Evaluation of ethanol-gasoline blends in SI engines using experimental and ANN techniques

The growing demand for energy has led to a significant reliance on crude oil to meet our increasing energy needs. However, fossil fuels like crude oil result in the emission of harmful gases and sulfur contents, whose effects on the environment are detrimental. There are many trends towards the use of alternative fuels to preserve the environment from pollution. As biofuels contain less carbon, which makes them produce fewer emissions, it affects global warming and climate change in a lesser proportion than fossil fuels [1–4]. The massive use of fossil fuels in recent years has led to an increase in global temperatures [5, 6]. It has been observed that the use of this fossil fuel affects the growth of the fetus, due to the presence of various emissions such as metals, particles, SO_X, NOx, and polycyclic aromatic hydrocarbons [7]. Also, it has been noted that the combustion of these fuels leads to an increase in carbon dioxide, which in turn leads to heat emission [8, 9]. With regard to NOx emissions, they are compounds formed when nitrogen interacts with oxygen at high temperatures and pressures of the engine, which in turn leads to the destruction of the ozone layer and the formation of acid rain [10, 11]. As for carbon monoxide, it results from incomplete combustion, which in turn leads to poisoning which is responsible for more than 50% of deaths related to poisoning [12]. Due to the various emissions mentioned earlier, the use of alternative and sustainable fuels in IC engines has proven to have many advantages [13]. There are many solutions to use alternative fuels to reduce the use of fossil fuels in engines. One of these solutions is the use of a mixture of fossil and alternative fuels together. Research has shown that the use of blended fuels can significantly decrease the consumption of conventional fuels [14, 15].

Several researchers have suggested that alcohol fuel is an applicable alternative fuel option for SI engines, as it has the potential to achieve the desired performance while emitting minimal exhaust emissions [16]. By making use of the high enthalpies of vaporization and octane numbers of alcohols, alcohol-gasoline dual injection engines can effectively prevent engine knocking. This is achieved through the injection of both gasoline and alcohol into the engine, allowing for optimal combustion and engine performance [17]. This high enthalpy also produces lower quantities of NOx emissions and leads to a reduction of HC,CO and CO₂ emissions additionally, it increases both engine torque and BSFC (brake-specific fuel consumption) [18, 19]. Ethanol is a clean fuel that may be mixed with petroleum, this can increase the amount of oxygen available for combustion in spark ignition engines, resulting in lower emissions [20]. Using ethanol as a fuel source is a promising alternative to conventional fuels, and its production is steadily increasing worldwide. Ethanol is an oxygen-containing compound that has been shown to produce lower particulate matter emissions and significantly reduce greenhouse gas emissions. By increasing the fuel's level of oxygen, ethanol facilitates more complete combustion [21]. Zhao et al indicated that increasing ethanol percentage leads to a reduction of harmful exhaust gases, also, they have found that octane number is higher for ethanol blends, however, using pure ethanol at lower loads, fuel atomization and mixture formation may be inadequate and poor. Pure gasoline was found to have a higher heating value, but adding ethanol to the fuel enhances the performance parameters [22]. A Verma, N S Dugala, and S Singh studied A 4-stroke 4-cylinder MPFI SI engine and they found that it can run on up to 20% ethanol-blend premium petrol without any changes. Raising the ethanol blend content causes the fuel consumption for brakes to increase significantly, this might be explained by ethanol's reduced calorific value, which is less 30% of gasoline's [23]. The usage of ethanol is becoming increasingly popular due to its renewable nature and lower CO₂ emissions compared to gasoline. Production of ethanol from biomass is especially promising in this regard. One of the biggest advantages of using ethanol in SI engines is that it raises the fuel octane and combustion speed, but one of its disadvantages is that it does not ignite at low temperatures [24]. Kumbhar and Khot used different percentages of ethanol (E0, E20, E40, and E60) in a combustion engine without any modifications to it, they observed that increasing the percentage of ethanol has a positive effect on NOx, CO, and HC emissions [25]. Mohammed et al indicated that when using gasoline fuel with ethanol in a one-cylinder, 4-stroke, SI engine, it was found that the increase in ethanol improves BP, BTE, BSFC, and octane number and reduces harmful emissions [22]. Najafi, Ghobadian et al used a 4-stroke SI engine with different ethanol percentages (E0, E5, E10, E15, and E20), it was found that there was a small increase in BP, T, BTE, and volumetric efficiency (VE), while there is a decrease in BSFC, because ethanol contains more percentages of oxygen, which led to lower concentrations CO2 and NOx emissions. However, one drawback was noted, the use of ethanol caused an increase in emissions of carbon dioxide and nitrogen oxides [26]. While gasoline (E0) and ethanol with ethanol (E50 and E85) were used in a single-cylinder 4-stroke SI engine the engine ran at two compression ratios (10:1 and 11:1) and was tested across a range of engine speeds (1,500 to 5,000 rpm). When the throttle is wide open, it was observed that increasing the percentage of ethanol led to an increase in BP, T, and fuel consumption, but it led to a decrease in emissions of CO, NOx, and HC. Furthermore, it has been observed that mixing ethanol with gasoline allows an increase in compression ratio (CR) without causing knocking [18]. Yu et al have studied a 4-cylinder SI engine with GPI and EDI using gasoline port injection plus ethanol direct injection, it has been noted that the use of 12% EGR improves engine performance and emissions, using an open-loop control system. The engine is tested at 1500 rpm and DIT by 120 °CA BTDC, with different levels of EGR and EDIr, maintaining an excess air ratio λ value of 1 [27].

