0 Introduction

Quenching is a vital process used to improve carbon steel's microstructure, hardness, and other mechanical properties. The AISI 1035 steel alloy has found widespread application in the manufacture of mechanical structures for forged parts. It is utilized for couplings, gears, links, and shafts. The phase changes and the microstructures of the materials are impacted by the sudden cooling. The improvement of the mechanical characteristics produced by the quenching process depends on a number of variables, including the quenching temperature, the chemical composition of the steel alloy, the thermophysics of the cooling medium, and the rate of cooling^[1-3]. The influence of carbon nanotubes (CNTs) on the microstructure, surface roughness, and hardness of AISI 1010 is investigated in this study via quenching. Three different levels of cooling medium are used, namely distilled water and surfactant with CNTs, distilled water with CNTs, and distilled water. The outcome showed that surface roughness, hardness, and microstructures vary with the cooling medium and CNT concentrations, this demonstrated that quenching by CNTs imposes higher roughness than quenching by water. However, specimens quenched by CNTs with a surfactant have harder surfaces^[4]. Previous literature has proposed the quenching of medium carbon steel for the design of optimal heat treatment conditions using the Taguchi technique. Whereby the coolant fluids were engine oil, de-ionized water, and salt solution as base fluid and with dispersed nanoparticles of TiO₂, αAL, and CuO. The experimental temperatures were 400, and 550 ℃. The soaking times were 60, 45, and 30 min, respectively. The experimental outcomes showed that the nanoparticle type TiO₂, tempering time 30 min, tempering temperature (400 ℃), and base fluid engine oil and salt solution positively affect yield strength and ultimate strength. Meanwhile, nanoparticle type CuO, tempering time of 60 min, tempering temperature of 550 ℃, and base fluid deionized water have a significant effect on elongation in medium carbon steel^[5]. The lubricating viscosity can be increased by incorporating copper oxide nanoparticles into 10w40 engine oil. A multilayer perception algorithm-based ANN prediction model was proposed^[6]. In order to maximize the hardness, laboratory-grade carbon nanoparticles were utilized with water as the base fluid for AISI 1045. The concentrations of carbon content varied from 0.1% to 0.5%. The outcome showed that a maximum hardness of about 885.34 HV might be achieved with a concentration value of 0.2%. At the same time, the microstructure of AISI 1045 after quenching mainly consisted of martensite^[7]. The carbon steel was heated to 1000 ℃ and quenched in different mediums to determine a suitable quenching medium for carbon steel. The results presented that increasing the ferrite phase structures leads to a decreased hardness of steel. Therefore, after quenching using water and nanofluids, a martensite phase is formed, which results in an increase in the hardness of steel by about 728 HV^[8]. There are two ways to make nanofluid: the one-step method and the two-step method. In the one-step procedure, nanoparticles are created while being spread evenly throughout the base liquid. While in the later, the nanoparticles are created and subsequently mixed with the basic liquid^[9]. For medium carbon steel, hybrid heat treatment was utilized to enhance the mechanical qualities of CK 35. Samples of carbon steel were subjected to a standard heat treatment at 830 ℃ and quenched in an oil medium. In order to enhance the mechanical properties of the steel, these samples were once again treated by generating impulse signals on the surface^[10]. Tempering temperature values have an impact on the microstructures and characteristics as well^[11]. It has been demonstrated that different quenching media affect the mechanical behavior of carbon steel AISI 1030 differently. Specimens were heated in an electrical furnace for 1 h to 950 ℃. The specimens were cooled using polyvinyl chloride, water, and oil. After an hour tempered at 200 ℃, the specimens were quenched in air. The outcome showed that quenching by water increased the hardness constant, yield stress, and ultimate stress while lowering Young's modulus E. Additionally, the pearlite and ferrite phases of the crystalline structure, as well as the grain size (refining), were improved by water quenching followed by tempering^[12]. An extensive study addressing all facets of quenching using vegetable oil was discussed^[13]. An analysis of the distortion, phase change, hardness, and microstructure of AISI 4410 as a result of the quenching process was conducted numerically using the finite element method. The validity of the simulation was confirmed by experimental verification of the simulated results. The outcomes demonstrated strong convergence between the simulation and experimental results^[14]. The volume fraction of pearlite, bainite, and ferrite in the steel rotor shaft 30CrMoNiV5-11 was predicted using ANN^[15]. The impact of heat treatment on the mechanical characteristics of A357 alloy was estimated using an ANN model with a back-propagation technique. According to the results, a back-propagation model might be adopted in the prediction approach^[16]. ANN was used to predict the anisotropic mechanical properties of Inconel 718 alloy at a temperature between room temperature to 600 ℃, and the result of the model confirmed the result of the f-test and t-test^[17]. The mechanical properties of titanium alloy have been widely exploited in engineering applications. Therefore, a model was proposed to minimize the cost and time for predicting mechanical properties using ANN and a back-propagation algorithm. The percentage error was less than 10%; this might be used to predict and estimate the mechanical properties of titanium alloy^[18]. An ANN model was utilized to analyze sensitivity and parametric based on experiment's result of shot peening process^[19-21]. Deep Cryogenic process was utilized to improve residual stresses and hardness of too steel. The result observations were increasing micro, and macro hardness about 27 HV, and 0.5 HRC respectively. While the amount of residual stress was descend about 35%^[22]. An experimental and statistical investigation on steel alloy was done. The effect of machining parameter was optimized and analyzed by Taguchi method and ANOVA^[23].

