Artificial Intelligence Techniques and Response Surface Methodology for the Optimization of Methyl Ester Sulfonate Synthesis from Used Cooking Oil by Sulfonation

Abstract

Herein, the impacts of sulfonation temperature (100–120 °C), sulfonation time (3–5 h), and NaHSO₃/methyl ester (ME) molar ratio (1:1–1.5:1 mol/mol) on methyl ester sulfonate (MES) yield were studied. For the first time, MES synthesis via the sulfonation process was modeled using the adaptive neuro-fuzzy inference system (ANFIS), artificial neural network (ANN), and response surface methodology (RSM). Moreover, particle swarm optimization (PSO) and RSM methods were used to improve the independent process variables that affect the sulfonation process. The RSM model (coefficient of determination (R²) = 0.9695, mean square error (MSE) = 2.7094, and average absolute deviation (AAD) = 2.9508%) was the least efficient in accurately predicting MES yield, whereas the ANFIS model (R² = 0.9886, MSE = 1.0138, and AAD = 0.9058%) was superior to the ANN model (R² = 0.9750, MSE = 2.6282, and AAD = 1.7184%). The results of process optimization using the developed models revealed that PSO outperformed RSM. The ANFIS model coupled with PSO (ANFIS-PSO) achieved the best combination of sulfonation process factors (96.84 °C temperature, 2.68 h time, and 0.92:1 mol/mol NaHSO₃/ME molar ratio) that resulted in the maximum MES yield of 74.82%. Analysis of MES synthesized under optimum conditions using FTIR, ¹H NMR, and surface tension determination showed that MES could be prepared from used cooking oil.

1. Introduction

The search for biodegradable and renewable feedstock for commercial production has been a research focus in recent time due to the depleting nature of crude oil and the necessity to safeguard the environment from the toxicity of petrochemical products.^1,2 Vegetable oil, algal oil, animal fat, and used cooking oil (UCO) are renewable and biodegradable feedstocks that can replace finite petroleum-based chemicals. UCO is a low-quality raw material whose use could reduce the cost of producing biodiesel (an intermediate product of methyl ester sulfonate). Besides, the processing of UCO aids in the effective conversion of biomass to useful products.^3,4 Currently, large volumes of UCO are produced around the world and pose a waste disposal issue.^5,6 As a result, there is a need to develop an innovative technology for collecting UCO from various locations and using it as a feedstock in the commercial production of fatty acid methyl ester (FAME) for surfactant (methyl ester sulfonate, MES) synthesis.^1,7

Surfactants are derived from either petrochemicals or oleochemicals.¹ Surfactants derived from bio-oils have been found to be nontoxic, less viscous, biodegradable, soluble in water, and slightly irritant to humans^4,8 when compared to petroleum-based surfactants such as cetyltrimethylammoniumbromide (CTAB), internal olefin sulfonate (IOS), and sodium dodecyl sulfate (SDS), which may soon be phased out owing to the depletion of crude oil reserves.^7,9 One of the kinds of oleochemical-based surfactants is fatty acid methyl ester sulfonate (MES).

MES is an anionic surfactant made by either direct sulfonation of FAME with a sulfonating agent such as SO₃, chlorosulfonic acid, oleum, or NaHSO₃ or neutralization of fatty acid methyl ester sulfonic acid (MESA) with sodium hydroxide.^4,9 Although the air-SO₃ falling film sulfonation process is efficient for MES production, it can only be done on a large scale continuously. However, because the former can be run batch-wise or continuously, the NaHSO₃-based sulfonation process has been proposed as a viable alternative to the air-SO₃ falling film sulfonation method. Furthermore, unlike air-SO₃ falling film sulfonation, the NaHSO₃-methyl ester sulfonation process does not release heat.^10,11 Wibowo et al.¹² reported MES synthesis via sulfonation of palm oil methyl esters with NaHSO₃, with an MES yield of 93.2%. Using the NaHSO₃ sulfonation process, a novel castor oil fatty acid methyl ester ethoxylate sulfonate was synthesized.¹³ When UCO methyl ester was sulfonated with NaHSO₃, a sulfonated product with improved properties was formed.⁷

Numerous studies have explored the use of conventional methods to assess the factors influencing the sulfonation process. Using this approach, experimental runs were performed by methodically changing the examined variable while maintaining the other variables constant. The performance ability of response surface methodology (RSM) in the sulfonation process has previously been reported.⁷ The use of conventional methods is characterized by several uncertainties. For example, although RSM can be used to lower experimental runs required to examine the impacts of various input factors and their combined influence on the response, it is unable to capture a chemical or biological process’ nonlinear behavior. Moreover, all of the contributing factors must be tested again, which result in a doubtful number of experimental runs.