Chaimanatsakun et al used a 1.3-L four-cylinder Gasoline Direct Injection (GDI) engine to study the impact of ethanol-gasoline blends (E10, E20) and reformed exhaust gas recirculation (REGR) on engine performance and emissions. Results showed significant improvements with ethanol blends and REGR. Brake-specific energy consumption (BSEC) improved by up to 11%, and brake thermal efficiency (BTE) increased by up to 12.4%. Carbon monoxide (BSCO) emissions decreased by up to 38%, and nitrogen oxides (BSNO) by up to 86%. Additionally, total particulate matter (PM) was significantly reduced, particularly in the accumulation mode, by up to 48% with E20 [28]. Al-Hasan used a Toyota Tercel-3A SI engine and was tested with ethanol-unleaded gasoline blends. Ethanol addition improved engine performance, increasing BP, BTE, and VE by 8.3%, 9.0%, and 7% on average, respectively. Fuel consumption rose by about 5.7%, while BSFC and air-fuel ratio decreased by approximately 2.4% and 3.7%. Exhaust emissions were significantly reduced, with CO and HC emissions decreasing by 46.5% and 24.3%, respectively. However, CO₂ emissions increased by 7.5%. The 20% ethanol blend yielded the best results [29]. Kumar used a four-stroke single-cylinder multi-fuel engine with variable compression, testing ethanol-blended fuels (E0, E5, E10, E20) to assess their impact on engine performance and emissions. It found that ethanol blends enhance engine efficiency, with E20 showing optimal BP. Brake power increased by 5% with 10% ethanol, indicating ethanol's positive effect on engine output. Thermal efficiency peaked at 24.5% for E20, while E10 and E5 followed closely. Ethanol's inclusion improved BTE and reduced SFC, illustrating its potential as a gasoline substitute for improved engine performance and reduced emissions [30]. Verma, Dugala, and Singh used a four-stroke, four-cylinder MPFI SI engine, and the impact of ethanol-premium gasoline blends on engine performance was assessed. Key parameters like Brake Torque (BT), Brake Power (BP), and Brake Specific Fuel Consumption (BSFC) were measured at speeds of 2200, 3200, and 4200 rpm under loads of 5, 10, 15, and 20 kg. The study revealed that the optimal blend (F2 sample with 6.25% ethanol) achieved a maximum BT of 104 Nm at 3000 rpm and 20 kg load, and a maximum BP of 24.5 kW at 4200 rpm and 20 kg load. Higher ethanol concentrations generally led to increased BSFC, with the lowest being 0.268 kg kW⁻¹h⁻¹ at 3200 rpm and 20 kg load for the F2 sample [23].

Machine learning is an important tool for prediction of SI engine performance and emissions due to their numerical capabilities [31, 32]. ANN (artificial neural network) is a machine learning algorithm designed to simulate the structure and operations of the human brain. ANN is made up of cells linked together to analyze different processes, so they are a powerful tool for tasks such as analyzing and predicting complex data [33]. These networks are used such as classification, prediction, and pattern recognition. It is able to process complex data and extract meaningful insights from it [34]. These networks have proven to be highly efficient in predicting IC engines as they act as an effective alternative to traditional methods [35]. Najafi et al studied the relationship between BP, T, BSFC, BTE, VE, and emission values in a 4-stroke SI engine operating on different gasoline with ethanol percentages and speeds, ANN model was developed. The training data for the model was 70% and 30% of the total experimental data, respectively. In the model, the back propagation method was used. Engine performance and exhaust emissions might be predicted by the ANN model with a high value of (R²) ranging from 0.97 to 1. It was found that MRE was between 0.46% and 5.57%, while RMSE was very low. These results suggest that the ANN approach can be a reliable method for precisely forecasting SI engine emissions as well as performance [26]. Uslu and Celik. used experimental data for the ANN model, to predict BMEP, BSFC, BT, NOx, HC, and CO based on engine speed (rpm), fuel blending ratio, and CR. RSM was also applied to identify the optimal engine operating conditions. It is shown that the ANN with RSM support can accurately estimate engine performance and emissions with R² between 0.94 and 0.99, and a MRE of less than 7% as compared to the experimental outcomes [36]. Deh Kiani, et al created an ANN model using SBPA for the engine, with some engine experimentation results used (KIA 1.3 SOHC engine) for training. To confirm the ANN's accuracy, the prediction dataset was compared to the target data. The results indicated that the ANN had high performance in predicting the emission indices, with R² of 0.98, 0.96, 0.90, and 0.71 for CO, CO₂, HC, and NOx, respectively. R² for T and BP were 0.99 and 0.96, respectively [37]. S H Hosseini et al developed an ANN model for an engine (Diesel engine single-cylinder, 4-stroke, air-cooled) using SBPA. The input and target parameters were mapped nonlinearly using an MLP network. Fuel blend, S, fuel density, fuel viscosity, LHV, intake manifold pressure, fuel consumption, EGT, oxygen content in exhaust gases, oil temperature, relative humidity, and ambient air pressure were among the input parameters. Engine torque, power, CO, CO₂, UHC, NO, RMS, and kurtosis of engine vibration were the desired outcomes. The (logsig-logsig) activation function with 25–25 neurons in hidden layers was found to be the best ANN model for the back-propagation training procedure. The R² for training, validation, and testing were 0.9999, 0.9994, and 0.9995, respectively, indicating that the model was able to predict various engine parameters for diverse scenarios with high accuracy [38]. W Liu, M Safdari Shadloo et al used a neural network model to predict engine performance and exhaust emissions, the necessary data for testing and training was obtained by varying the percentage of alcohol at various speeds and compression ratios. The training process utilized six experimental datasets, and the resulting ANN model was developed using a standard program. The performance of ANN was evaluated using MSE, R², and AARD%. The AARD% values for CO and HC emissions were 10.50% and 15.45%, respectively, while those for T and fuel consumption were 10.50% and 3.13%, respectively. These error values were deemed acceptable when compared to experimental uncertainty [39].