In this study, experimental work was conducted to figure out the effect of different quenching mediums using nanofluids such as water/TiO₂, CuO/Oil, Al₂O₃/Oil engine, water/hybrid nanoparticles on the mechanical properties of AISI 1035 steel alloy. Taguchi array is utilized to design the experiments set and analyses the effect of quenching conditions. Sixteen experiments were conducted to optimize the effect on the residual stresses, hardness, and microstructures. The results of experiments have been analyzed by MANOVA in the MINITAB 19 to figure out the largest effect of quenching parameters on mechanical properties. Moreover, an ANN is used to develop original numerical model to estimate how mechanical characteristics will behave during the quenching process. The proposed numerical model for predicting the correlation of mechanical properties is supplemented by experimental data. Both theoretical and experimental investigations can benefit the usage of the experimental data and the suggested ANN model. The ANN performance is evaluated by different statistics criteria such as R², and RMSE.

1 Materials and Experimental 1.1 Materials

Medium carbon steel AISI1035 was purchased from Alhadad for Materials Co., Ltd, Baghdad, Iraq. The chemical composition and mechanical properties of the specimens are listed in Tables 1 and 2. The chemical composition analysis was carried out at Al-Nasser Company for Engineering Inspections.

1.2 Nanofluid Preparation

Four different categories of nanofluids were used in this investigation. Nanoparticles were acquired from US Research Nanomaterial, Inc. Details about the materials, oxides, and base fluids are included in Table 3. Metallic oxides increase their thermal conductivity and viscosity as their specific size decreases^[25]. The nanoparticle size has an impact on the thermophysics of nanofluids. Al₂O₃, TiO₂, and CuO were consequently mechanically milled for a period of 60 min. The nanoparticles were gradually added to the base fluid, which was either distilled water or engine oil, and mixed for 180 min using a magnetic stirrer to produce a homogenous liquid. Water distillation was used to produce the distilled water. To prevent the deposition of nanoparticles, the samples were submerged directly in the nanofluid. The L16 orthogonal array was chosen in this study to optimize the samples' hardness, microstructures, and residual stresses. Nanofluid type, nanoparticular concentrations, and soaking period were all taken into account.

2 Experimental Design

The Taguchi method is an efficient technique that is frequently used in engineering analysis to help reduce the number of experiments and save time. An orthogonal array is regarded by this method's experiment design as being reliable, simple to analyze, and easy to comprehend. The number of elements in the study will determine how many experiments need to be conducted, so it is important to carefully control the significant factors. The signal to noise function (S/N) can be used to measure how far the sleeved parameters are from their ideal values using the loss function^[26]. Small S/N is best utilized for grain size and residual stresses, while large S/N is better used for hardness response. Eqs.(1) and (2) are used to calculate the responses of S/N^[27]. In this study, the design of experiments used the L16 orthogonal array. Four factors were chosen in order to examine their effects on the mechanical properties, namely concentration ratio, nanofluid type, quenching temperature, and soaking time. Only one factor had two levels, while the other three factors had four levels. Tables 4 and 5 include an explanation of the factors and their levels, as well as the number of required runs. A mixed design (1^2 3^4) of the L16 orthogonal array included sixteen experiments.Where 1^2 3^4 represents one factor has two levels, and three factors have four levels.

$ \mathrm{S} / \mathrm{N} =-10 \log \frac{1}{n} \sum\limits_{i=1}^n y_i^2 $

(1)

$ \mathrm{~S} / \mathrm{N} =-10 \log \frac{1}{n} \sum\limits_{i=1}^n 1 / y_i^2 $

(2)

where n is number of runs; y is experimental output.

