Prediction of Tool Eccentricity Effects on the Mechanical Properties of Friction Stir Welded AA5754-H24 Aluminum Alloy Using ANN Model

The current study uses three different pin eccentricities (e) and six different welding speeds to investigate the impact of pin eccentricity on friction stir welding (FSW) of AA5754-H24. To simulate and forecast the impact of (e) and welding speed on the mechanical properties of friction stir welded joints for (FSWed) AA5754-H24, an artificial neural network (ANN) model was developed. The input parameters for the model in this work are welding speed (WS) and tool pin eccentricity (e). The outputs of the developed ANN model include the mechanical properties of FSW AA5754-H24 (ultimate tensile strength, elongation, hardness of the thermomechanically affected zone (TMAZ), and hardness of the weld nugget zone (NG)). The ANN model yielded a satisfactory performance. The model has been used to predict the mechanical properties of the FSW AA5754 aluminum alloy as a function of TPE and WS with excellent reliability. Experimentally, the tensile strength is increased by increasing both the (e) and the speed, which was already captured from the ANN predictions. The R2 values are higher than 0.97 for all the predictions, reflecting the output quality.


Introduction
As experimentally proven by Thomas et al. [1] in 1991, friction stir welding (FSW) is a solid-state welding process that joins two plates using a non-consumable rotating tool to heat and stir the material plates at the interface of the joint. FSW generates frictional heat through mechanical stirring between the rotating tool and materials to produce strong joints. This process can be utilized for various materials, including aluminum, titanium, steel, and copper alloys, and is known for its ability to produce high-quality defect-free weldments [1]. Friction stir welding (FSW) provides high-integrity joints for welding metal alloys in a solid state. FSW is used in various applications, including shipbuilding, aerospace, automotive, and railways [2][3][4][5][6][7]. For low, medium, and high deformation resistance alloys, investigations have been conducted on how the welding parameters and tool shape affect the quality of the weld [8][9][10][11]. When compared to conventional arc welding methods, FSW offers a number of benefits [12][13][14]. FSW does not experience issues such as thermal distortion, gas The remainder of this article is structured as follows. Section 2 discusses the data collection methodology for the experimental data used to train the ANN. Section 3 outlines the development and training of the ANN model, as well as the produced results. Section 4 presents a discussion of the produced results from the ANN model. Finally, an overall conclusion of this article is presented in Section 5.

Collecting the Experimental Data
At Suez University, the friction stir welding of AA5754-H24 aluminum alloy butt joint was carried out using a locally produced FSW machine [48][49][50][51]. The welding samples are produced on two plates that are 5 mm thick, 100 mm broad, and 120 mm long. Table 1 provides the chemical breakdown of the AA5754-H24 aluminium alloys. Table 2 is a list of the measured mechanical characteristics of the employed Al alloys. The welding tool is made of an HRC62 heat-treated cold-worked tool steel rod W302 (0.1% Si, 0.39% C, 5.2% Cr, 0.40% Mn, 1.4% Mo, 0.95% V, and 90.6 wt%). Three tool designs are prepared, as shown in Figure 1. The first is the tool without eccentricity (T0), in which the pin and shoulder axes are aligned. The second tool contains a pin eccentricity of 0.2 mm (T2), i.e., the pin axis is shifted by 0.2 mm from the tool axis. The third tool contains a pin eccentricity of 0.8 mm (T8), i.e., the axis of the designed pin has been shifted 0.8 mm from the axis of the tool. In every instance, a pin that is 4.6 mm long and 6 mm in diameter with 19 mm shoulder diameter and a 2 • concavity is employed.
The remainder of this article is structured as follows. Section 2 discusses the data collection methodology for the experimental data used to train the ANN. Section 3 outlines the development and training of the ANN model, as well as the produced results. Section 4 presents a discussion of the produced results from the ANN model. Finally, an overall conclusion of this article is presented in Section 5.