Eckart et al examined and compared ANN with other machine-learning techniques like GLM, SVM, and RF. The focus is on their effectiveness in predicting laminar burning velocities in hydrogen-methane mixtures. ANN is highlighted for its accuracy and computational efficiency, emphasizing its importance in combustion research and engine optimization [40]. G Najafi et al used ANN to predict the performance and emissions of a 4-stroke SI engine running on various ethanol-gasoline blends and speeds. The ANN model, trained with 70% of the experimental data and tested with the remaining 30%, employed the back propagation method. The model showed high accuracy with R-squared values ranging from 0.97 to 1. The Mean Relative Error (MRE) varied between 0.46% and 5.57%, and the Root Mean Square Error (RMSE) was notably low. These findings indicate that ANN is a reliable tool for precisely forecasting emissions and performance of SI engines [26]. Sayyed, Das, and Kulkarni utilized an ANN to predict NOx emissions in a Direct Injection Compression Ignition (DICI) engine using multiple biodiesel blends. The study focused on the performance and emission characteristics of the engine. The ANN model, based on the physicochemical properties of the fuel, showed high predictive accuracy with correlation coefficients (R²) ranging from 0.996 to 0.998 across the training, validation, and testing phases. The results indicate the ANN's effectiveness in estimating NO_x emissions, providing a robust tool for predicting engine performance and emissions using biodiesel blends [41].

According to previous studies, many types of alternative fuels can be used in SI engines, on top of which is ethanol. The purpose of this experimental study is to indicate the effect of ethanol-gasoline blends on various performance and emission parameters of a GX35-OHC 4-stroke, air-cooled, single-cylinder gasoline engine. These various parameters are BP, BTE, BSFC, EGT and engine emissions of CO, CO2 and HC. When the percentage of ethanol in the gasoline is changed, the study discusses how this affects the engine's overall performance and emissions. In addition, the ANN technique is created to produce a dependable and effective tool for predicting engine performance and emissions in a SI engine. Python-based code is introduced for ANN known for its consistent and stable results. This code promises to elevate predictive accuracy in IC engine parameters and provide dependable solutions, marking a significant advancement in predictive analysis. The ANN model will be trained and validated using experimental data collected for ethanol blends of 15% by volume. The aim is to provide accurate predictions of engine performance and emissions under different operating conditions and ethanol gasoline blend ratios. The results of this study can guide future research and development of engines that can operate on alternative fuel sources.

A GX35-OHC (Honda) one-cylinder, four-stroke, air-cooled, and spark ignition engine (SI) was tested experimentally to study the effect of ethanol-gasoline blends on the performance and emissions of that engine, see table 1 for its specifications. The engine is fixed on a stand with a model of CT 159 (Gunt Hamberge basic module) that includes flow meters to measure fuel and air consumptions, and sensors to measure intake and exhaust temperatures, these measurements are displayed on digital screens as shown in figure 1. To measure the engine rotational speed and torque, the engine is connected to a dynamometer with a model HM 365 universal brake (max. speed: approx. 3000 min⁻¹, max. torque: approx. 12Nm), as shown in figure 1. All obtained data from the dynamometer are recorded and analyzed by its software. Additionally, this study measures engine emissions, such as HC, CO, and CO2 by using an exhaust gas analyzer of MHC 222. This analyzer uses the principle of infrared radiation absorption to detect the emissions levels [42]. Different ethanol-gasoline blends are used, such as E0, E5, E10, and E15, where E0 represents pure gasoline 92 (With specifications shown in table 2) and other blends represent the volume percentage of ethanol compared to the gasoline. The starting temperature of the engine was not consistent when collecting data for different speeds; it was cold before testing the first speed and then gradually warmed up during the data collection process. However, it's important to note that the engine performance may vary with temperature. Nevertheless, for each parameter such as BP, BSFC, BT, etc, the starting temperatures were the same for the first speed tested.

Zoom In Zoom Out Reset image size

To perform the experimental investigations, firstly, we studied the effect of ethanol-gasoline blends on the performance and emissions of the engine at different engine speeds ranging from 1600 to 2800 rpm. Secondly, the current research focused on the effect of the intake air flow rate and temperature at E15 ethanol-gasoline blend on the engine performance and emissions. To control the air intake conditions, the CT 159 module is provided with an air blower and flow control valve to control the intake air flow rate, also it is provided with an air heater to change the intake air temperature.