Hybrid nanoparticles were used to obtain further thermal properties of single nanoparticles^[28]. Nanofluid comprising nanoparticular TiO₂ + CuO based water was used to quench one category of the experimental specimens. An Olympus SZ61, an optical microscope, was used to observe the microstructures and calculate the grain size. In order to conduct the microstructure test and to determine the morphology of the surface by scanning electronic microscopy (SEM), the research specimens were prepared in three steps: mounting, grinding, and polishing. The surfaces of the specimens were wet ground on SiC sheets with grit sizes 180, 320, 600, 800, and 1100. On a rotary pre-grinder polisher, the specimens were polished using Al₂O₃ alumina powder (0.6, 0.3, 0.1 m). In the preliminary stage, the surface was polished to remove contamination using an ultrasonic machine. Finally, to reveal the morphology and grain structures of the specimens, a 2% Nital solution was used to etch the polished surfaces. An optical microscopy device with a 100 X magnification factor was used to gather the results. Fig. 1 illustrates the AISI1035 microstructure. A Zwick & Co.'s Z323 hardness tester was used to determine the surface hardness. Standard Test Method ASTM E92 is utilized for hardness test. Under a weight of 0.9 kg, the hardness of the quenched samples was evaluated. Three different hardness measurements were made, and the average value was calculated. The XRD method was used to measure the residual stresses near the surface of the alloy using an XRD 6000 (ORION, model RKS 1500 F-V -SP).

3 Artificial Neural Network (ANN) Model

ANNs can be reformed for huge, complex, and nonlinear databases. Moreover, it can be used for prediction and optimization requirements^[29-31]. ANN is consisting of input layer, hidden layer, and output layer^[32]. Traditionally, ANN is consisting of one input layer, one or more hidden layer based on the ANN performance may modified, output layer is result of ANN model^[33-34]. Activation function is used based on the maximum R² value, and minimum Root Mean Square (RMS) value^[35-36]. ANN architecture is feeding with r, p, d and s parameters as input parameter, weight w, and bias b, linear combiner factor, transfer function f, and output a.

To determine the best outcomes from the experimental attempts, four input subsets of parameters (time, temperature, nanofluid type, and concentration ratio) were used, which was expensive and time-consuming. Since previous researchers have utilized the artificial numerical approach, it might be possible to determine the ideal input parameters to achieve optimal outputs. In this context, Fig. 2 provides a summary of the ANN model employed in this investigation. As one of the tests for each network system, it contains one output, one hidden layer, and four input parameters. The four neurons in the input layer represent the concentration values, the type of nanofluid, the temperature, and the soaking duration, respectively. Ten neurons make up the buried layer. The output layer, however, only contains one neuron because it is accurate in estimating mechanical features.

The training function Levenberg-Marquardt backpropagation(trainFcn='trainlm') with learning rate 0.1, training time 1000 s, learning epoch 10000 epoch, stopping criterion based on MSE criteria goal equal to zero, and transfer function sigmoid, these parameters are adopted to establish ANN model.

For improved training results^[37], the feed-forward background was taken into consideration when developing the network scheme. Only 16 experimental data could be applied to the network for the training goal due to the extremely constrained experimental conditions. The results of the observations of the output neurons from the ANN algorithm included residual stress, grain size, and hardness. The ANN estimation was tested using the same 16 experimental data, and the method's errors were determined.

4 Results 4.1 Quenching Result

The results of the proposed quenching method exhibited a significant enhancement in hardness, a refinement of the grain size, and the removal of unwanted residual stresses, the results are listed in Table 6.

The outcome also indicated the enhancement of the hardness, grain size, and residual stresses of the AISI 1035 steel alloy during analysis of ANOVA. The heat treatment process is descending to refine grain size and becomes harder. Also, it is lead to remove the tension stresses^[38-41]. Phase transformation, especially in the steel, from the austenite to metastable martensite phase takes place. This action is highly recommended because it enhances hardness of steel^[2]. The obtained result is in agreement with previous studies on hardness improvement. Thus, the best conditions for quenching by a nanofluid of water/TiO₂ is by adding 0.2% of TiO₂ nanoparticles and using a holding time of one hour. The grain size was measured based on the result of optical microstructures. The results showed refinements of grain size for most specimens. Fig. 3(a) shows the morphology of the surfaces of the steel AISI 1035 under heat treatment (quenching) in a nanofluid of water/TiO₂. The percentage of elements (C) appears to be higher than those of the other elements, as seen in the EDS for the light region. Because of the portion (C), the residual stress was reduced, and the grain size was large. The high proportion of (C), resulting from reaction of the media(water/TiO₂) during the quenching operation, indicates that the nanofluid is less perfect for producing reasonable hardness and refined grain size. Fig. 3(b) shows the effect of heat treatment in the media (Oil/CuO). The percentage of (C) was still high, but the rate of (Mn) increased, and (Fe) decreased, which caused a high reduction in residual stress and a low grain size; hence, this medium was better than the other. It can be seen in Fig. 3(c) that the percentage of element (Fe) was high; it is suggested that the residual stress was so high that the effect of the quenching media was not enough to spread the heat regularly. The (Fe) in Fig. 3(d) was extremely high. The main change involved the elements (Mn), (Fe), and (C). Numerous studies have been conducted to identify these phase transitions and establish a link between them and the mechanical properties of steel alloys^[42].