More recently, robust statistical modeling tools of artificial intelligence techniques such as the adaptive neuro-fuzzy inference system (ANFIS) and artificial neural network (ANN) have been exploited for modeling processes because they can be used to approximate the nonlinearity associated to a biochemical process system. ANN imitates the brain process mechanism and hence consists of a group of neurons that are linked together in multiple layers. These multiple layers form the basis of the ANN and are referred to as multilayered perceptron (MLP) comprising three layers, viz., input, hidden, and output layers. A benefit of ANN is that it may estimate a variety of nonlinear functions without the need for a specific fitting function to be specified.¹⁴ ANN has been employed in various chemical processes, such as enzymatic-catalyzed reactions,^15,16 esterification and transesterification reaction for biodiesel synthesis,¹⁷⁻²⁰ polymerization reactions,^21,22 and the photocatalytic process.²³ ANFIS is a technique that combines both fuzzy systems and neural networks in a single framework. This offers the ANFIS an advantage such as the ability to demonstrate ambiguity, learning steps, as well as the computational power of neural networks.²⁴ It has also been used in various chemical processes.²⁵⁻²⁸ Advantages of ANN and ANFIS compared to RSM have been extensively reported in many studies.^{18,19,23,29,30}

To increase the effectiveness of the process, it is essential to optimize the process parameters.³¹ Due to the local optimization method by RSM that is only capable of searching local optima, a global optimization method such as particle swarm optimization (PSO) is needed that would locate the global optimum of a given function. PSO is a renowned metaheuristic population-based approach. It is a stochastic optimization technique that is based on the swarm behavior such as flocking of birds and schooling fish. Whereas RSM has its own inbuilt optimization algorithm, ANN and ANFIS developed models need to be coupled with PSO to estimate the global optima of a process. It is noteworthy to state that there has not been a report in the open literature in the optimization of the sulfonation process using the trio of RSM-PSO, ANN-PSO, and ANFIS-PSO.

Thus, methyl ester sulfonate synthesis via sulfonation of methyl ester (ME) with NaHSO₃ was modeled using the trio of ANFIS, ANN, and RSM optimization techniques. The influence of temperature, time, and molar ratio of NaHSO₃ to ME on the sulfonation process was investigated. The effectiveness of the methods was determined statistically by employing the average absolute deviation (AAD), correlation coefficient (R), coefficient of determination (R²), adjusted R², and mean square error (MSE). Furthermore, ME and MES produced under optimum conditions were characterized using FTIR, ¹H NMR, GC-FID, and surface tension determination analyses.

2. Methodology

2.1. Materials

UCO utilized for the experiment was obtained from an eatery in Dehradun, India. The procedure used in pretreating the oil sample and its physicochemical properties have been previously reported.⁷ All the chemical compounds (KOH, CH₃OH, NaHSO₃, Al₂O₃, and NaOH) herein were provided by Thermo-Fisher Scientific Industries, Mumbai, India.

2.2. Preparation of ME and MES

2.2.1. ME Synthesis

The methanolysis process was used to convert UCO to methyl esters in a round-bottom flask. The process was carried out at a temperature of 65 °C for a duration of 1 h. The molar ratio of UCO to methanol was 1:6, and KOH was used as catalyst with a concentration of 1.0 wt %. After completing the reaction process, the resulting product was separated into two layers (upper and lower layers) via a separating funnel. Subsequently, the upper layer (a mixture of biodiesel and unreacted methanol) was evaporated in a rotary vacuum evaporator to remove methanol. Thereafter, the produced biodiesel was washed severally with warm water to remove dissolved KOH and finally dried to remove moisture.

2.2.2. MES Synthesis

MES was synthesized using the method described by Wibowo et al.,¹² that is, sulfonation of UCO methyl esters with NaHSO₃ in a two-neck round bottom flask with a magnetic stirrer. ME, NaHSO₃, and alumina (as catalyst) were fed to the reaction vessel and stirred at 400 rpm. The operating parameters (molar ratio of NaHSO₃ to ME, time, and temperature) were adjusted to the desired operating values (see Table 1). When the sulfonation reaction was completed, a centrifuge was employed to remove the residual NaHSO₃ at a rotating speed of 7500 rpm for 20 min. Afterward, the methyl ester sulfonic acid (MESA) obtained from the supernatant was purified with methanol for 1.5 h at 55 °C. Finally, the purified product was neutralized by adding the 20% NaOH solution dropwise while stirring until a pH of about 8.0 was achieved. After methanol recovery with a rotary evaporator, a sticky pale yellow liquid product (MES) was kept in a covered container for quality analysis. The MES yield (Y₁) was calculated as the percentage of the mass of MES obtained (M_MES) to the mass of ME used (M_ME).

1

Table 1. Experimental Ranges and Levels of the Operational Parameters.

		level
parameter	description	–α	–1	0	+1	+α
T	sulfonation temperature (°C)	96.84	100	110	120	123.16
t	sulfonation time (h)	2.68	3.0	4.0	5.0	5.32
M	NaHSO3/ME molar ratio (mol/mol)	0.92:1	1:1	1.25:1	1.5:1	1.58:1

2.3. Model Development

2.3.1. Sulfonation Process Optimization by RSM

The central composite design (CCD) of RSM was utilized to generate the data set for conducting the experiment. To examine the influence of operational parameters on the sulfonation process output (MES yield), three numeric parameters (sulfonation temperature, sulfonation time, and NaHSO₃/ME molar ratio) were selected. Fifteen experimental runs were suggested by the CCD as indicated in Table 1, which were replicated twice, and the average result of the decolorization efficiency was reported.

