Evaluating the Performance of Aluminum Oxide Nanoparticle-Modified Asphalt Binder and Modelling the Viscoelastic Properties by Using Artificial Neural Networks and Support Vector Machines

,


Introduction
Bituminous materials are found as natural deposits or obtained as a product from the process of the distillation of crude oil. Te quality of bitumen relies on its adhesive properties and depends on the source it is obtained [1]. Bitumen is a viscoelastic material, meaning that it softens when subjected to heat and hardens when cooled. Tey are categorized by their dependability or capability to fow at wide ranges of temperatures [2]. Due to the limitation of bitumen regarding the temperatures, adding additives to base bitumen is one of the best techniques to enhance the properties of bitumen. Many materials were used as modifers of bitumen, namely, polymers, plastic, steel slag, and glass. It was indicated that the usage of polymers to modify bitumen was able to improve the quality and mitigate the distresses on the mixture which leads to an increase in the durability [3][4][5]. On the other hand, the incompatibility between the bitumen and polymers leads to phase separation among the blends which reduces the strength of the pavement [6,7].
As a result, nanomaterials have been introduced as another way to enhance the bitumen properties and to improve compatibility between the bitumen and the polymers. Nanomaterials have been developed and incorporated rapidly in the design of fexible pavement roads as it has unique properties. Tese properties include high surface work, a large fraction of surface atoms, structural features, quantum efects, and spatial confnement [8]. It was predictable that modifcation of bitumen with nanomaterials would improve the properties of bitumen including an increase in stifness, less temperature susceptibility, and improved strength against weather-induced and moisture damage. Some nanomaterials that have been used to modify polymer-modifed bitumen includes nanoclay, nanofbers, carbon nanotubes, and nanosilica. In the previous studies conducted by using the aforementioned nanomaterials, it has been reported that the performance of bitumen-modifed with nanomaterials showed signifcant improvement in the complex modulus while the phase angle was reduced, indicating that the rutting parameter of modifed bitumen could be minimized this way [9][10][11].
On the other hand, although nanomaterials are promising, some of them are expensive and demand further research for exploration and optimization for bitumen modifcation purposes prior to feld testing and applications [9]. In addition, laboratory testing of materials is time-and resource-intensive. In this vein, advanced statistical modeling and soft computing modeling approaches are in vogue in the prediction of the performance characteristics of asphalt binders [12].
Artifcial neural networks (ANN) is a nonlinear, nonparametric modelling technique that is among the soft computing methods which have proved to ofer a high degree of accuracy in computational prediction of the results observed in laboratory testing [13]. A signifcant amount of research has been dedicated to modeling of asphalt mixtures with the ANN, however, investigations on binder performance were mostly limited to be addressed by the experimental procedures. Lui et al. [14] studied the ANN and Iowa models to predict dynamic modulus of base and recycled shingle asphalt mixtures. El-Badawy et al. [15] compared regression models and ANN models with Witczak and Hirsch predictive models for forecasting dynamic modulus of asphalt mixture. Tapkın et al. [16] presented a study to predict the strain accumulation in polypropylene (PP)modifed marshal samples. Abedali [17] conducted a comparison study among the performance of multiple linear regression models (MLR) and ANN with base asphalt binder. Ziari et al. [18] perfomed a similar study to Abedali with carbon nanotube (CNT)-modifed asphalt binders to predict the rutting performance. Venudharan and Biligiri [13] employed the ANN to predict the rutting performance of asphalt binders with diferent crumb rubber (CR) gradations. Support vector machines is another subfeld of artifcial intelligence methods which focuses on the development of data modelling algorithms which can learn from trained data points. Recently, SVM is gaining the attention of researchers towards developing classifcation and regression models due to its performance and attractive features such as embodying the structural risk minimization (SRM) principle which is known to perform better than the empirical risk minimization (ERM) principle employed by the neural networks [19]. Previously, Maalouf et al. [20] studied the potential use of SVM to predict resilient modulus of asphalt mixes and compared their model performance with the traditional least squares (LS) method. Gopalakrishnan and Kim [19] utilized the support vector machine approach to predict the stifness for the hot mix asphalt. Zhang et al. [21] conducted a research study to predict Marshall parameters of fexible pavement design by using SVM and genetic programing (GP). Another study conducted by Uwanuakwa et al. [22] utilized Gaussian process regression (GPR), recurrent neural networks (RNN), ANN, and SVM to predict the rutting and fatigue parameters in modifed asphalt binders.
Te objective of the current study was to conduct experimental procedures to assess the rheological and physical characteristics of base and nanoalumina-modifed binders and to utilize artifcial intelligence models by considering the rheological behavior of the asphalt binders under diferent mechanical test conditions in order to forecast the experimental outcomes at a more extensive range of testing conditions which the asphalt is expected to undergo during its lifetime.

