately 20 g) and appropriately stored in refrigerated conditions without moisture and light.

Each viscera batch was subjected to a freeze-drying (lyophilisation) process to eliminate the high moisture content in pre-defined matrixes. The high moisture content in eviscerated material was undesirable in the closed system SC-CO₂ – viscera matrixes, as it directly affects the overall polarity and the mechanisms of the extraction process. The freeze-drying process was designed and regulated to produce treated material portions with moisture content not higher than 5% (w/w) and enhance material stability.

Lab-scale freeze dryer − Freeze Zone Freeze Dryer LABCONCO was used to lyophilise the eviscerated material. The optimal freeze-drying process conditions were experimentally determined at 72 h and an operating P of 13.3 kPa. The process runs in three consecutive steps:

• Initial material freezing at an operating temperature of 233 K;

• Online regulated lyophilisation at 233 K and 13.3 kPa with a duration of 60 hours;

• Cascade-regulated freeze-drying within a temperature region of 233 K to 293 K for 10 hours.

A laboratory-scale extraction in a supercritical fluid unit produced by NOVA-Swiss, HPEP 565.0156 (Nova Werke LTD, Effertikon) was used for the isolation of PUFA from the investigated raw material – carp viscera samples (Figure 1).

Figure 1.

Lab-scale SFE-CO₂ unit HPEP 565.0156; GC − gas tank (CO₂), CU − compressor unit, C − diaphragm compressor, E − extractor, S − separator, UT1, UT2 − ultra thermostats, HE − heat exchanger, RD − ruptured disk, FI − flow measuring instrument, PI – pressure measuring instrument, TI – temperature measuring instrument

A pre-prepared matrix of lyophilised viscera material (20±1 g) was placed between two layers of 4 mm glass beads to eliminate the void volume in the extractor. The total capacity of the SFE extractor unit is 200 ml. A stabilisation period of 30 min was defined to provide uniform thermal distribution within the extractor and the necessary initial contact time in the system solvent − natural matrix. The SC-CO₂ extraction process was performed according to a pre-defined experimental plan regarding the operating values of the process parameters: P (20, 30, 35, and 40 MPa), T (313, 323, and 333 K), and Q_CO2 (3.23, 4.62, and 5.9 g/min). The SFE-CO₂ separator unit was operated at constant P (2 MPa) and T (298 K), as these values of operating parameters are in the range of optimal P-T conditions for the described closed system. The resulting total extracts (252 separate samples) were collected and stored in borosilicate dark glass vials at 253 K for further instrumental analysis.

The total extract yield (Y_TOTAL), as well as the yield of PUFA (Y_PUFA) and PUFA solubility (R_PUFA), were mathematically defined according to equations (1), (2) and (3):

$Y_{TOTAL}\left[\%\right]=\frac{m_{extract}}{m_{sample} \left(dry basis\right)}\times 100$ (1)

$Y_{PUFA}\left[\%\right]=\frac{m_{extract}\times x_{PUFA}}{m_{sample} \left(dry basis\right)}\times 100$ (2)

$R_{PUFA}\left[\text{mg PUFA}/{\text{dm}}^3{}_{\text{N}}{\text{CO}}_2\right]=\frac{m_{extract}\times x_{PUFA}}{V_{C{O}_2}}\times 100$ (3)

Where: m_extract – mass of extract[g]; m_sample (dry basis) – mass of lyophilised raw material [g]; x_PUFA – mass fraction of PUFA in the extract; Y_TOTAL – yield of extract [g/100 g lyophilised raw material]; Y_PUFA – yield of PUFA [g/100 g lyophilised raw material]; R_PUFA – solubility of PUFA [mg PUFA/dm³ _NCO₂], where _NCO₂ – CO₂ at NTP conditions, 101.325 kPa and 293 K; V_CO2 – volume of CO₂ [dm³] used in the extraction process, calculated from the formula $V_{C{O}_2}=\left(Q_{C{O}_2}/{\rho }_{C{O}_2}^{NTP}\right)\times t$, whereQ_CO2 – mass flow rate of CO₂ used in the extraction $\left[g/\min \right],{\rho }_{C{O}_2}^{NTP}$ – density of CO₂ at NTP [g/dm³], t – extraction time [min].

The SFE-CO₂ extracts from the viscera matrixes were investigated to characterise their fatty acid profile and quantify the presence of ω-3 (including EPA and DHA) and ω-6 PUFA. The analysis followed the AOAC Official Method 996.06 Gas Chromatography with Flame Ionisation (GC/FID) [2] using an Agilent 7890A unit equipped with a J&W HP-88 GC column: 60 m (length), 0.25 mm (internal diameter), 0.20 µm (particle size), 7-inch cage. The exact volume of 1 µl total extract was used for each chromatographic analysis, performed according to the previously stated AOAC Official Method: initial temperature of 393 K (1 min), sequenced by 10 K/min gradient rise to 448 K (10 min), then 5 K/min rise to 483 K (5 min) and a final 5 K/min rise to 503 K (12 min). The working temperatures for the injector and detector were 523 K and 573 K, respectively. The carrier gas was helium, with a 0.8 ml/min flow rate.