The response and operational parameters were correlated using a second-order polynomial response equation (eq 2) with all model terms.

2

where Y stands for the process output (MES yield), β_o symbolizes the constant coefficient; β₁, β₂, and β₃ are the linear terms; β₁₂, β₁₃, and β₂₃ are the coefficients of interaction terms; β₁₁, β₂₂, and β₃₃ denote the coefficients of quadratic terms; and T, t, and M are the coded values of the sulfonation process variables.

2.3.2. Development of the Sulfonation Process’ ANN Model

Levenberg–Marquardt’s algorithm (LMA) was used to create a feedforward, multilayer ANN because of its ability for quick convergence and function modeling. Pure-linear (purelin) and hyperbolic tangent sigmoid (tansig) transfer functions were used for the input–output–hidden layers, respectively. The chosen ANN had an input layer consisting of three neurons (sulfonation temperature, sulfonation time, and NaHSO₃/methyl ester (ME) molar ratio) and a hidden–output layer with one neuron (MES yield). Iteratively testing different numbers of neurons (2–20) until the mean square error (MSE) value of the target data was minimal and high R, which is nearly equal to unity, was attained was used to obtain the optimal hidden neuron number. The tangent sigmoid function is described by eq 3, whereas the purelin transfer is given by eq 4.

3

4

Every input and output data set was scaled back to a value between −1 and +1. Because the tangent sigmoid transfer function varies from −1 to +1, normalization is required. In addition, the normalization ensures that overflow that may result to very large or very small weight anomalies is avoided.³² The normalization was performed by using eq 5.

5

where X^A, X^min, and X^max signify the actual, minimum, and maximum, respectively.

The data set points of input and output (15 in total) were grouped into three subsets, viz., training (60%), validation (20%), and testing (20%).³³ This is crucial to determine the model’s capacity for predicting hidden data that were not used for training and to evaluate the ANN’s capacity for generalization.³⁴

2.3.3. Development of the Sulfonation Process’ ANFIS Model

To prognosticate the MES yield of the sulfonation process, the multiple-input–single-output (MISO) fuzzy model was used to implement the ANFIS model. Three inputs (sulfonation temperature, sulfonation time, and NaHSO₃/ME molar ratio) and one output (MES yield) were employed to develop the ANFIS model. Three generalized bell-shaped membership functions (gbellmf) of first-order Sugeno fuzzy logic were employed for each of the input factor. The output prediction of the ANFIS model was done using hybrid learning algorithms integrated with the defuzzifier formular. The framework was built employing normalization, defuzzification and fuzzification, overall summation, and product.³⁵ The Takagi–Sugeno IF–THEN rules with regard to the input variables can be defined using rules 1 and 2, assuming a two input variable fuzzy inference system (FIS) (u and v) and an output (w).³⁵ The following expression is given for rules 1 and 2:

Rule 1: IF u is A₁ and v is B₁,

Rule 2: IF u is A₂ and v is B₂,

where the fuzzy sets are A₁, A₂, B₁, and B₂ and the outputs of the system are u₁ are u₂. The controllable parameters of the FIS are g₁, g₂, h₁, h₂, k₁, and k₂.

2.3.3.1. First Layer

This layer contains adaptive nodes with three input parameters. Each node n is defined by the following function:

6

where the input parameter to node n is designated as u and ψ_n¹ (symbolizes MF) is the fuzzy set. A_n is the membership class that implies when the provided input n satisfies A.

The expression of gbellmf is given as

7

where gbellmf premise parameters are g_p, h_p, and k_p. The width of the curve is modified by g and h (both must ≥0), and k is the curve’s midpoint. The MF values vary between 0 and 1.

2.3.3.2. Second Layer

This region contains nonadaptive nodes. Product of the incoming signals is performed in this layer to subject all the weight (μ) to scrutiny. Each output node demonstrates the firing strength of the weight.

8

2.3.3.3. Third Layer

Each node runs the required fuzzy rules, and this layer computes each level activation rule. This layer is evaluated by dividing the firing strength of each rule by the aggregate number of rules. This layer’s node is not adaptable.

9

2.3.3.4. Fourth Layer

Defuzzification is used in this layer to calculate the output of the membership function. This layer’s nodes are adaptable.

10

where [g_p, h_p, k_p] refers to a consequent parameter set.

2.3.3.5. Fifth Layer

This has the total of the individual node’s outputs from the defuzzification layer. A single node that represents the output indicates that the layer is not adaptable.

11

where μ̅_n. Q_n denotes the output of node n in the defuzzification layer. Table 2 summarizes the ANFIS features that were employed in this study. The modeling of the ANFIS was carried out in MATLAB 2018a using the fuzzy logic toolbox.