Sample Preparation of Modifed Asphalt Binders.
Te base asphalt cement (BAC) used in the current study was 60/ 70 penetration of grade and it was provided by Petronas Petroleum while the additives ASA and Al 2 O 3 were purchased from the Shijiazhuang Chanchiang Corporation Company in China. Physical properties regarding the material properties are listed in Table 1.
Te blending of aluminum oxide nanoparticles (Al 2 O 3 ) into BAC was performed by adding diferent concentrations: 3%, 5%, and 7% by the weight of base asphalt. Te asphalt was heated to a desirable temperature (150°C) then the nanoparticles were incorporated into the blend gradually, and the whole samples were mixed under the temperature of 170°C by using a high shear mixer at a speed of 5000 rpm for 90 minutes.

Physical Properties.
Assessments of physical characteristics of base and modifed asphalt binders were performed by following the American Society for Testing and Materials (ASTM) criteria. Te penetration test was performed according to ASTM D5 and softening point test was implemented according to ASTM D36, while the viscosity test was performed according to ASTM D4402.

Viscoelastic Properties of Modifed Asphalt Binders.
A dynamic shear rheometer (DSR) was used to conduct frequency sweep test to evaluate the changes of viscoelastic properties, stifness (G * ), and phase angle (δ) for base and modifed asphalt binders. Te tests were conducted under various temperatures (10, 15 +10 up to 75°C) by using two diferent spindles; an 8 mm and a 25 mm in diameter for samples that have a height of 2 mm and 1 mm for low and high temperatures, respectively. Te testing frequencies ranged from 0.1592 Hz to 15.92 Hz. In addition, DSR was used to conduct multiple stress creep and recovery (MSCR) at a temperature of 64°C by using a 25 mm diameter spindle with the thickness of 1 mm. In MSCR test, twenty cycles were used with one second of a loading time and nine seconds ofloading and the test was carried out using two stress levels 100 Pa and 3200 Pa [23].