The experimental analysis correlated to the isolation yield of multicomponent extract through the SFE-CO₂ on analysed freeze-dried viscera samples produced satisfactory results for this output. The value of the whole extract yield was maximal at 40 MPa and 333 K, Y_TOTAL = 51.5 g/100 g lyophilised material. All experimental results were calculated as the average of three experiments. Discussed experimental data points indicate that each obtained result for the observed response was confined to statistically insignificant processing error (less than 2%), thus providing sufficient validity of the results and adequacy of the employed extraction system design. This repeatability is valid for all values of the entire data set. The results from the established experimental plan produced the required data set for the ANN model. The initial experimental work provided the needed data depicting the experimental conditions and obtained extraction yield outputs, eq. (1). Each produced extract was subjected to GC/FID analysis to determine the fatty acid profile. Figure 2 presents one of the chromatograms, generated at optimal operating conditions regarding the presence and resulting R_PUFA as a function of the operating process variables.

Figure 2.

Chromatogram of SFE-CO₂ extract (P = 35 MPa, T = 313 K, Q_CO2 = 3.23 g/min, t = 180 min)

The interpretation of the results obtained from the presented chromatogram regarding the resulting fatty acid profile of the fish oil is given in Table 1. Considering all aspects regarding the operating variable values and the comparative analysis of the resulting Y_PUFA and R_PUFA values serve as a basis for process modelling and optimisation. The optimisation includes the economic aspect of the investigated process, and therefore, Table 1 and Figure 2 present the fatty acid profile and the chromatogram obtained at P = 35 MPa, T = 313 K, Q_CO2 = 3.23 g/min. These operating conditions require low operating costs − lowest values for operating temperature and CO₂ flow rate, whereas the operating P enables high value for Y_PUFA extracted per Q_CO2 (4.77±0.043 mg/g CO₂).

Analysis of the experimental results suggests the satisfactory presence of highly sought-after ω-3 (including EPA and DHA) and ω-6 PUFA, thus confirming the established SFE-CO₂ process as an adequate separation system for the analysed raw material. Experimentally derived results concerning the total extraction yield coupled with the generated chromatograms of implemented instrumental analysis were the basis for the calculation of the Y_PUFA, eq. (2). Considering the physical properties of gaseous CO2, as the solvent mass flow rate was measured at the plant exit (at NTP operating conditions), PUFA yield results were used to calculate the solubility of these bioactive components in SC-CO2, eq. (3). An excerpt of these results is presented in Table 2. According to conducted experimental work, the data regarding Y_PUFA obtained at various operating conditions and total extraction time of 180 min represents the calculation basis for R_PUFA, according to previously defined equations (1) to (3). Initial calculations present the mass of polyunsaturated fatty acids obtained per unit mass of CO₂(m_PUFA/m_CO2 [mg PUFA/g CO₂]) and these values are then used to express the studied response, R [mg PUFA/dm³ _NCO₂].

The process of ANN creation consisted of three steps: training by the introduction of experimentally obtained data, validation, and determining the deviation of the ANN-generated output from the experimental data. The task of the ANN model is to predict the solubility of PUFA in supercritical CO₂ during SFE of biologically active components from lyophilised tissue samples. For this purpose, MATLAB − Neural Network Toolbox software kit was employed. The created ANN consists of three neural layers: the input layer, the hidden (operative) layer, and the output layer. The input layer is defined by the experimental variables (P, T, Q_CO2) matrix, whereas the output layer represents the resultant PUFA solubility (R_PUFA). From an engineering point of view, the main goal was to design a hidden (operative) layer and to define the transfer functions for the artificial neurons to generate an optimal ANN architecture for the investigated system. The structural architecture of the operational (hidden) layer of the designed ANN in terms of the number of neurons resulted from the minimisation of the mean squared error (MSE) function of output values versus experimental data (Figure 3).

Figure 3.

MSE as a function of the number of neurons in the hidden layer

Considering the network performance related to the number of neurons in the operative (hidden) neural layer presented in Figure 3, one can conclude that the optimal number of neurons in this ANN layer is 10. Thus, the minimum value of the MSE is achieved using the 3-10-1 architecture of the ANN. Table 3 shows the designed ANN’s architectural properties, including the designated transfer functions for the hidden and output neural layers. The input layer consists of 3 neurons (operating P, T, and Q_CO2) and the data matrix includes the operating values of these parameters: 20, 30, 35, and 40 MPa (P), 313, 323 and 333 K (T), and 3.23, 4.62 and 5.9 g/min (Q_CO2). The output layer consists of one neuron (R_PUFA).

