Development of nano-colloidal system for fullerene by ultrasonic-assisted emulsification techniques based on artificial neural network

2.1 Materials

Fullerene (C₆₀), Tween® 80 (T80), Span® 80 (S80), white beeswax, and xanthan gum were purchased from Sigma–Aldrich. Phenonip was purchased from Bramble Berry® Inc. Palm kernel oil esters (PKOEs) were synthesized from palm kernel oil and oleyl alcohol via enzymatic transesterification according to the literature (Keng et al., 2009). Deionized water was purified using Milli-Q® water system. All other chemicals were of analytical grade and were used as received without further purification.

2.2 Formulation of fullerene nanoemulsions

Formulations containing fullerene were prepared using ultrasonic dispersion. 12.0 mg of fullerene powders was magnetically stirred in PKOEs to ensure quick and complete dissolution. This formulation constituted of 12.5% (w/w) PKOEs-C₆₀, 1.0% (w/w) beeswax, and 0.7% (w/w) phenonip as the oil (dispersed) phase whereas the aqueous (continuous) phase consisted mainly of deionized water with 7.7% (w/w) T80:S80 (4:1), and 0.9% (w/w) xanthan gum. Both phases were stirred independently under heating condition at 75.0 °C. 100.0 g of fullerene nanoemulsions was prepared by gradual addition of aqueous phase into the oil phase while homogenized using Polytron high shear homogenizer (PT3100, Kinematica AG) at room temperature (25.0 ± 2 °C) for 15 min. Pre-mixed emulsions were then subjected to sound waves generated by ultrasonic processor (UP400S, Hielscher Ultrasonics GmbH) at 24 kHz with maximum power output of 400 W. 14 mm × 100 mm (diameter × length) titanium probe tip was dipped into the samples while sonicated at specific amplitude with on:off cycle per second. Samples were immersed in ice bath to mitigate the ultrasound thermal effect. Sonication amplitude and time were varied according to the experimental design. Homogenization rate was controlled as a secondary technique in the production of nanoemulsion.

2.3 Droplet size and viscosity analysis

A dynamic light scattering droplet size analyzer (Zetasizer Nano ZS90, Malvern) was used to determine the droplet size of fullerene nanoemulsions. Standard measurement was executed with an argon laser (λ = 488 nm) at the angle of 173° in room temperature (25.0 ± 0.5 °C). Samples were diluted with deionized water at the ratio of 1:200 before injecting the samples into the capillary of zeta disposable cells using 3 mL syringe. Dilution was necessary to avoid multiple scattering effects and slow injection to restrain air bubbles from forming inside the cells before inserted into the module. Instrument was equilibrated for 120 s to acquire stable temperature for improved accuracy of reading. Five replicates (n = 5) from each sample were analyzed twice after 24 h of sample preparation and the results were reported as an average for each analysis.

Viscosities of fullerene nanoemulsion measurement were performed by dynamic shear rheometer (Kinexus Rotational Rheometer, Malvern) under constant shear rate of 0.1 s⁻¹ at 25.0 °C. Analysis was undertaken within 24 h after sample preparation with 0.15 mm of gap in between the cone and plate geometry (4°/40 mm). Once the sample was loaded, it was allowed to reach equilibrium for 5 min prior to the measurement. Viscosities of each sample were recorded every 5 s in Pascal second (Pa s) for 5 min and the mean value of viscosities was taken as the representative viscosity of the sample.

2.4 Artificial neural network modeling

Modeling was carried out using NeuralPower® version 2.5 by CPC-X Software (Carnegie, USA). Input variables, include sonication amplitude (30–70%), sonication duration (60–120 s), and homogenization rate (4000–5000 rpm), and would be investigated in response to the droplet size and viscosity of the nanoemulsion system. 26 experiments were categorized into two data sets which consisted of training (18 points) and testing (8 points) sets as shown in Table 1. Through different learning algorithms, input variables are multiplied by connection weights and summed before applying the sigmoid activation (transfer) function to generate the output. Experience gained from the learning phase would be used in data fitting of model. Validation data set (8 points), which was excluded from the learning phase, was utilized to estimate potential model that has been trained (Table 2). Eventually, testing phase was conducted simultaneously to check the fitness of selected model as well as to avoid over-fitting by controlling errors.