Table 2. Features of the Developed ANN and ANFIS Model.

model	property	value/remark
ANN	training function	Levenberg–Marquardt backpropagation
	performance function	MSE
	learning	supervised
	input layer transfer function	no transfer function is used
	output layer transfer function	purelin
	hidden layer transfer function	hyperbolic tangent sigmoid (tansig)
	number of training iterations	120
	number of best iterations	70
	number of input neurons	3a
	number of hidden neurons	10
	number of output neurons	1b
ANFIS	fuzzy type	Sugeno
	input	3a
	output	1b
	membership function	generalized bell-shaped
	input/ouput membership function	3
	and method	product
	or method	probabilistic
	implication method	product
	aggregation	sum
	number of rules	27
	number of of linear parameters	54
	number of nonlinear parameters	27
	output membership function type	linear

^aSulfonation temperature, sulfonation time, and NaHSO₃/ME molar ratio.

^bMES yield.

2.4. Operational Parameters’ Sensitivity Study

Conducting a sensitivity study will ensure how well the model behaves. It is used to examine the impact of each input parameter on the model response (output).³⁶ The sum of squares for each input term and the total of squares for all input parameters were used to conduct the sensitivity analysis for the RSM model.³⁷ These sums of squares were utilized to compute the percentage contribution of each input parameter on the response (MES yield) following eq 12.

12

where SOS_i and SOS_o, respectively, represent the sum of squares for a particular input factor and the total sum of squares for all the input factors.

Sensitivity analysis of the input factors for the implemented ANN model was carried out employing Garson’s method following eq 13.³⁸ The weights computed for both the input parameters and the response (MES yield) are displayed in Table 3. The calculated weights were employed to analyze the relative importance of the operational factors.

13

where the terms n_i and n_h symbolize the input and hidden neurons, respectively. The connecting weight is represented by β. The layers consisting of the input, hidden are designated by i, h, and o superscripts, respectively, whereas neurons embedded in the input, hidden, and output are symbolized by the k, a, and b subscripts. Ω_t is the t_th input parameter’s influence on the output parameter’s relative importance.

Table 3. Weights of ANN Model Employed for Analysis Results.

	input weights			output weight
neuron	sulfonation temperature	sulfonation time	NaHSO3/ME	MES yield
1	2.3174	1.6698	–1.0946	0.6126
2	2.2292	1.9752	–1.3210	1.1358
3	–2.3596	1.1338	–0.7228	0.3221
4	–2.3942	–1.1341	1.2867	–0.0173
5	–3.6142	1.1097	0.1621	1.8098
6	–1.4673	2.9082	0.9966	–1.1542
7	1.2220	–1.7860	1.7769	0.3842
8	2.5357	–0.8107	1.8620	0.9540
9	1.1485	1.7165	–1.9031	–0.2252
10	–1.4492	–1.3038	1.6817	–0.0526

The ANFIS model’s input parameters were analyzed for sensitivity by determining the maximal value of the selected input parameters while retaining the remaining input parameters at their nominal levels (the most frequent values). The sensitivity study was carried out to validate the effect of input parameters on the model response (MES yield) of the implemented ANFIS model. To depict the response of the ANFIS model, one variable was changed, whereas other variables remained constant.³⁹Table 4 displays the nominal, minimum, and maximal values that were employed to perform the sensitivity analysis by the ANFIS model.

Table 4. ANFIS Sensitivity Analysis Parameters.

input variable	minimum	nominal	maximum
temperature	96.84	110	123.16
time	2.68	4	5.32
NaHSO3/ME molar ratio	0.92	1.25	1.58

2.5. Process Parameter Optimization

The optimization procedures of RSM and PSO were engaged to determine the optimal blends of process variables to attain the highest MES yield to improve the sulfonation process. The implemented ANN, ANFIS, and RSM models were used as the optimization algorithm’s fitness function. Table 5 illustrates parameters of PSO employed for the variables’ optimization. The optimal values prognosticated by RSM and PSO were validated in the lab by completing the experiment in triplicate separately. By averaging the values, the observed values and predicted values were compared. PSO codes were developed using MATLAB R2018a, while optimization by RSM was executed using version 7.0.0 of the Design Expert software.

Table 5. PSO Features for RSM, ANN, and ANFIS Developed Models.

property	value/comment
swarm size	10–15
initial range	[0.1, 0.1]
self-adjustment	2
social adjustment	2
iteration	10–50

2.6. Appraisal of the Developed ANN, ANFIS, and RSM Models

Statistical measures including the correlation coefficient (R), coefficient of determination (R²), adjusted R², mean square error (MSE), and average absolute deviation (AAD) were engaged to evaluate the developed models’ predictive efficacy. The statistical indicators were determined using eqs 14–18.^33,40

14

15

16

17

18

where v is the number of operational parameters, n is the experimental number points, x_i, pr is the predicted value, x_i, ob is the experimental value, and x_ob, av is the mean of the observed value.