Artifcial Intelligence Modelling and Analysis.
In addition to the mechanical empirical evaluation of the base and modifed asphalt binders, current study proposed artifcial intelligence research methods, namely, artifcial neural networks (ANN) and support vector machines (SVM) as predictive models for the estimation of high and intermediate temperature performance characteristics of asphalt binders. Te database for the artifcial intelligence modelling was built by using the outcomes of the mechanical tests conducted under laboratory conditions. Te data points were frstly normalized in order to improve data integrity and reduce data redundancy and MATLAB 2019b, Math-Works was used to develop the ANN and SVM models.
Te ANN is a system for processing data by mimicking the idea of the way biological neurons work in a human brain. Te ANN is composed of highly interconnected processing constituents called artifcial neurons, which operate in parallel logic and transmit information from one layer to others in serial operations [24]. Information is processed at the neurons, and the signals are passed through connection links, which are associated with weight vectors that determine the strength of the connections. Te functions of these neurons are the frst to sum up all the initially assigned weights from the lower layers and then to process the sum by a linear or nonlinear activation function to determine the output signal from the given input signals [25]. According to Venudharan and Biligiri [13], pure-linear function is categorized as linear, whereas hyperbolic tangent and logarithmic-sigmoidal functions are classifed as nonlinear activation functions. As adopted in this study, hyperbolic tangent activation function is given in the following equation.
where s j � input jth neuron, y j � output jth neuron.
Tis type of machine learning method is referred to as a feed-forward neural network model (FFNNM) and it is adopted in the current study due to its efciency in generalization [26]. As illustrated in Figure 1, the structure of the neural network was composed of three main layers, namely, the input layer, a hidden layer, and the output layer where the fow of information was from the input layer to the output layer. FFNNM was utilized with a backpropagation method where the aim was to fnd the optimum weights as an iterative process to minimize the error between the target and computed outputs with a selected accuracy [27].
A number of alternative ANN model structures were developed with 762 set of data points collected from the experimental investigations. Te dataset was randomly divided as the training, validation, and testing. Te training dataset was used for ftting the model. Te validation dataset was used to provide an unbiased evaluation of the model ft and to adjust the model hyperparameters. Herein, K-fold cross validation technique was utilized to minimize the bias and variance [28]. Finally, the test dataset was used in the evaluation of model performance by using untrained dataset.
Two diferent scenarios were modelled based on the experimental outcomes for G * and δ of base and modifed asphalt binders. Levenberg-Marquardt (LVM) was the training algorithm adopted in the ANN modeling. Te input parameters (x i ) considered to have infuence on the target parameters were the mechanical test conditions, temperature (T) and frequency (F) for both models. Te output Advances in Materials Science and Engineering parameters (y i ) were G * and δ for scenario 1 and scenario 2, respectively. Te approach to fnd the optimum model was employed by the trial and error method by using diferent ANN architectures. Support Vector Machines (SVM) is another signifcant computational modelling tool that has been used in solving numerous engineering problems including classifcation and regression. Te aim of this technique is to learn the boundaries between the two classes through mapping the inputs to a high dimensional region. Initially, SVM's were used as a classifer for character and object recognition tasks; however, more recently the technique has also been proven to be efective in regression and time series prediction applications. SVM modelling technique has been derived from the statistical learning theory which has a number of distinct characteristics as compared to the traditional neural networks such as using a set of linear functions defned in a high dimensional space, carrying out risk minimization by using the loss function, and utilizing a risk function which consists of the empirical error and a regularization term which is derived from structural risk minimization (SRM).
Assuming that, where x i is the input and y i is the output parameters, the optimum SVM function is expressed as follows.
where w is the weight factor, φ(x) represents the activation function, and b is the bias. Te accuracy of the proposed model relies on the minimization of the regularized risk function, which is as expressed as follows .
With ε-insensitive loss function, where the frst term 1/2‖ω‖ 2 measures the fatness of the function, and the second term C1/n n i�1 L Ɛ (y, f(x)) is the empirical error. Te regularization parameter C determines the trade-of between the empirical error and the fatness of the model.
In other words, the suitable selection of user-defned parameters, which are regularization parameter C, radius of insensitive tube ε, and the parameter of RBF kernel function c is crucially important to improve the model accuracy. Regularization parameter C controls the approximate function smoothness or fatness for the SVM where higher values of C represents higher penalty of errors and vice versa. Another parameter that controls the accuracy of the approximation function is the ε parameter which has significant efects on the number of support vectors. Te ε parameter smoothens or complexifes the approximate function where a smaller value of ε leads to less support vectors and results in less complex learning machines and vice versa.
Te SVM modelling approach adopts the idea of kernel functions. Herein, the radial basis kernel function (RBF) as expressed in equation (5) was utilized.
where K is the kernel, M is the training input, N is the unlabeled input, and c is the kernel specifc parameter.
In order to select the most suitable user-defned parameters, the manual method involving larger number of trials was used. Parameters combination resulting in the smallest value of root mean squared error (RMSE) and higher value of coefcient of determination (R 2 ) was selected. After developing both prediction models by using the ANN and SVM, the systematic evaluation of the model efciencies was performed in conjunction with statistical performance indicator metrics. Model prediction capacity with training and the testing datasets were evaluated by performance indicator metrics, which were mean squared error (MSE), root mean squared error (RMSE), and coeffcient of determination (R 2 ) as given in the following equations.
where c i is the ith experimentally observed data and c i is the ith target data predicted by the ANN and SVM models.

Efects of Nanoparticles on Physical
Properties. Te infuences of nanoparticles on the physical properties of modifed binders were illustrated in Figure 2. It was noted that, modifed binders have higher stifness compared to the base binder as the penetration values were declined and the softening point values were risen. Te reduction in penetration was nearly 63% for 5% Al 2 O 3 and 60% for 3% Al 2 O 3 . At 7% Al 2 O 3 , the modifed binders have shown a diferent reaction as the reduction in penetration was slightly less than 3 and 5% concentrations but higher than the base asphalt with a reduction of penetration around 45%. As a response to the reduction in penetration, the softening points of modifed asphalt binders were improved by nearly 15% for 5% Al 2 O 3 , followed by 11% increment for 3% and 7% Al 2 O 3 modifed binders. In addition, the efects of Al 2 O 3 nanoparticles on the viscosity of modifed binders at diferent temperatures are shown in Figure 3. Te Superpave requires that the viscosity of asphalt binders must not be more than 3 Pa·s at 135°C. It was noted that all of the asphalt binders passed the requirement of the Superpave method as the highest number among the blends was 1.04 Pa·s at 135°C for 7% Al 2 O 3 while lowest viscosity noted to be for the base asphalt binder which was 0.50 Pa·s. Te additional increment of the modifer content led to a rise in the viscosity value which was an expected outcome since the modifcation process has led to an increase in stifness for the asphalt binders.