In this study, universal learning algorithms applied were QP, BBP, IBP, and LM. Each node in the hidden or output layers acted as a summing junction which transformed the inputs from the previous layer using Eq. (1).

2.5 Levenberg–Marquardt (LM) algorithm

LM is an analogous version of back propagation algorithm which was often applied to solve nonlinear minimum mean square in most modeling problems. While fundamental back propagation is the steepest descent algorithm, LM algorithm offers an alternative conjugation methods for the second derivative optimization (Marquardt, 1963). Due to its fast convergence and stability properties, it reduces the step length to attain the least square curve fitting during minimization process. In this particular algorithm, the update function, f_n, can be computed using Eqs. (5) and (6) as shown below.

3.1 Algorithm topology analysis

Architecture of ANN was designed in a fashion of three controlling parameters in the input layer and two features of developed nanoemulsions as the output layer. Number of nodes ranging from 1 to 15 served as the hidden layer to study the topologies by applying different algorithms independently. During the learning process, testing data set would determine the function of error by calculating the minimum RMSE value. When RMSE values stopped decreasing, or alternatively began to increase, it indicated that the model was over-fitting the data. This marked the end of the learning phase. All data sets were trained for ten times for every runs in each node to eliminate random correlation in stochastic weight initialization (Salama and Abdelbar, 2014). Fig. 1 shows the minimum RMSE values yielded for each node in hidden layer of different algorithms. Topologies with the lowest RMSE were selected as the best topology representing the respective algorithms for comparison.

Close

Figure 1 RMSE versus node number of hidden layer in neural networks for (A) droplet size and (B) viscosity.

Export to PPT

Neural network with very low node numbers in the hidden layer would restrict the learning performance due to the network size. It can be seen in Fig. 1 that 3 nodes and below gave exceptionally high RMSE values. Yet, excessive number of nodes in the hidden layer would cause the learning duration to be lengthy and may be compromised by local minima (Ghoreishi and Heidari, 2013). According to the graph, hidden node numbers that constructed the topologies using LM, IBP, BBP, and QP algorithms with minimum RMSE were 4, 6, 7, and 8, respectively. From Table 3, RMSE values from the testing data sets revealed that LM had the lowest RMSE value among other algorithms with 2.083 and 2.226 for droplet size and viscosity, respectively. Hereinafter, LM derived network would be the representative model in the prediction of determining process conditions for nanoemulsion formulation.

This representative model was specifically chosen from the trained models by different learning algorithms based on the RMSE, AAD, and R² values. In order to test how well the data fit along the regression line, R² values were derived from the scatter plot of predicted against actual values of droplet size and viscosity for both training (Fig. 2) and testing (Fig. 3) data sets. These statistic values are presented in Table 3. LM algorithm with 4 hidden nodes exhibited the lowest RMSE among the rest of algorithms. Besides, high R² values of droplet size and viscosity obtained using LM were 0.9809 and 0.9743, respectively. Despite just having excellent correlation between data, AAD values further supported that LM model was able to describe well the formation of this particular system. It is worth mentioning that LM algorithm did perform better than BBP, IBP, and QP in this case. For these reasons, it was selected as optimum model for the production of nanoemulsion system.

Close

Figure 2 Predicted versus actual values (A) droplet size and (B) viscosity of fullerene nanoemulsions in the learning phase using ANN: QP, BBP, IBP and LM algorithms.

Export to PPT

Close

Figure 3 Predicted versus actual values of (A) droplet size and (B) viscosity of fullerene nanoemulsions in the testing phase using ANN: QP, BBP, IBP and LM algorithms.

Export to PPT

3.2 Validation of selected model

Optimum model was validated by eight different experimental runs which were conducted independently apart from the training and testing data sets. Results are tabulated in Table 2. It was observed that the experimental values had no significant difference as compared to the predicted values in terms of droplet size but not always in the case for viscosity. Residual standard error (RSE) percentage for droplet size was lower than 3% in all cases indicating the high fitness of model. However, a slight deviation in prediction of viscosity which caused the RSE percentage of three out of eight runs to be over 5%. Nonetheless, the selected model (LM-3-4-2) was demonstrating great predictive ability for both the droplet size and viscosity of nanoemulsions.