2.7. Analysis and Characterization of ME and MES

Transesterification and sulfonation processes’ products were analyzed using different characterization techniques. A gas chromatography-flame ionization detector (Agilent 7890A, USA) and capillary column (size: 0.32 mm × 15 m, 0.10 mm thickness) were used in determining the FAME profile of the methanolysis product, as discussed in our previously reported study.⁷ Furthermore, functional groups present in both methanolysis and sulfonation products were evaluated using Fourier transform infrared (FTIR) spectrophotometer (Perkin Elmer, USA), whereas the chemical compositions of the two products were analyzed with the aid of a nuclear magnetic resonance spectrometer (¹H NMR, Bruker Avance III-HD 500 MHz) with CDCl₃ used as solvent. A surface tensiometer (KRUSS Scientific, USA) was used to estimate the surface tension of the MES solution at different concentrations.

3. Results and Discussion

3.1. Analysis of FAME Profile of Biodiesel Produced from UCO

The GC-FID was used in analyzing the UCO methyl ester composition, and the analysis results are shown in Figure 1 and Table 6. As indicated in the results, the major methyl esters detected in the UCO biodiesel were palmitic, oleic, linoleic, linolenic, and stearic acids. The methyl linoleate accounted for 47.2% of the whole intermediate product, making it a dominant component in the synthesized UCO FAME. Furthermore, the GC-FID results revealed that the UCO biodiesel contained 19.10% saturated and 77.6% unsaturated methyl ester contents, indicating that the intermediate product satisfied the minimum requirement of the ASTM and EN 14214 standard.⁴¹

Figure 1.

Chromatogram of UCO methyl ester.

Table 6. Composition of the FAME Intermediate Derived from UCO.

FAME profile	chemical formula	retention time (min)	composition (wt %)	class
methyl myristate	C15H30O2	11.308	0.7	unsaturated
methyl palmitate	C17H34O2	12.672	13.7	saturated
methyl palmtoleate	C17H32O2	13.589	0.20	unsaturated
methyl stearate	C19H38O2	14.662	5.4	saturated
methyl oleate	C19H36O2	14.839	23.7	unsaturated
methyl linoleate	C19H34O2	15.319	47.2	unsaturated
methyl linolenate	C19H32O2	16.000	5.8	unsaturated
total saturated (%)			19.1
total unsaturated (%)			77.6
total FAME (%)			96.7
other			3.3

3.2. Sulfonation Process Modeling via RSM

The CCD matrix and the estimated value of MES yield by the RSM model are shown in Table 7. The RSM model predicted a range of MES yields between 37.14 and 68.63%. The polynomial regression model obtained for the process using Design Expert version 7.0.0 is described by eq 19.

19

where T is sulfonation temperature (°C), t is the sulfonation time (h), and M is the NaHSO₃/ME molar ratio (min).

Table 7. The Three-Factor CCD Matrix and the Value of Prediction by the Developed Models.

run	sulfonation temperature, T (°C)	sulfonation time, t (h)	NaHSO3/ME molar ratio, M (mol/mol)	MES yield (wt %)	RSM prediction (wt %)	ANN prediction (wt %)	ANFIS prediction (wt %)
1	123.16	4	1.25	58.1	59.68	62.71	58.10
2	96.84	4	1.25	38.5	40.08	38.50	38.50
3	100	3	1	70	68.63	69.98	70.00
4	120	3	1.5	38.51	37.14	38.51	38.51
5	110	2.68	1.25	49.8	51.37	49.80	49.80
6	100	5	1.5	57.2	55.83	57.21	57.20
7	110	4	1.58	41.5	43.07	41.50	41.50
8	110	5.32	1.25	65	66.57	65.02	65.00
9	110	4	0.92	42.8	44.37	41.78	42.80
10	120	5	1	60	58.63	61.38	60.00
11	110	4	1.25	45.7	45.05	45.72	45.84
12	110	4	1.25	45	45.05	45.72	45.84
13	110	4	1.25	45.8	45.05	45.72	45.84
14	110	4	1.25	49	45.05	45.72	45.84
15	110	4	1.25	43.72	45.05	45.72	45.84

Table 8 displays the ANOVA and statistical significance test outcomes for the model. The Fisher test (F value) and p value at a 95% degree of confidence were employed to determine statistical significance. The model’s p value and F value are 0.0027 and 17.87, respectively, indicating that it is significant overall. All the model terms are significant except for the NaHSO₃/ME molar ratio (M), interaction between sulfonation temperature and NaHSO₃/ME (TM), quadratic of sulfonation temperature (T²), and quadratic of NaHSO₃/ME molar ratio (M²). The model’s high R² (0.9698) and adjusted R² (0.9685) values confirmed its validity. The Pareto chart was employed to estimate each model term’s significance and their interactions (Figure 2). The significance of each model term on the chart depends on the length of the bar.⁴² Hence, from the Pareto chart, the insignificant model term was the quadratic term of the sulfonation temperature followed by the interaction between sulfonation temperature and NaHSO₃/ME molar ratio term and the quadratic and linear terms of the NaHSO₃/ME molar ratio term, whereas other model terms are significant. This corroborates the ANOVA evaluations depicted in Table 8. The actual and predicted values have low variance, as indicated by the low coefficient of variance (CV), which is 5.66%. The signal/noise ratio is gauged by adequate precision; a value >4 is desirable.⁴³ The value of 13.606 computed in this work (Table 9) implies an adequate signal.