Efects of Nanoparticles on the Isochronal Plots.
Te isochronal plots which represent stifens and elasticity of binders at various temperatures and at a frequency of 1.592 (10 rad/s) are demonstrated in Figures 4 and 5, respectively. Te results showed that the base asphalt binder has the lowest stifness and the highest elasticity, 5% Al 2 O 3 has the maximum stifness and the minimum elasticity among the blends while this was followed by 7% Al 2 O 3 and 3% Al 2 O 3 of the modifer contents. Te results indicated that 5% of Al 2 O 3 has the optimum viscoelastic properties among the blends. In addition, the highest complex modulus resulted in higher stability under permanent deformation at high temperatures, whereas the lower phase angle indicated to a higher resistance to thermal cracking at low temperatures. It is noteworthy to mention that, 7% Al 2 O 3 showed a diferent behavior as the complex modulus was decreased and the phase angle was enhanced at the same time compared to 5% Al 2 O 3 . Tis might be due to the occurrence of agglomeration among the nanoparticles during the mixing process which have afected the dispersion of nanoparticles in the asphalt matrix [11].

Efects of Nanoparticles on the Master Curves.
To better characterize and understand the rheological performance of nanoparticles-modifed asphalt binders, a master curve was constructed. A master curve is a representation of stifness (G * ) and elastic modulus (δ) over a range of temperatures and frequencies by using convenient shifting factors. Te reference temperature (T ref ) was 25°C and data points at other temperatures were shifted horizontally to construct a single curve. Figure 6 illustrates the master curves of G * and δ for the base and modifed asphalt binders. Te greater stifness and elasticity denoted that, the binder has superior viscoelastic properties which indicated better resistance to permanent deformation and fatigue cracking at elevated temperatures and low temperatures. Figure 6 demonstrates that, the base asphalt has the highest elasticity and the lowest stifness, while 5% Al 2 O 3 has the lowest elasticity and highest stifness among the blends. Based on the results revealed by the master curves, it can be concluded that 5% Al 2 O 3 can be considered as the optimum concentration for the modifcation process particularly at high temperatures although the improvement in high temperature performance characteristics was at the expense of losing elastic characteristics for the modifed asphalt binders.

Efects of Nanoparticles on High Temperatures Permanent Deformation.
Bitumen is a thermoplastic material that performs as solid and liquid at low and high temperatures, respectively. Te permanent deformation (rutting) is one of the efects of the most common issue of the pavement at elevated temperatures, and the ASTM revealed that to stand with this issue, the bitumen (unaged) must have permanent deformation more than or equal to 1 kPa. Te efects of rutting on base and modifed binders are displayed in Figure 7. It was noted that, the modifed binders have less temperature susceptibility compared to the base binder. Tis indicated that the modifcation of bitumen using nanoparticles can reduce the permanent deformation and increase the service life of the pavement. Te best performance at 65°C was achieved by 5% Al 2 O 3 concentration with enhancement up to 210%, which was followed by 7% and 3% Al 2 O 3 with improvements around 118% and 44%, Advances in Materials Science and Engineering respectively. Moreover, all the blends passed the requirements of Superpave at 65°C while the 5% Al 2 O 3 has also passed it at 75°C.

Efects of Nanoparticles on Low Temperatures Termal
Cracking. Te Superpave technique considered 5000 Pa as the maximum limit for fatigue-cracking at low and intermediate temperatures. Figure 8 illustrates the infuence of nanoparticles on the fatigue resistance parameter for base and modifed binders. Te results showed that, the modifed asphalt binders have higher stifness in comparison with the base asphalt binder which was designated as the base asphalt and has better elastic properties compared with the modifed binders. Moreover, nanoparticles have a signifcant efect on visco-properties, and it does not have an excellent impact on elastic properties, but all binders were within the specifcation of Superpave which is less than 5000 Pa at 25°C.

Efects of Nanoparticles on the MSCR.
MSCR test was used to simulate the movement of trafc fow on the highway surface and to evaluate the resistance and recovery of asphalt binder to rutting after the applied loads are removed [29]. Figure 9 presents the results of MSCR for the base and modifed asphalt binders under stress levels of 100 and     respectively. Te reduction of the creep compliance parameter was also anticipated from the increase in hardness of modifed asphalt binders as observed from the physical tests. Meanwhile, the base asphalt binder has the maximum sensitivity to permanent deformation at high temperatures with compliance values of 4.04, 168.38 Pa for both stress levels respectively. In addition, the improvements in resistance of permanent deformation were 35% and 44% for 3% Al 2 O 3 and 69% and 73% for 5% Al2O3 and 61% and 69% for 7% Al 2 O 3 at 100 and 3200 Pa individually.