ANN can be used to predict engine performance and emissions that operate on E15 fuel using Python programming language. At this experiment, there are five parameters that can affect engine performance and emissions. These parameters are intake air temperature, intake air flow rate, rotational speed, brake torque, and fuel flow rate. The values of these parameters can be determined and used as input neurons for ANN.

The engine performance and emissions are evaluated at different ethanol blends, such as E0, E5, E10, and E15, and at different engine speeds ranging from 1600 to 2800 rpm. In the following sections, we will verify the effect of the ethanol blends on the performance parameters that include; BP (brake power), BSFC (brake specific fuel consumption), BTE (brake thermal efficiency), and EGT (exhaust gas temperature), and also their effect on the emissions (such as, HC, CO, CO₂).

3.1. Brake power (BP)

Brake power (BP) is defined as the product of brake torque and engine rotational speed, as indicated in equation (1) is measured at different engine speeds and different ethanol blends. Figure 2 shows that BP has the same curve trend for all tested ethanol blends, as BP increases with the engine speed. At engine speed of 1600 rpm, the difference in BP between E0 and E15 is 0.0044 kW (3.58% increase), while this difference equals 0.0056 kW (1.821% increase) at 2800 rpm. Thus, when the engine speed is low, the change in BP between ethanol blends is very low, while with increasing engine speeds, the effect of ethanol blends is noticed. This can be explained as at relatively low engine speeds, the engine is not required to give its full power, so the difference in energy output between ethanol and gasoline is not very apparent. However, at relatively high engine speeds, an increase in ethanol blends leads to an increase the output power of the engine compared to pure gasoline. This increase produces a noticeable percentage increase in brake power between E0 and E15.

The results suggested that blending ethanol with gasoline can increase engine performance by increasing output power. This outcome can result from several factors, including the elevated oxygen content in ethanol, this higher oxygen concentration has the capacity to enhance the combustion process, ultimately resulting in a greater energy release during combustion [44]. In addition, ethanol possesses an increased capacity to absorb heat during its transition from liquid to gas due to its higher heat of vaporization. This property offers potential benefits for engines by contributing to improved engine cooling and a reduced risk of engine knock. Consequently, these attributes collectively facilitate the engine's ability to generate higher power levels [45].

Zoom In Zoom Out Reset image size

3.2. Brake specific fuel consumption (BSFC)

Brake Specific Fuel Consumption (BSFC) is defined as the fuel consumed by the engine, measured in kilograms per hour (kg/hr), divided by the brake power in kilowatts (kW), as indicated in equation (2). This metric is recorded across various engine speeds and for different ethanol blends, as illustrated in figure 3. It is noticed that increasing engine speed leads to a decrease in BSFC for all ethanol blends. Furthermore, as the ethanol concentration in the fuel increases, BSFC decreases [46]. At 1600 rpm, the decrease of BSFC between E0 and E15 is 0.0379 kg kW⁻¹ (4.043% decrease), while at 2800 rpm the decrease of BSFC between E0 and E15 is 0.01326 kg kW⁻¹ (2.15% decrease). Thus, at low speeds, there is a noticeable effect of ethanol blends on BSFC, however, this effect decreases with increasing the engine speed. The reason for variations of BSFC between E0 and E15 at lower engine speeds is that the engine is not operating at its maximum fuel consumption rate. Consequently, when ethanol is added to the fuel blend, it enhances the combustion process, resulting in more efficient utilization of fuel and thus a more significant decrease in BSFC. However, at higher engine speeds, the engine is operating at its maximum fuel consumption rate, which means that it is already burning a larger amount of fuel to generate the same amount of output power.

Zoom In Zoom Out Reset image size

3.3. Brake thermal efficiency (BTE)

Brake Thermal Efficiency (BTE) is indeed calculated as the ratio of Brake Power to Fuel Power as indicated in equation (3). Figure 4 shows the variations of the brake thermal efficiency along the engine speed at the different ethanol blends. BTE increases are found to increase with the engine speed for all tested ethanol concentrations. Additionally, an increase in ethanol blends leads to an increase in BTE [47]. This denotes that the engine's efficiency in converting fuel energy into useful mechanical work is improved with the higher concentration of ethanol in the fuel blend. At an engine speed of 1600 rpm, the difference in BTE between E0 and E15 is low (around 0.93417% increase), although BTE difference between E15 and E0 equals 1.1322%, at 2800 rpm. Thus increasing the engine speed enhances the effect of ethanol blends on BTE.

Zoom In Zoom Out Reset image size

3.4. Exhaust gas temperature (EGT)

Exhaust gas temperature (EGT) gives an indication of engine performance as it describes the actual extracted energy from the fuel during the combustion process that is transformed into mechanical work by the engine. Clearly, EGT increases with the engine speed for all tested ethanol blends, as illustrated in figure 5. In addition, at 1600 rpm, EGT decreases by 20 °C from E0 to E15, while at 2800 rpm, EGT is reduced by 36 °C when using E15 compared to E0. Therefore, increasing the ethanol blends leads to a decrease in EGT, and consequently increases engine performance.