Neural network of LM-3-4-2 was mapped out to relate the connectivity of different layers via feed-forward control as shown in Fig. 4. Input data of hidden nodes are calculated by weighted summation as shown in Eq. (7).

Close

Figure 4 Architecture of neural network for multilayer normal feed forward connection type according to optimum topology from Levenberg–Marquardt algorithm.

Export to PPT

3.3 Relative impact of processing parameters on nanoemulsions

Effective percentage of processing parameters in formulating nanoemulsions at optimum level is shown in Fig. 5. From the analysis, sonication time has the largest effect on shaping the droplet size of the nanoemulsion system (44.17%) followed by sonication amplitude (39.92%) and homogenization rate (15.91%). Similarly, sonication amplitude has the most substantial effect on viscosity (45.26%) followed by homogenization rate (27.38%) and sonication time (27.36%). It was evident that ultrasonic cavitation dominated the control on the physical attributes of the nanoemulsion system over the homogenization techniques. Nonetheless, both processing methods have significant effect on the response variables which cannot be omitted.

Close

Figure 5 Effective percentage of processing parameters toward droplet size and viscosity of fullerene nanoemulsions.

Export to PPT

3.4 Synergistic effect of ultrasonic parameters on droplet size and viscosity

Fig. 6 shows the 3D plot representing the effect of ultrasonic cavitation toward the droplet size and viscosity of fullerene nanoemulsion. In Fig. 6A, extended sonication time reduced the droplet size in an inverse sigmoidal pattern when exposed to high irradiation power above applied sonication amplitude of 50%. This phenomenon creates multiple cavities by sound pressure followed by immediate collapse which eventually breaks the emulsion droplet into a smaller size (Shahavi et al., 2019). Higher applied amplitude intensified the frequency of this event which leads to consecutive breakdown into nano-droplets. When the liquid surrounding the vibrating probe tip was split into fine droplets, ultrasonic cavitation occurred. A correlation was proposed for the prediction of droplet size, D_p produced by sonication amplitude, Am as described below (Dalmoro et al., 2013; Rajan and Pandit, 2001).

Close

Figure 6 Synergistic effect of ultrasonic parameters on (A) droplet size and (B) viscosity of fullerene nanoemulsions.

Export to PPT

In Fig. 6B, as the sonication amplitude increased toward 70%, the viscosity was reduced steadily at different rates. The effect of sonication amplitude in reducing the viscosity decreases over time but remained significantly effective within the tested range. As the sonication amplitude increases, local heat was generated and dissipated within the system which led to temperature hike. Consequently, this caused the interfacial tension to be disrupted where viscosity was directly affected. Chemical bonds between monomer units were broken down as well which resulted in fragmentation of structural polymer network (Gogate and Prajapat, 2015). Beyond the consideration of sonication amplitude, longer sonication time led to further reduction in viscosity of nanoemulsions due to the longer exposure to cavitation. The viscosity of nanoemulsions was mainly governed by sonication amplitude but ample time of sonication is a prerequisite for this case to achieve the desired viscosity. Likewise, Fig. 5 was further supported by this analysis which indicated that the sonication amplitude has indeed the most prominent influence on regulating the viscosity.

As the amount of each component was maintained at a constant mass, the viscosity increased as the droplet size decreased in the system. Large surface area of smaller droplets would result in higher frequency of interaction between adjacent droplets which leads to resistance in flow. The presence of non-hydrodynamic forces such as electrostatic, steric, van der Waal, and Brownian forces naturally caused the overall viscosity of colloidal system to become higher as droplet size reduces. Since the linear viscoelastic properties in oscillatory shear flow are largely dependent on droplet size, the storage modulus increases with the decrease in droplet size due to an increase in interfacial stress (Pal, 2011). Nevertheless, fullerene nanoemulsion was able to remain stable and homogeneous at 4 and 25 °C over the period of 90 days with the mean droplet size below 200 nm and the textural change in viscosity was almost negligible.