Figure 4.

ANN model for the prediction of PUFA solubility in SC-CO₂, R_PUFA = f (P, T, Q_CO2)

Figure 4 presents a scheme of the complete architecture of the generated ANN. The results generated after completing the training procedure are illustrated in Figure 5 and Figure 6. Figure 5 depicts the successful completion of the training procedure using the selected Levenberg-Marquardt back propagation training algorithm. Figure 6 demonstrates the ANN training and validation performance; the presented graph suggests that the designed ANN provides the minimal MSE value of 0.17912 for the validation data set. This result confirms that the created ANN and completed training give a highly accurate predictive model to produce a good fit of experimentally observed values R_PUFA in the range of experimental operating conditions and extrapolated data points. The ANN model generated R_PUFA output values that fit the experimental data adequately. The results of regression analysis of the ANN model are presented in Figure 7 and Figure 8. From these results for the employed simulations of the ANN model for the prediction of R_PUFA, it can be concluded that the created ANN predicts the experimentally obtained data with high precision with a correlation coefficient of r = 0.9905 for the entire data set.

Figure 5.

Training procedure for the ANN model R_PUFA = f (P, T, Q_CO2)

Figure 6.

ANN training performance - deviation of model output values from the experimental data for PUFA solubility for each data matrix set

Figure 7.

Correlation of model and experimental data values − validation of the ANN model for the prediction of R_PUFA = f (P, T, Q_CO2)

Figure 8.

Comparative plot − model and experimental data values for the entire data set

The mathematical interpretation of the created ANN is:

$=a_0+a_1\text{×}T+a_2\text{×}P+a_3\text{×}Q+a_4\text{×}{T}^2+a_5\text{×}T\text{×}P+a_6\text{×}T\text{×}Q+a_7\text{×}{P}^2+a_8\text{×}P\text{×}Q+a_9\text{×}{Q}^2$ (4)

Where the dimensions of variables are: P [MPa], T [K] and Q [g/min]. Table 4 shows the values of the model coefficients that were determined.

The 3D-RSM was designed for a detailed analysis of the influence of the experimental variables and resulting variable interactions on the predicted R_PUFA in SC-CO₂ extraction of ω- 3 and ω-6 from freeze-dried viscera tissues, and to initiate process optimisation. The analysis of the defined mathematical model involves interpreting the values of resulting regression coefficients (Table 4). The study suggests that the operating pressure positively impacts the output − R_PUFA. On the other hand, the operating temperature has a more complex impact on the extraction process and mass transfer phenomena occurring in the extractor. Initially, the operating temperature has a negative effect due to the decreased viscosity of the supercritical fluid. Further temperature increase positively impacts R_PUFA due to increased vapour pressure of extracted compounds. The complex impact of model interactions can be observed and elaborated through the analysis of the 3D surfaces generated by the software package Statgraphics Centurion, shown in Figure 9 to 11.

Figure 9.

3D response surface R_PUFA = f (P,T); Q_CO2=3.23 g/min

Figure 10.

3D response surface R_PUFA = f (T, Q_CO2); P=30 MPa

Figure 11.

3D response surface R_PUFA = f (P, Q_CO2); T=323 K

The presented response surface in Figure 9 suggests the highest values for R_PUFA in the operating areas of the most elevated P and T values of 313 and 333 K at constant values of Q_CO2=3.23 g/min. The interpretation of these outcomes originates from the complex interaction between the operating P and T, influencing the physical properties of the extracted components and the supercritical fluid. The function maximum, located in these regions, is the optimal operating region for this system.

Operating values of the experimental variables (P and T) exhibit a significant impact on the observed physical characteristics of the designed extraction setup supercritical carbon dioxide – freeze-dried viscera and, consecutively, on the solvability of isolated ω-3 and ω-6 in SC-CO₂. These process variables affect the vapour pressure of the isolated material and the physical behaviour of the introduced supercritical fluid.

The temperature had a complex influence on the solubility of PUFA in supercritical CO₂ at constant pressure (P=const). In fact, at lower constant P, the increase in the operating temperature decreases the solubility of PUFA in supercritical CO₂. At the higher constant operation P (from 35 to 40 MPa), the result of the increase in the operating T is an increase in the solubility.

The 3D response surface in Figure 10 indicates a function maximum or highest R_PUFA value in the operating region of high values for temperature and CO₂ flow rate, as well as in the area of low values for T and CO₂ flow rate, at a constant pressure of 30 MPa. The analysis of the 3D response surface presented in Figure 11 shows the highest values for PUFA solubility were obtained, as expected, in the operating region of the highest values of the working P and Q_CO2 at a constant operating T of 323 K.