Table 8. Test of Significance for Every Regression Coefficient and ANOVA^a.

source	SS	df	MS	F value	p value
model	1292.22	9	143.58	17.8686	0.0027	significant
T	192.08	1	192.08	23.9045	0.0045
t	115.52	1	115.52	14.3766	0.0127
M	0.845	1	0.845	0.10516	0.7588
Tt	121.156	1	121.156	15.078	0.0116
TM	24.0891	1	24.0891	2.99791	0.1439
tM	396.736	1	396.736	49.3742	0.0009
T2	48.1694	1	48.1694	5.99473	0.0580
t2	403.473	1	403.473	50.2126	0.0009
M2	3.77958	1	3.77958	0.47037	0.5233
ANOVA
R2	0.96985
adjusted R2	0.91557
residual	40.1765	5	8.0353
lack of fit	24.9698	1	24.9698	6.56809	0.0625	not significant
pure error	15.2067	4	3.80168
total SS	1332.39	14
adeq precision	13.6056
CV (%)	5.66

^aSS, sum of squares; MS, mean square; df, degree of freedom; CV, coefficient of variation.

Figure 2.

Pareto chart of standardized effects for the RSM model. T, sulfonation temperature; t, sulfonation time; M, NaHSO₃/ME molar ratio; L, linear; and Q, quadratic.

Table 9. Statistical Performance of the Developed Models.

indicator	RSM	ANN	ANFIS
R	0.9846	0.9874	0.9943
R2	0.9695	0.9750	0.9886
adj R2	0.9584	0.9659	0.9844
MSE	2.7094	2.6282	1.0138
AAD (%)	2.9508	1.7184	0.9058

3.2.1. Variables’ Interactive Effect on Sulfonation Process Response

Figure 3 depicts the interactive effect of temperature, time, and NaHSO₃/ME molar ratio on the sulfonation process. There was a corresponding increase in MES yield with increasing sulfonation temperature and time, as illustrated in Figure 3a. Temperature was thought to determine the extent of heat transfer by the catalyst (Al₂O₃) to improve the diffusion of the NaHSO₃ (sulfonating agent) into the liquid phase (methyl ester), thus increasing the reaction rate. This finding was consistent with previous research.⁹

Figure 3.

Surface plots on the impact of (a) sulfonation temperature (°C) and sulfonation time (h), (b) NaHSO₃/ME molar ratio and sulfonation time (h), and (c) NaHSO₃/ME molar ratio and sulfonation temperature (°C) on MES yield (%).

The ME was sulfonated with NaHSO₃ at a fixed temperature of 110 °C to evaluate the combined effect of NaHSO₃/ME molar ratio and sulfonation time on MES yield. The MES yield decreased with increasing molar ratio of NaHSO₃ to ME and sulfonation time (Figure 3b). This result indicated that the highest MES yield was obtained at a low molar ratio of sulfonating agent to intermediate product, implying that there was no reaction of MES with unconsumed NaHSO₃ in the liquid phase.⁹ However, the minimum MES yield obtained at a higher molar ratio of NaHSO₃ to ME suggested that a high concentration of unreacted sulfonating agent in the reaction medium drove the reaction toward the formation of more intermediates.⁴⁴ This observation was corroborated by Figure 3c (effect of NaHSO₃/ME molar ratio and temperature on MES yield), where the maximum MES yield was achieved at low NaHSO₃/ME molar ratio, confirming that sulfonation temperature was significant to MES yield (sulfonation process output variable).

3.3. Sulfonation Process Modeling via ANN

The experimental data set produced by the CCD was used to execute the neural network model. Fifteen experimental data points were adequately employed in total, of which 60% were engaged to train the network model, 20% to test it, and the remaining 35% to validate the model. Table 5 shows the predicted results by the developed ANN model. To compute the weights and biases for this network, the Levenberg–Marquardt (LM) back-propagation procedure was engaged during the training. The neural network was trained heuristically with different hidden neurons (2 to 20). The optimal hidden neuron selected was 10 because it gave the MSE and highest R as shown in Figure 4. In this study, the sulfonation process’ network structure is shown in Figure 5, where 3 represents the input variables, 10 represents hidden neurons, and 1 represents the output (MES yield). The results of the regression for validation, training, testing, and overall are displayed in Figure 6. The results revealed that the observed and predicted values had a good degree of agreement.

Figure 4.

Optimal hidden neuron number.

Figure 5.

ANN topology with input, hidden (tansig transfer function), and output layers (pure linear transfer function).

Figure 6.

Regression plots for training, testing, validation, and overall data set for the developed ANN model.