Artifcial Intelligence Model
Results. Support vector machines (SVM) and artifcial neural network (ANN) models were developed by using MATLAB (MathWorks Inc R2019b). Dataset included 852 data points from the mechanical test conditions (testing temperature and loading frequency) as the input parameters and the G * and δ as the output parameters. A total of four individual models were constructed to predict two diferent outputs; G * and δ by using two diferent modelling techniques; ANN (model 1 and model 2) and SVM (model 3 and model 4, respectively). A number of diferent ANN architectures with diferent learning algorithms, activation functions, and a various number of hidden neuron layers were trained iteratively to fnd the best performing network structure. Te optimum model was obtained by using the Levemberg Marquadt (LM) learning algorithm with two inputs, three hidden neurons, and one output topology. Te network utilized backpropagation method to minimize the errors in prediction and the activation function was selected as the hyperbolic tangent activation function. On the other hand, the radial basis kernel function was utilized for the model development by the SVM. After setting the suitable userdefned parameters, the 10-fold cross validation technique was utilized to improve model prediction accuracy. By using this technique, the data points were divided into ten diferent sets. Each algorithm was then trained by using nine of the sets and evaluated using the tenth set. Tis process was repeated 10 times and each time a diferent set was used for testing. Te mean of the predictions from ten diferent models were then used to evaluate the performance of the algorithm.
Te statistical metrics utilized in this study to evaluate the model efciency were the coefcient of determination (R 2 ) and the root mean squared error (RMSE). Figure 10 illustrates the best performing networks for predicting targets of G * and δ with training datasets based on R 2 . Te R 2 values close to +1 indicated that, the model has high capacity for predicting the experimental outcomes. From Figure 10, it was observed that, the ANN model slightly performed better than the SVM model for predicting the G * while the SVM model outperformed the ANN model in predicting the experimentally observed δ with the training datasets.
On the other hand, it should be noted that, a smaller diference between the R 2 values obtained for the training and testing datasets is signifcantly important in order to avoid model overftting. On this basis, the model     performances with the training and testing datasets are presented in Table 2. It was observed that the R 2 obtained for model 1 was 0.989 and 0.873 for training and testing datasets and 0.911 and 0.816 for model 2, respectively. On the other hand, for the SVM models, the R 2 for training and testing datasets were 0.984 and 0.952 for model 3 and 0.929 and 0.861 for model 4, respectively. Based on these results, it was concluded that a smaller gap between the training and testing datasets for the SVM models suggested that those models have not sufered from the overftting phenomenon which could have negatively afect the performance of the models for untrained datasets.
Moreover, besides evaluating the model performance by the R 2 , the residuals distribution of the models have signifcant infuence on the model prediction performance. On this basis the RMSE values for the models are also computed and tabulated in Table 2. A smaller RMSE value indicated smaller distribution of the errors and therefore resulted in better performing models. From Table 1, it was observed that the smallest RMSE values for the testing datasets were  In addition, considering that high R 2 values do not always mean a good regression model, the residual distribution test decides whether the selected model is suitable for describing the empirical data, or if another model is required [30]. Te residuals distribution of the two models by using the testing datasets for predicting the G * and δ are shown in Figures 11 and 12, respectively. As observed from both the fgures, in the SVM model, the residual variations were minimal and consistent, while the deviance in the ANN was signifcantly larger than in the SVM models.

Conclusion
Te physical properties and rheological performance characteristics of aluminum oxide nanoparticle-modifed asphalt binder were investigated by using conventional and DSR testing procedures. Heuristic modelling techniques were utilized to predicting the experimental outcomes namely, G * and δ from the mechanical testing conditions, testing temperatures and frequencies. Based on the mechanical test results and the heuristic analysis to predict the experimental outcomes, the following conclusions were drawn.
(i) Te investigation of rheological properties shows an increase in the complex modulus and decline in the phase angle, indicating better viscoelastic properties which led to mitigate the asphalt distress. (ii) Nanoparticles can increase the hardness of modifed binders which reduce the rutting efects at high temperatures, but it has lacked the efects at low temperatures due to low elastic characteristics. (iii) Both ANN and SVM models were able to predict the targeted outputs, G * and δ, accurately while SVM models were able to capture the nonlinear relationship in the dataset and was shown to be more efcient in predicting the untrained data points.

Data Availability
Te data used to support the fndings of this study are available from the corresponding author upon request.

Conflicts of Interest
Te authors declare that they have no conficts of interest.