Collecting the Experimental Data
At Suez University, the friction stir welding of AA5754-H24 aluminum alloy butt joint was carried out using a locally produced FSW machine [48][49][50][51]. The welding samples are produced on two plates that are 5 mm thick, 100 mm broad, and 120 mm long. Table 1 provides the chemical breakdown of the AA5754-H24 aluminium alloys. Table 2 is a list of the measured mechanical characteristics of the employed Al alloys. The welding tool is made of an HRC62 heat-treated cold-worked tool steel rod W302 (0.1% Si, 0.39% C, 5.2% Cr, 0.40% Mn, 1.4% Mo, 0.95% V, and 90.6 wt%). Three tool designs are prepared, as shown in Figure 1. The first is the tool without eccentricity (T0), in which the pin and shoulder axes are aligned. The second tool contains a pin eccentricity of 0.2 mm (T2), i.e., the pin axis is shifted by 0.2 mm from the tool axis. The third tool contains a pin eccentricity of 0.8 mm (T8), i.e., the axis of the designed pin has been shifted 0.8 mm from the axis of the tool. In every instance, a pin that is 4.6 mm long and 6 mm in diameter with 19 mm shoulder diameter and a 2° concavity is employed.   The welding speeds used in FSW are 50, 100, 150, 200, 300, and 500 mm/min. In each case, a rotation speed of 600 rpm is used along with a 0.2 mm tool plunge depth and a 3 • tilt angle.
By evaluating the joint tensile characteristics and calculating the Vickers hardness in the heat-affected zone (HAZ) and weld nugget zone (WNZ), the FSW joints are evaluated. A "Instron" tensile testing machine with a 300 KN capability is used for tensile testing. Flat tensile specimens with the dimensions specified in Figure 2 are cut perpendicular to the welding direction. The moving head of the machine moves at a velocity of 0.1 mm/s in accordance with ASTM: E8/E8M-16a standard. With a 1 kg load close to the section centerline, the Vickers hardness of the weld nuggets and the thermomechanically affected zone (TMAZ) of the cross sections perpendicular to the welding line direction are measured. To evaluate the mechanical performance of the welding, measurements were made of the ultimate tensile strength, elongation, and hardness of FSWed nugget zone (WNZ) and thermomechanically affected zone (TMAZ). The measured data are shown in Table 3. The welding speeds used in FSW are 50, 100, 150, 200, 300, and 500 mm/min. In each case, a rotation speed of 600 rpm is used along with a 0.2 mm tool plunge depth and a 3° tilt angle.
By evaluating the joint tensile characteristics and calculating the Vickers hardness in the heat-affected zone (HAZ) and weld nugget zone (WNZ), the FSW joints are evaluated. A "Instron" tensile testing machine with a 300 KN capability is used for tensile testing. Flat tensile specimens with the dimensions specified in Figure 2 are cut perpendicular to the welding direction. The moving head of the machine moves at a velocity of 0.1 mm/s in accordance with ASTM: E8/E8M-16a standard. With a 1 kg load close to the section centerline, the Vickers hardness of the weld nuggets and the thermomechanically affected zone (TMAZ) of the cross sections perpendicular to the welding line direction are measured. To evaluate the mechanical performance of the welding, measurements were made of the ultimate tensile strength, elongation, and hardness of FSWed nugget zone (WNZ) and thermomechanically affected zone (TMAZ). The measured data are shown in Table 3.  For FSW joints made with T0 and T2 tools, it has been found that the tensile strength increases as the welding speed rises. The highest possible tensile strength is reached at 500 mm/min welding speed, and it lowers for the FSW joints made with tool T8 (highest speed). This may be caused by the rate at which heat is applied, which regulates the amount of plastic deformation that occurs. Moreover, a tool pin that is eccentric by more than 0.3 mm produces more heat input [42,52]. This causes a stress concentration, which greatly worsens the mechanical characteristics of the weld joint. Additionally, lowering  