3.4. Sulfonation Process Modeling via ANFIS

Figure 7 presents the framework of the ANFIS model. About 27 (3 × 3 × 3) rules consisting of three linguistic terms, viz., low, medium, and high, were used for the ANFIS model (Figure 8). Table 9 displays the ANFIS model’s predicted MES yields for various experimental conditions. Figure 9 depicts graphs of experimental and prognosticated values vs run numbers, illustrating the model’s accuracy. The R, R², and adjusted R² values for the ANFIS model were 0.9943, 0.9886, and 0.9844, respectively, confirming a strong correlation between the experimental and predicted values. A good fit model is also defined by a high R² value.⁴⁵ This suggests that the developed model can account for 98.8% of the variation between experimental and prognosticated values.⁴⁶

Figure 7.

Developed ANFIS model framework.

Figure 8.

Developed ANFIS model rule viewer.

Figure 9.

Experimental and predicted values vs run numbers for the ANFIS model.

3.5. Developed Models’ Performance Evaluation

By computing the MSE R, AAD, R², and adjusted R², it was statistically determined that the generated models were capable of predicting the MES yield. Table 9 displays the outcomes calculated from these statistical measures for the three models. Figure 10 illustrates the regression plots of predicted and observed values, which agree with the high values of R obtained for the three models (Table 9). According to some reports, the correlation between predicted and observed values should be at least 0.8.⁴⁵ Additionally, the three models’ R² values were high, indicating strong model fitness.⁴⁵ The adjusted R² was utilized to verify R² overestimation, and all of the models’ estimated values were high, demonstrating their importance. The error terms (MSE and AAD) computed for all the developed models have low values, demonstrating that all the models have good precision and accuracy. The ANFIS model was superior to RSM and ANN models, as indicated in Table 9, whereas the RSM model has the lowest precision and accuracy for predicting MES yield.

Figure 10.

Experimental vs predicted values’ parity plots for the developed models.

3.6. Optimization of the Input Process Variables for the Sulfonation Process

The process input variables (temperature of sulfonation, sulfonation time, and molar ratio of NaHSO₃/ME) were optimized using RSM, RSM-PSO, ANN-PSO, and ANFIS-PSO. Employing the developed models as objective functions, PSO was applied to obtain the best combination of the investigated input process variable for the highest MES yield. The established optimum conditions for each of the techniques are presented in Table 10. Figure 11 displays the optimization result for RSM-PSO, ANN-PSO, and ANFIS-PSO. The estimated values were validated in duplicate in the laboratory, and the mean value of the MES yield is reported in Table 10. The order of the optimization is ranked as ANFIS-PSO, ANN-PSO, RSM-PSO, and RSM (see Table 10). ANFIS-PSO gave the highest MES yield (74.8%) under favorable conditions of reaction temperature 96.8 °C, reaction time 2.68 h, and molar ratio of NaHSO₃/ME molar ratio 0.92.

Table 10. Optimization Techniques and Model Validation.

method	reaction temperature (°C)	reaction time (h)	NaHSO3/ME (mol/mol)	predicted MES yield (%)	experimental MES yield (%)
RSM	101.38	3.06	1:1	65.43	67.03
RSM-PSO	96.84	2.68	1.18:1	67.25	68.01
ANN-PSO	99.16	2.68	0.99:1	72.35	73.22
ANFIS-PSO	96.84	2.68	0.92:1	74.82	77.96

Figure 11.

Convergence and optimization result plots for the PSO. (a) RSM model, (b) ANN model, and (c) ANFIS model.

3.7. Sensitivity Analysis of the Input Variables on MES Yield

Figure 12 displays the sensitivity analysis for the ANFIS and ANN models. The outcomes exhibit a similar pattern to that of RSM, where sulfonation temperature is the most important input factor on the response (MES yield) followed by sulfonation time and finally NaHSO₃/ME molar ratio. It was observed that the levels of importance varied for the different modeling techniques (see Figure 12). Although the level of importance for NaHSO₃/ME molar ratio for the RSM model was very low compared to ANN and ANFIS, none of the process input factors, however, could be disregarded.

Figure 12.

Level of importance of process input variables on MES yield.

3.8. Analysis of MES Produced under Optimum Conditions

3.8.1. FTIR Analysis

Figure 13 presents the FTIR analysis conducted on both ME and MES to determine the functional groups. The FTIR spectra of ME and MES revealed some peaks around 2939–2911 cm^–1, indicating −CH₃ (methyl) stretching vibration.^4,13 Also, the peak at 1745 cm^–1 (C=O stretching) appeared in the spectra of both samples, suggesting that the sulfonation process did not affect the ester group, as reported for ME and MES obtained from sesame oil.¹ The new peaks at 1351, 1168, and 1089 cm^–1 on the spectrum of the UCO MES were all attributed to S=O vibration reductions, indicating that sulfonic acid (−SO₃H) was successfully incorporated into the MES structure in the form of C-SO₃H.⁴ These FTIR data confirmed that the sulfonation product was methyl ester sulfonate.

Figure 13.

FTIR spectra of ME and MES samples.