The process of quenching results in the production of martensite. The ideal hardness value was introduced when carbon was present during the quenching of oil with alumina oxide. The formation of martensite and the carbon content led to an increase in hardness. Increased carbon content makes an alloy harder, but the martensite proportion must be properly managed. Additionally, manganese has a weaker ability to increase the steel's hardness. The properties of steel alloy are influenced by residual stresses. Exchanging of heat transfer among quenching parameters and steel surface is considering a key source to relief residual stresses^[2]. As a result, eliminating undesired residual stresses from the manufacturing process is one of the main goals of the heat treatment procedure^{[2, 43-44]}. Residual tensile stresses are eliminated, and compressive residual stresses are generated by quenching in copper oxide oil. The slight increase in manganese allowed for the greatest compressive stresses. The mechanical properties of a steel alloy may be influenced by the grain size microstructure. As can be seen from the findings of the experimental study, the refinement of the grain size increases the resistance and hardness of steel. Fig. 4 shows a plot of how the concentration ratio, nanofluid medium, temperature, and soaking duration affect the hardness, grain size, and residual stresses. The outcome showed that for entire cases, agreed with literature studies, the concentration ratio significantly influenced the hardness, residual stresses, and grain size. The type and temperature of the nanofluids had less of an impact than the concentration ratio. On the other hand, the amount of time spent soaking only had an impact on the outcome of hardness. A statistical analysis MANOVA is implanted by MINITAB 19 software to study the significance of the studied factors on hardness, grain size and residual stresses. Three criteria Wilks', Lawley-Hotelling, and Pillai's are used. The main significant factor is nanofluid type at α=0.05 with p value of 0.029 based on Wilks' criteria, 0.01 based on Lawley-Hotelling criteria.

4.2 Prediction Results of ANN Model

In order to make an estimation of hardness, Residual stresses, and Grain size, the ANN training has been realized using 16 different samples, as shown in Table 5. Figs. 5(a)-(c) display how each ANN network is implemented. These charts demonstrate how closely the output of the model matches the actual goal values. The best outcome, where outputs=targets, is shown by the dashed line in each figure. A solid line illustrates the difference between the target and output linear regressions.

The relationship between the target and the output is shown by the R-value. When R = 1, the relationship between the objectives and the outcomes is perfectly linear. There is no linear relationship between targets and outputs when R is close to zero^[29]. The predicted data are listed in Table 7.

4.3 Predicting of Hardness using ANN

The test values for the hardness experiment results are contrasted with the ANN findings displayed in Table 8. The ANN estimation gives the maximal hardness for experiment No. 14, whereas experiment No. 2 gives the lowest one. This demonstrates that the middle hardness levels can be predicted by the ANN more accurately than the lower and higher values. In the upcoming sections, this circumstance will be covered in more detail along with accompanying error graphs. Fig. 6 represents the experimental and ANN results in accordance with all of these. Although the ANN values are generally a little greater than that of the experimental values.

4.4 Predicting of Grain Size using ANN

Similar to the prior section, 16 specimen values from Table 6 were used for the ANN analysis of grain size. The ANN estimation outcomes shown in Table 9 were discovered via the trained algorithm, as shown in Fig. 7. The experiment with soaking time of 30 min and a concentration value of 0.5% using of category one achieved the biggest grain size value, whereas Experiment No.11, with a concentration value of 2.5% using of category three and a soaking time of 60 min, produced the smallest grain size value. These results exhibit the same trend as the experimental findings. In that regard, it is predicted that the error values will be lower than the hardness values. Thus, it can be concluded that this ANN model is most effective for estimating grain size.

4.5 Predicting of Residual Stress Using ANN

Based on the experimental findings in Table 6, the residual stress values were predicted, as shown in Fig. 8. The ANN findings shown in Table 10 were discovered by the trained ANN algorithm. There is a clear parallel between the ANN results and the overall experimental outcomes. However, Experiment No.4 revealed the maximum tensile residual stress, while Experiment No.8 presented the minimal compressive residual stress case.