For FSW joints made with T0 and T2 tools, it has been found that the tensile strength increases as the welding speed rises. The highest possible tensile strength is reached at 500 mm/min welding speed, and it lowers for the FSW joints made with tool T8 (highest speed). This may be caused by the rate at which heat is applied, which regulates the amount of plastic deformation that occurs. Moreover, a tool pin that is eccentric by more than 0.3 mm produces more heat input [42,52]. This causes a stress concentration, which greatly worsens the mechanical characteristics of the weld joint. Additionally, lowering the rotational speed and increasing the welding speed (by increasing the revolutionary pitch) reduces the heat input. This increases the tensile strength as the welding speed is increased. However, decreasing the welding speed causes more heat input and a low tensile strength [36]. The FSWed nugget zone and the TMAZ are used to measure the hardness values of the FSW welded joints. The junction with the maximum hardness is one that has been welded using a T2 tool with an eccentricity of (e 0.2) at a welding speed of 500 mm/min. The TMAZ is found to have less hardness than the weld nugget zone. Figure 3 shows the top surface view of the AA5754-H24 FSW joints at rotating speeds of 600 rpm and various welding speeds of 50, 100, 150, 200, 300, and 500 mm/min using the FSW tools T0, T2, and T8 previously discussed. All joints appear nearly identical on the top surface, which is mostly affected by the shoulder, with the exception that the top surfaces of joints fused at low welding speeds are smoother and have less flash on both the advancing side (AS) and the retreating side (RS). On the other hand, it can be observed that the welding speed has a noticeable impact on the way the surface of welded joints looks. At these high welding speeds, beginning at 300 mm/min, the semicircular banding features visibly grow broader. The measured semicircular banding features spacing agrees with the calculated ones, claims Krishnan [53]. For the welding speeds of 50, 100, 150, 200, 300, and 500 mm/min, the measured semicircular spacing (revolutionary pitch) in this study is 0.083, 0.167, 0.25, 0.333, 0.5, and 0.83 mm, respectively. This spacing appears to be almost identical to the spacing in Figure 3.
the rotational speed and increasing the welding speed (by increasing the revolutionary pitch) reduces the heat input. This increases the tensile strength as the welding speed is increased. However, decreasing the welding speed causes more heat input and a low tensile strength [36].
The FSWed nugget zone and the TMAZ are used to measure the hardness values of the FSW welded joints. The junction with the maximum hardness is one that has been welded using a T2 tool with an eccentricity of (e 0.2) at a welding speed of 500 mm/min. The TMAZ is found to have less hardness than the weld nugget zone. Figure 3 shows the top surface view of the AA5754-H24 FSW joints at rotating speeds of 600 rpm and various welding speeds of 50, 100, 150, 200, 300, and 500 mm/min using the FSW tools T0, T2, and T8 previously discussed. All joints appear nearly identical on the top surface, which is mostly affected by the shoulder, with the exception that the top surfaces of joints fused at low welding speeds are smoother and have less flash on both the advancing side (AS) and the retreating side (RS). On the other hand, it can be observed that the welding speed has a noticeable impact on the way the surface of welded joints looks. At these high welding speeds, beginning at 300 mm/min, the semicircular banding features visibly grow broader. The measured semicircular banding features spacing agrees with the calculated ones, claims Krishnan [53]. For the welding speeds of 50, 100, 150, 200, 300, and 500 mm/min, the measured semicircular spacing (revolutionary pitch) in this study is 0.083, 0.167, 0.25, 0.333, 0.5, and 0.83 mm, respectively. This spacing appears to be almost identical to the spacing in Figure 3.