Zoom In Zoom Out Reset image size

The decrease in EGT with increasing ethanol blends can be justified by ethanol's higher octane rating and its cooling effect due to its higher heat of vaporization compared to gasoline. The higher octane rating reduces the probability of engine knock, allowing for more efficient combustion at higher compression ratios. Ethanol's cooling effect also means that it absorbs more heat during vaporization, lowering the temperature of the air-fuel mixture and subsequently the combustion temperature. These factors contribute to a lower EGT, which suggests that the engine runs cooler and potentially more efficiently with higher ethanol blends. While lower EGT is generally associated with better engine efficiency and lower NOx emissions due to reduced combustion temperatures, it is essential to also consider other performance metrics.

3.5. HC emissions

The hydrocarbon HC emissions refer to the amount of unburned or partially burned fuel that the engine emits. An increase in the engine speed causes of decrease in the (HC) emissions at different ethanol blends, as verified in figure 6. Furthermore, as more ethanol is added to the fuel, HC emissions decrease at all engine speeds [48]. At the engine speed of 1600 rpm, HC emissions equal 188 ppm, and 175 ppm at ethanol blends of E0, and E15, respectively. This gives an indication that increasing the ethanol blends leads to an improvement in combustion efficiency. Moreover, HC emissions decrease with increasing ethanol percentage when the engine speed increases. For instance, at 2800 rpm, the HC emission values drop from 88 ppm for E0 to 71 ppm for E15, indicating that the decrease in HC emissions is 17 ppm. This means that using ethanol in the fuel blend can substantially impact reducing HC emissions at higher engine speeds.

Zoom In Zoom Out Reset image size

The presence of oxygen in ethanol's molecular structure contributes to more complete combustion, so it reduces the emission of unburned hydrocarbons. Ethanol has a higher latent heat of vaporization than gasoline, which means that it absorbs more heat from the surroundings when it vaporizes. This can cool the intake charge, leading to a denser air-fuel mixture and a reduction in combustion temperatures. Lower combustion temperatures typically result in less incomplete combustion and thus lower HC emissions.

3.6. CO emissions

Carbon monoxide (CO) emissions were measured at different engine speeds for various ethanol blends. The results indicated that the increase in engine speed leads to a decrease in CO emissions for all ethanol blends, as shown in figure 7. CO emission is produced during incomplete combustion that results from insufficient Oxygen during the combustion process. Additionally, increasing the ethanol percentage leads to a reduction in CO emissions. At 1600 rpm, the CO emission values decrease from 1.62% for E0 to 1.45% for E15 with a decrease in CO emissions of 0.17%. Increasing the engine speed by adding ethanol resulted in a significant drop in CO emission values, for instance, at 2800 rpm, the CO emission values decreased from 1.02% at E0 to 0.83% at E15 with a decrease in CO emissions of 0.19%. Thus, adding the quantity of ethanol to the fuel has a greater potential for reducing CO emissions at higher engine speeds. This can be described in terms of how increasing the amount of ethanol in the fuel blend can lead to a more efficient combustion process, which lowers the amount of carbon monoxide (CO) released into the atmosphere. This is because ethanol burns more thoroughly than gasoline and produces less carbon monoxide (CO), as ethanol contains more oxygen than gasoline.

Zoom In Zoom Out Reset image size

3.7. CO₂ emissions

An illustration of the Carbon dioxide emissions along the engine speed at various ethanol blends is presented in figure 8. The results show that an increase in the engine speed and ethanol blends leads to an increase in the amount of CO₂ emission that is released from the engine. At 1600 rpm, the CO₂ emission changes from 2.44% for E0 to 3.46% for E15 with an increase of 1.02%, while, at 2800 rpm, the CO₂ emission changes from 6.24% for E0 to 7.45% for E15 with an increase of 1.21%. This indicates that the use of ethanol in the fuel blend can result in higher CO₂ emissions. These findings imply that increased ethanol content and engine speeds can result in higher carbon dioxide emissions. Since ethanol contains oxygen, it facilitates more efficient burning and boosts energy output per unit of volume. These findings imply that increased ethanol content and engine speeds can result in higher carbon dioxide emissions. Since ethanol contains oxygen, it facilitates more efficient burning and boosts energy output per unit of volume. Ethanol contains oxygen in its molecular structure, which contributes to more complete combustion of the fuel-air mixture. When ethanol is added to gasoline, it can increase the stoichiometric combustion efficiency, leading to more CO₂ and less CO and HC emissions. Although the combustion is more complete, the carbon in the fuel still oxidizes to CO₂, which can result in higher CO₂ emissions by volume.

Zoom In Zoom Out Reset image size

Several measurements are taken in order to measure the 'n' number of experimental parameters. The necessary outcome of the experiment is estimated using this collection of measurements. Thus, equation (4) can be used to determine the mean of the amount measured in the studies [49].

Xm is the measured value and n is the number of measurements. The formula for standard deviation (SD) is indicated in equation (5) and the uncertainty (U) can be calculated by equation (6) As indicated in table 3. [50, 51].