3.8.2. ¹H NMR Analysis

The ¹H NMR results of the ME and MES samples are shown in Figure 14. The signals at around 0.88–0.97 ppm in the ME spectrum were attributed to the methyl (CH₃−) proton of the fatty acid chains, whereas the signals at around 1.41–2.10 ppm were attributed to the methylene (−CH₂−) protons of saturated acyl chains.¹ Peaks at 3.71 and 5.42 ppm, respectively, indicated the presence of protons of CH₃–O– in ester and −CH=CH of the vinyl group.⁴⁷ However, after the conversion of ME to MES via the sulfonation process, some signal disappeared while some new signals were formed, indicating that the target product (MES) was successfully formed. New signals formed at 0.83 ppm (terminal methyl (CH₃−) proton), 2.38 ppm (allylic (=CH–CH₂–CH=) protons), 3.37 ppm (−OCH₃– linked with the ester group), and 3.68 ppm (proton linked with carbon atom bearing the −SO₃ (sulfonate) group) were detected. Interestingly, the ¹H NMR signals of the produced MES confirmed the presence of methyl, ester, and sulfonate groups.

Figure 14.

¹H NMR spectra of ME and MES.

3.8.3. Surface Tension Analysis of MES

Figure 15 displays the plot of surface tension against MES concentration. The result revealed a decrease in surface tension with increasing MES concentration. However, the values of MES surface tension stabilized as saturation of solution/air interface with molecules of surfactant set in.⁴ In addition to this, the values of the critical micelle concentration (CMC) and corresponding surface tension were estimated from Figure 15. The point where the descending and horizontal lines intercept was taken as the value of CMC. Regarding the results obtained, the values of the CMC and corresponding surface tension were 118 mg/L (0.118 g/L) and 32.1 mN/m, respectively. Comparing the MES synthesized in the current work and MES obtained from mango kernel oil by Sathi Reddy et al.,⁴⁸ the CMC of the former was slightly greater than the CMC of the latter (80 mg/L). However, the CMC of MES synthesized from castor oil using chlorosulfonic acid (sulfonating agent) was 205 mg/L,⁴⁹ and this was higher than the CMC of MES produced herein. The discrepancy might be due to the different source of intermediate product and sulfonating agent used. Nevertheless, MES with lower CMC often possesses improved hydrophobic surface and reduced surface tension capability as the surfactant molecules can arrange themselves closer to the surface.⁴

Figure 15.

Surface tension–concentration plot for MES.

3.9. Reaction Mechanisms for Methanolysis of Used Cooking Oil and Sulfonation of Methyl Ester

Reaction Scheme 1 shows the mechanism for the transesterification of UCO into fatty acid methyl esters. During methanolysis of UCO in the presence of a catalyst (KOH), the triglycerides (TG) present in UCO convert to diglycerides (DG) followed by the conversion of DG to monoglycerides (MG) and then finally to glycerol (G). Overall, 1 mol TG reacts with 3 mol methanol in the presence of a KOH catalyst to produce 1 mol glycerol and 3 mol methyl esters (ME, desired product), which correspond to the various fatty acid methyl esters detected (see Table 6).

Scheme 1. Mechanism for the Transesterification of UCO into Fatty Acid Methyl Esters.

(a) Transesterification reaction steps. (b) Transesterification of triglyceride with methanol.

Scheme 2 displays the reaction mechanism for the sulfonation of fatty acid methyl esters with NaHSO₃. The hydrosulfite (SO₃H) anion replaces a hydrogen atom attached to the α-carbon of methyl ester, and Na⁺ is removed during sulfonation, resulting in the formation of an intermediate product (MESA), which is purified by methanol and subsequently neutralized by NaOH solution at controlled conditions (pH and temperature) to produce methyl ester sulfonate (RCH(SO₃Na)CO₂CH₃) and water.

Scheme 2. Reaction Mechanism for the Sulfonation of Methyl Ester with NaHSO₃.

4. Conclusions

This work focused on the significance of selecting the right modeling and optimization strategies to convert UCO to MES through the transesterification–sulfonation process. Detailed modeling of the process was executed using ANFIS, ANN, and RSM. Moreover, RSM, RSM-PSO ANN-PSO, and ANFIS-PSO were used in optimizing the three investigated input process variables (reaction temperature, sulfonation time, and NaHSO₃/methyl ester molar ratio) to obtain the highest MES yield. According to statistical measures for evaluating the efficiency of the created models, the RSM model recorded the least efficiency in predicting MES yield (R = 0.9846, MSE = 2.7094, and AAD = 2.9508%). The ANFIS model outperformed the ANN model (R = 0.9874, MSE = 2.6282, and AAD = 1.7184%) in terms of R, MSE, and AAD. The results of the optimization demonstrate that ANFIS-PSO provided the best combination of operation parameters (sulfonation temperature 96.84 °C, sulfonation time 2.68 h, and NaHSO₃/ME molar ratio 0.92:1 (mol/mol)) with the highest MES yield (74.82%). Finally, sensitivity analysis of the input variables on MES yield depicts that the ranking of the level of importance of the process variables is sulfonation temperature > sulfonation time > NaHSO₃/ME molar ratio.

The authors declare no competing financial interest.