Methodology of the ANN Model
Most manufacturing processes are complex in nature, highly non-linear, and have a large number of input parameters. Currently, no mathematical models can describe the behavior of these processes. Due to the cost-effective and relatively easily understandable nature of ANN models and their ability to be trained using data collected from these complex manufacturing processes, they have been extensively used as predictive techniques [54,55].

Methodology of the ANN Model
Most manufacturing processes are complex in nature, highly non-linear, and have a large number of input parameters. Currently, no mathematical models can describe the behavior of these processes. Due to the cost-effective and relatively easily understandable nature of ANN models and their ability to be trained using data collected from these complex manufacturing processes, they have been extensively used as predictive techniques [54,55].
The input layer, the hidden layer, and the output layer are the three layers that constitute an ANN. All of the input parameters are contained in the input layer. The hidden layer processes data from the input layer, after which the final (output) layer computes the next output vector. In Figure 4, the three layers of the ANN model utilized in this study are shown schematically. The input layer, the hidden layer, and the output layer are the three layers that constitute an ANN. All of the input parameters are contained in the input layer. The hidden layer processes data from the input layer, after which the final (output) layer computes the next output vector. In Figure 4, the three layers of the ANN model utilized in this study are shown schematically. An ANN consists of simple synchronous processing elements that are inspired by the biological nervous system. The basic unit in the ANN is the neuron. Neurons are connected by links known as synapses, and associated with each synapse is a weight factor. In this work, two training algorithms (gradient descent with momentum algorithm, and Levenberg-Marquardt Algorithm) were used with a single hidden layer (eight neurons). The inputs and outputs were normalized in the range of 0:1. Thus, the network architecture consisted of two input neurons, eight hidden neurons with a nonlinear activation function, a logistic sigmoid (logsig), and four output neurons with a linear activation function. Table 4 shows the trained network weights and biases. The ANN is trained and tested using MATLAB. After training the ANN successfully, it is tested using measured data that are different from what was used in the training processes. The number of neurons is increased through the training process from five to nine to define the output accurately. An ANN consists of simple synchronous processing elements that are inspired by the biological nervous system. The basic unit in the ANN is the neuron. Neurons are connected by links known as synapses, and associated with each synapse is a weight factor. In this work, two training algorithms (gradient descent with momentum algorithm, and Levenberg-Marquardt Algorithm) were used with a single hidden layer (eight neurons). The inputs and outputs were normalized in the range of 0:1. Thus, the network architecture consisted of two input neurons, eight hidden neurons with a nonlinear activation function, a logistic sigmoid (logsig), and four output neurons with a linear activation function. Table 4 shows the trained network weights and biases. The ANN is trained and tested using MATLAB. After training the ANN successfully, it is tested using measured data that are different from what was used in the training processes. The number of neurons is increased through the training process from five to nine to define the output accurately.
To train and test an ANN neural network, we need ANN training (patterns), input data, and corresponding target values (measured data). A total of 18 patterns have been obtained in this work from the experiments. The welding speed and tool pin eccentricity are inputs to the network, while the outputs include tensile strength, elongation, TMAZ hardness, and weld nugget zone hardness. As a result, the architecture of the ANN changes to 2-8-4, where 2 represents the input values, 8 represents the number of hidden layer neurons, and 4 represents the outputs. Four measured results have been utilized as test data, and fourteen out of a total of eighteen measured results have been used as data sets to train the network (75% training data against 25% test data). The measured results are used to develop and test the ANN model. Figure 5 shows the variation in the mechanical properties of AA5754-H24 FSW joints with tool pin eccentricity and welding speed, and the predicted new data obtained from the network after the training process. As can be seen from the figure, the predicted values match very well with the measured data.

Results and Discussion
changes to 2-8-4, where 2 represents the input values, 8 represents the number of hidd layer neurons, and 4 represents the outputs. Four measured results have been utilized test data, and fourteen out of a total of eighteen measured results have been used as d sets to train the network (75% training data against 25% test data). The measured resu are used to develop and test the ANN model. Figure 5 shows the variation in the mechanical properties of AA5754-H24 FSW joi with tool pin eccentricity and welding speed, and the predicted new data obtained fr the network after the training process. As can be seen from the figure, the predicted valu match very well with the measured data. The results of the training and testing of the ANN model for the measured a predicted data are given in Figure 6a,b. The results show that both the measured a predicted data sets are highly similar, with variance values (R) of 0.9986 and 0.9889 training and testing, respectively. Figure 6c-f compare the measured and predicted resu of the test data. From these figures, it can be observed that the measured and predic values are also highly similar. The ability of a neural network to reliably anticipate results of the test data that have not yet been seen is, nevertheless, its main criterion quality.