5.1. ANN architecture

The artificial neural network (ANN) is a computational model for information processing, similar to biological nervous systems. ANN consists of an input layer, a hidden layer, and an output layer, formed by nodes. Neurons in the input layer receive incoming data and pass it to the neurons in the hidden layer for processing [52]. The input parameters that are fed into the input layer of the required network are engine speed (S), engine torque (T), intake air temperature (T_air), intake air flow (Q_air), and fuel consumption (). These parameters have an impact on both the performance and emissions of the engine. The output parameters of the network include BP, BSFC, BTE, EGT, HC emissions, CO emissions, and CO₂ as shown in figure 9. Each output neuron will be provided in a separate ANN for all input neurons to predict the value according to all input parameters. Each neuron in hidden and output layers can be determined by summation and activation functions. The summation function can be expressed mathematically as the sum of the products of the input values and their respective weights, with the addition of a bias term as shown in equation (7), which is then used as the input to the activation function. The activation function is responsible for calculating the output of a neuron based on the weighted sum of its inputs as shown in perceptron (figure 10). The selection of an appropriate activation function depends on the problem being solved and the desired range of output values. The logistic sigmoid activation function has been widely used in multilayer perceptron models due to its differentiability, continuity, and non-linearity [53, 54]. The sigmoid activation function used in this ANN is shown in equation (8).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

5.2. Normalization of experimental data

Normalization involves scaling the input and output data to a common range, typically between 0.1 and 0.9 , to ensure that they are on a similar scale and to facilitate training and improve the performance of the model. This is achieved by applying equation (9) [55].

The terms 'd_max' and 'd_min' refer to the maximum and minimum values of the input data, respectively. The variable 'd_i ' represents the ith data point in the input data. This normalization is achieved by Python programming languge during running the code of the network.

5.3. ANN application

In our regression analysis, we use a feedforward backpropagation network. The Levenberg–Marquardt (LM) backpropagation training algorithm is employed to minimize the error by reducing the difference between the actual outputs and the predicted outputs. We evaluate the configuration with the highest overall performance based on the mean of the performance metrics: Root Mean Square Error (RMSE), proportion of the variance (R²), and Mean Absolute Error (MAE). For all output parameters, configurations with the lowest RMSE and MAE values, as well as the highest R² values, are selected. The network's configuration starts with 3 and goes up to 12 neurons in the hidden layer to determine the optimal number of neurons. These criteria ensure superior overall performance in terms of prediction accuracy, variance explanation, and the absolute error between the predictions and actual values, as detailed in tables 4 and 5.

Table 4. Selection of the highest performance ANN for BP, EGT, HC emissions, and CO emissions.

	BP			EGT			HC emissions			CO emissions
Number of Neurons	RMSE		MAE	RMSE		MAE	RMSE		MAE	RMSE		MAE
3	0.0046	0.9952	1.51	1.3054	0.9999	0.20	1.1989	0.9990	1.36	0.0065	0.9996	0.56
4	0.0048	0.9949	1.57	2.6751	0.9995	0.38	0.6944	0.9997	0.75	0.0068	0.9996	0.64
5	0.0046	0.9952	1.52	5.0993	0.9980	0.81	0.5559	0.9998	0.63	0.0075	0.9995	0.71
6	0.0054	0.9935	1.79	2.6969	0.9995	0.39	0.5300	0.9998	0.54	0.0076	0.9995	0.72
7	0.0050	0.9945	1.67	1.3202	0.9999	0.21	0.5553	0.9998	0.63	0.0071	0.9995	0.65
8	0.0051	0.9941	1.72	2.6683	0.9995	0.38	0.5438	0.9998	0.55	0.0077	0.9995	0.72
9	0.0052	0.9940	1.75	5.0743	0.9980	0.80	0.5575	0.9998	0.55	0.0080	0.9994	0.77
10	0.0047	0.9951	1.55	1.6563	0.9998	0.30	0.5940	0.9998	0.60	0.0081	0.9994	0.79
11	0.0049	0.9947	1.63	1.2164	0.9999	0.22	0.5307	0.9998	0.55	0.0082	0.9994	0.82
12	0.0051	0.9942	1.71	1.6502	0.9998	0.30	0.5965	0.9998	0.61	0.0086	0.9993	0.87

Table 5. Selection of the highest performance ANN for CO2 emissions, BSFC, and BTE.

	CO2 emissions			BSFC			BTE
Number of Neurons	RMSE		MAE	RMSE		MAE	RMSE		MAE
3	0.0402	0.9992	0.56	0.0095	0.8540	0.70	0.0639	0.8966	0.53
4	0.0370	0.9993	0.50	0.0114	0.7944	0.75	0.0864	0.8224	0.71
5	0.0351	0.9994	0.48	0.0120	0.7679	0.85	0.0355	0.9672	0.29
6	0.0348	0.9994	0.47	0.0099	0.8377	0.76	0.0574	0.9160	0.49
7	0.0347	0.9994	0.47	0.0110	0.8119	0.74	0.0558	0.9176	0.49
8	0.0344	0.9994	0.47	0.0100	0.8367	0.73	0.0493	0.9356	0.36
9	0.0347	0.9994	0.47	0.0082	0.8888	0.63	0.0560	0.9176	0.49
10	0.0343	0.9994	0.47	0.0074	0.9089	0.60	0.0334	0.9708	0.29
11	0.0342	0.9994	0.47	0.0097	0.8480	0.71	0.0566	0.9158	0.50
12	0.0345	0.9994	0.47	0.0064	0.9339	0.51	0.0554	0.9188	0.48

According to BP, EGT, HC emissions, CO emissions, and CO₂ emissions, segmoid function is used as an activation function. The learning rate is set to 0.001, momentum constant is 0.1, the number of epochs varies according to each neural network meaning it will update its weights and biases with this value based on the training data. The error goal is set to 1e⁻³⁰ with regularization parameter of 0.01 to prevent overfitting. BSFC and BTE have unstable values so, they use previous parameters but they have some difference as they use ReLU function as an activation function with a lower regularization parameter of 0.001, which may offer different regularization strength.