4.6 Errors of Proposed ANN Model

The ANN results presented in the previous sections have different errors, depending on the material mechanical properties. The absolute error is computed using the formula:

$ \text { Error }=\mid \text { real }- \text { expected } \mid $

(3)

And the percentage error is calculated using the below formula:

$ \text { Error } \%=\frac{\mid \text { real }- \text { expected } \mid}{\text { real }} \times 100 \% $

(4)

However, the maximal percentage error has been found to be around 1.59%, using the formula:

$ \text { Total error } \%=\frac{\sum\limits_{i=1}^N E_i}{N} \times 100 $

(5)

where, E_i and N indicate the ANN absolute percentage errors given in Tables 11, 12, and 13. The number of tests for all mechanical evaluations is N=3×16, respectively.

Since this study uses four separate experimental mechanical tests, the overall variance percentage is acceptable. As a result, every trial contains experimental errors that could have an impact on the ANN algorithm's outcomes.

4.7 Performance Criteria

In order to evaluate the performance of proposed ANN model, R², and RMSE criteria are adopted for assessment purposes^[27]. The experimental and ANN results listed in Table 7 are used in formula (6) to calculate R^{2[29, 45]}.

$ R^2=\frac{\sum\limits_{i=1}^n\left(k_{\exp , i}-K_{\mathrm{EXP}}\right)\left(k_{\mathrm{ANN}, i}-K_{\mathrm{ANN}}\right)}{\sqrt{\sum\limits_{i=1}^n\left(\left(k_{\exp , i}-K_{\mathrm{EXP}}\right)^2\left(k_{\mathrm{ANN}, i}-K_{\mathrm{ANN}}\right)^2\right)}} $

(6)

where k_exp, k_ANN are represent experimental, and predicted ANN results.

K_EXP, K_ANN are evaluated by Eqs.(7) and (8):

$ K_{\mathrm{EXP}}=\frac{1}{n} \sum\limits_{i=1}^n k_{\mathrm{EXP}, i} $

(7)

$ K_{\mathrm{ANN}}=\frac{1}{n} \sum\limits_{i=1}^n k_{\mathrm{ANN}, i} $

(8)

$ \mathrm{RMSE}=\sqrt{\frac{\sum\limits_{i=1}^n\left(k_{\exp , i}-k_{\mathrm{ANN}, i}\right)^2}{n}} $

(9)

The results of Table 7 are used to calculate RMSE value for the result of experimental and expected result of ANN model. The minimum value of RMSE is achieved by the predicted result of grain size which is 0.025.

5 Conclusions

The effects of heat treatment on the mechanical properties of AISI 1035 were examined in this study. To enhance the mechanical properties of a medium carbon steel alloy, various quenching mediums of nanofluid are proposed. Also, we suggest a model for predicting the mechanical properties of AISI 1035 during the quenching process using an artificial neural network (ANN). The following conclusions can be drawn from the obtained results:

1) Quenching AISI 1035 using nanofluids is promising but requires the careful adjustment of the volume concentration of the nanoparticles.

2) The highest compressive residual stress of 239 MPa was attained under the following conditions: oil/CuO, a CuO volume fraction of 5%, a temperature of 950 ℃, and a soaking period of 30 min.

3) The maximum hardness was 679.6 HV, which was achieved under conditions of water/hybrid nanoparticles(75% CuO + 25% TiO₂), a temperature of 950 ℃, and a soaking time of 60 min.

4) The refinement of grain size was 14.854 μm, which was achieved at a concentration ratio of nanoparticles of 2.5 vf %, engine oil/Al₂O₃, a temperature of 850 ℃, and a soaking time of 60 min.

5) The outcomes showed that all conditions led to a noticeable improvement in the alloy's hardness which reached 100%, the grain size is refined about 80%, and unwanted residual stresses is removed from 50 to 150 MPa

6) Statistical analysis of MANOVA showed that the nanofluid type had significant effect on the responses of hardness, grain size, and residual stresses. P values are 0.029, and 0.01 for Wilks', and Lawley-Hotelling criteria respectively.

7) Performance of ANN model is evaluated based on R² value which approached to 1. The value indicates a good performance for the proposed model.

8) For the parameters of residual stresses, hardness, and particle size, the provided ANN model fits well with the experimental findings.

9) The proposed ANN model has a better agreement with the expected results of residual stresses and hardness values than it does with the projected outcome of grain size.

10) Percentage error, R², and RMSE indicate an acceptable impression for proposed ANN model.

Acknowledgment

The authors would like to thankful Kut Technical Institute for their funding supports.