Results and Discussion
The arithmetical mean of the number of inputs and outputs is typically used determine how many neurons should be employed in the hidden layer. As a result, i advised to employ a scaled conjugate gradient (SCG) with eight hidden layer neurons the current application. Five to nine hidden layers are used in this application's testi Based on these outcomes, the model performed well. The network with nine hidden la neurons has produced the same accuracy of results for the eight hidden layer neuro Accordingly, only eight hidden layer neurons have been used in the present applicatio The results of the training and testing of the ANN model for the measured and predicted data are given in Figure 6a,b. The results show that both the measured and predicted data sets are highly similar, with variance values (R) of 0.9986 and 0.9889 for training and testing, respectively. Figure 6c-f compare the measured and predicted results of the test data. From these figures, it can be observed that the measured and predicted values are also highly similar. The ability of a neural network to reliably anticipate the results of the test data that have not yet been seen is, nevertheless, its main criterion for quality.
The arithmetical mean of the number of inputs and outputs is typically used to determine how many neurons should be employed in the hidden layer. As a result, it is advised to employ a scaled conjugate gradient (SCG) with eight hidden layer neurons in the current application. Five to nine hidden layers are used in this application's testing. Based on these outcomes, the model performed well. The network with nine hidden layer neurons has produced the same accuracy of results for the eight hidden layer neurons. Accordingly, only eight hidden layer neurons have been used in the present application. The prediction accuracy for the results produced by the network is based on statistical methods. Errors occurring during the training and testing processes are called the root-mean-squared (RMS), an absolute fraction of variance (R2), and mean error percentage values. These are calculated as follows [14]: The results are shown in Table 5. As can be seen in Table 5, the absolute fraction of variance (R2) values for the outputs of the tensile strength, elongation, hardness of TMAZ and weld metal for the training data are all greater than 0.99 and for the test data are all greater than 0.98, except for the The prediction accuracy for the results produced by the network is based on statistical methods. Errors occurring during the training and testing processes are called the rootmean-squared (RMS), an absolute fraction of variance (R2), and mean error percentage values. These are calculated as follows [14]: The results are shown in Table 5. As can be seen in Table 5, the absolute fraction of variance (R2) values for the outputs of the tensile strength, elongation, hardness of TMAZ and weld metal for the training data are all greater than 0.99 and for the test data are all greater than 0.98, except for the hardness of the TMAZ, which is 0.97, respectively. Based on these results, it can be concluded that the proposed model is highly accurate.

Conclusions
In this study, AA5754-H24 has been friction-welded using a wide range of welding speeds at three different tool pin eccentricities, and an ANN model has been developed to predict mechanical properties. The experimental results have been used to train the ANN model used afterward for prediction. Based on the obtained results, the following conclusions can be outlined:

•
The ANN model has been developed based on the FSW experimental work data of AA5754-H24, in which FS was welded using 0, 0.2, and 0.8 mm TPE and welding speeds of 50, 100, 150, 200, 300, and 500 mm/min. • The ANN model was successfully used to predict the effect of tool pin eccentricity on the mechanical properties of FSW AA 5547-H24, and the networks can be used as an alternative.

•
The RMS error values for the ultimate tensile strength, elongation, hardness of the TMAZ, and weld metal for the test data were 1.1346, 0.3515, 1.2759, and 0.3743, respectively; the R2 values are all greater than 0.98, except for the hardness of the TMAZ, which is 0.97. • It is found that the correlations between the measured and predicted values of the ultimate tensile strength, elongation, and hardness of the weld metal are better than those of the hardness of the TMAZ. Data Availability Statement: Data will be available upon request from the corresponding authors.

Acknowledgments:
The authors would like to acknowledge the support of Prince Sultan University for paying the Article Processing Charges (APC) of this publication.

Conflicts of Interest:
The authors declare no conflict of interest.