The performance of the ANN model is evaluated for different numbers of neurons in the hidden layer, and the evaluation metrics used are (RMSE, MAE). The evaluation is conducted for seven different output variables: BP, EGT, HC emissions, CO emissions, CO₂ emissions, BSFC, and BTE. Table 4 describes the performance ANN for BP, EGT, HC emissions, and CO emissions. In the case of BP, the model achieves the highest performance with 3 neurons in the hidden layer, demonstrating a low RMSE of 0.0046, and high of 0.9952, and a low MAE of 1.51. According to EGT, it is noticed that 11 neurons in the hidden layer are the best-performance network as it has the lowest values of RMSE of 1.2164 and MAE of 0.22 with the highest value R² of 0.9999. For HC emissions, it is noticed that 6 neurons in the hidden layer are the best performance network as it has the lowest values of RMSE of 0.5300 and MAE of 0.54 with the highest value R² of 0.9998. Similarly, 3 neurons in the hidden layer produce the best predictions of CO emissions, with an RMSE of 0.0065, an R² of 0.9996, and an MAE of 0.56.

Table 5 gives information on ANN performance for CO₂ emissions, BSFC, and BTE. According to CO₂ emissions, the ANN model reaches its highest performance with 11 neurons in the hidden layer with a low RMSE of 0.0342, and MAE of 0.47, and of 0.9994. The best performance is, however, achieved by BSFC, which uses 12 neurons in the hidden layer and gives a low RMSE of 0.0064, a low MAE of 0.48, and a high R² of 0.9339. Finally, it is found that BTE achieves exceptionally high performance with 10 neurons in the hidden layer, resulting in RMSE of 0.0334, MAE of 0.29, and of 0.9708. The performance trends for different output variables can be attributed to the complexity and nonlinearity of the relationships between the input variables and the corresponding outputs. The ANN model with an appropriate number of neurons can effectively capture and represent these complex relationships, leading to accurate predictions and lower errors.

6.1. ANN prediction

A neural network-based approach is employed to predict crucial engine performance: BP, EGT, HC emissions, CO emissions, CO₂ emissions, BSFC, and BTE. Firstly, regression plots were generated to visualize the correlation between the target and predicted data. These charts demonstrated how well ANN performed at predicting the relevant engine parameters. Notably, for BP, EGT, HC emissions, CO emissions, and CO₂ emissions, the model demonstrated outstanding performance with relatively small differences between the target and predicted values. According to the convergence between target and predicted outputs, these parameters stabilized and became consistent throughout an increasing number of epochs, from 1000 to 5000. However, predicting BSFC and BTE presented difficulties. The projected values for these metrics showed noticeable swings even using 1000 epochs, it was noticed that increasing epochs more than 1000 for these two parameters didn't make any modification so the model's performance was less stable than the other parameters. To increase the precision of the forecasts for BSFC and BTE, additional optimization and investigation of alternative modeling methodologies may be necessary. Plotting the convergence of the target and predicted outputs according to the sample size was the main target of the second analysis. We were able to see how the neural network model over time converged to make accurate predictions thanks to this investigation. The performance of the ANN was increased with an increasing number of epochs for BP, EGT, HC emissions, CO emissions, and CO₂ emissions, reiterating the model's dependability for these measures. However, the predicted values for BSFC and BTE remained with acceptable performance than other parameters with 1000 epochs, indicating the need for further investigation and possible model refinement as indicated in figures 11 and 12.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

6.1.1. Data visualizations and overfitting mitigation strategies

According to the mentioned ANN results, advanced visualizations are used to demonstrate the effectiveness of the model and the specific strategies implemented to avoid the risk of overfitting. By integrating some methods, it is required to present a transparent and robust analysis, reinforcing the generalizability and reliability of the results. As there are two different neural networks one for (BP, EGT, HC, CO, CO₂) and another for (BSFC, BTE), the results for both networks are studied using various methods such as validation set, cross-validation, and learning curve. The results of BP and BTE networks are selected to be studied as follows:

6.1.1.1. Validation set

Utilizing a validation set to compare the predicted values from the trained ANNs against actual values represents a critical step in evaluating model performance and checking if there is any overfitting as shown in table 6. The observed high accuracy in predicting the validation set indicates that the models have high reliability in their predictions, thereby enhancing confidence in their applicability to real-world scenarios. This validation process highlights the significance of comprehensive model evaluation and validation to ensure the reliability and generalization capability of machine learning models. Furthermore, it offers a valuable understanding of potential areas for further model refinement and improvement.

Table 6. Comparative study of actual and predicted outputs for BP and BTE ANNs.

Input Data
Engine Speed rpm	Inlet Temperature °C	Air Flow L/min	Engine Torque N.m	Fuel Consumption Kg/hr	Actual Output Data (BP)	Predicted Output Data (BP)	Actual Output Data (BTE)	Predicted Output Data (BTE)
2820	137.2	132.5	1.175	0.368	0.3469891	0.355137	8.31287121	8.412983
2840	138.6	134	1.186	0.374	0.3527213	0.346018	8.32102672	8.390891
2860	140	135.5	1.197	0.380	0.3584997	0.359898	8.31897356	8.298756
2880	141.4	173	1.208	0.386	0.3643242	0.357777	8.32304378	8.283941
2900	142.8	138.5	1.1219	0.392	0.3701948	0.368655	8.32264863	8.312896

6.1.1.2. Cross-validation

When applying the cross-validation technique on the two networks (BP and BTE) as shown in table 7, it is found that the R² scores for the BP ANN range from approximately 0.9957 to 0.9985 with an average R² of 0.9966 while the R² scores for the BTE ANN show more variability, ranging from approximately 0.7409 to 0.9978 with an average R² of 0.9517. The high R² scores obtained for the BP ANN across different folds of the dataset affirm its ability to generalize well to unseen data. The near values of R² with a high average support the conclusion that the BP ANN is not overfitting. While noticing the variability in the R² scores of the BTE ANN, we conducted thorough analyses to ensure its reliability. Although there is some variance due to the unstable values in experimental data, the BTE ANN still achieves a commendable average R², indicating substantial predictive capability. The use of K Fold Cross-Validation, a strong technique for evaluating model performance, helped validate the generalization of our models. The performance of the BP ANN and the overall satisfactory performance of the BTE ANN across folds provide confidence in their reliability.

Table 7. K Fold Cross-Validation R² Scores for BP and BTE ANNs.

BP	BTE
[0.997792416162308, 0.9940754076638034, 0.9985444831809122, 0.9956917941883376, 0.9965629490071887, 0.996868747853836, 0.9961488436017671, 0.9966151990612169, 0.9961036682792997, 0.9971131870507753]	[0.9746218273028279, 0.7408975860611189, 0.9908478484673438, 0.990501730853401, 0.9961219535571026, 0.9135953435763285, 0.9977687245566996, 0.993018193481204, 0.9231443727156228, 0.996064651556134]
Average R2: 0.9965516696049445	Average R2: 0.9516582232127784

6.1.1.3. Learning curves

The Learning Curve for BP in figure 13(a) shows a steady reduction in the MSE as the number of training examples increases. The curve gradually reduces, indicating that adding more training data reduces the training error to a stable, low value. The cross-validation score also shows a decreasing trend as more training examples are used, closely following the training score but at a slightly higher error rate throughout. The gap between the training and cross-validation scores is relatively small. This gives us an indication that the close proximity of the training and cross-validation curves suggests that the model generalizes well to unseen data. There is no sign of overfitting as the model performs similarly on both training and unseen data.

Zoom In Zoom Out Reset image size

The Learning Curve for BTE in figure 13(b) starts with a high error and then decreases as more training examples are added but with fluctuations. After initial sharp drops, it stabilizes around a low error value. The cross-validation score starts very high, much higher than the training score, and displays significant variability, especially with fewer training examples. As more training examples are added, the cross-validation score improves, suggesting that increasing the dataset size helps the model generalize better. However, the early high variance and the fluctuations indicate that the model might be sensitive to the specific data points in the training set.

This investigation into the performance and emissions of a GX35-OHC 4-stroke gasoline engine indicated that ethanol-gasoline blends positively influence engine operation. Increasing ethanol concentrations, from E0 to E15, demonstrated a slight improvement in BP and BTE, with a notable reduction in BSFC and EGT. Additionally, ethanol increase correlated with reduced HC and CO emissions across the board. Conversely, a rise in CO₂ emissions was observed with higher ethanol blends and engine speeds. Employing a neural network for predictive analysis, specifically a feedforward backpropagation neural network with an LM training algorithm, yielded high accuracy in predicting engine performance and emissions. The model achieved excellent prediction metrics, with R² values exceeding 0.99 for most parameters, signifying the efficacy of ethanol blends in optimizing engine performance while adhering to emission standards. Notwithstanding the success, some challenges persisted in accurately predicting BSFC and BTE, yet the results remained robust with R² values of 0.9339 and 0.9708, respectively. There are various statistical significance methods such as Validation set, Cross-validation indicated to avoid overfitting for the two different networks. While the conclusions presented here summarize the experimental results, it is necessary to acknowledge the study's limitations. The scope of engine performance parameters could be expanded to include additional metrics for a more comprehensive assessment. Future research may explore the long-term effects of ethanol blends on engine durability and maintenance, as well as the broader environmental impact of their use. Further, incorporating automation in data collection could significantly enhance the research methodology, facilitating extensive data acquisition through overnight trials with automated variable adjustments. The integration of real-time AI model training with this extensive dataset would support predictive analytics, allowing for the refinement of engine performance through multi-objective optimization. Such advancements could propel the efficiency, reliability, and environmental sustainability of engine design and operation.

The data cannot be made publicly available upon publication because no suitable repository exists for hosting data in this field of study. The data supporting this study's findings are available upon reasonable request from the authors.