Prediction of surface roughness of tempered steel AISI 1060 under effective cooling using super learner machine learning.

doi:10.21203/rs.3.rs-5189811/v1

Surface roughness is essential to evaluate the quality of the surface of the product. To predict the surface roughness researchers have been using statistical and empirical methodologies both of which lack generalizability when applied to unseen data. To overcome the limitation of existing models’ scholars have developed machine learning and artificial intelligence. Machine learning can predict the surface roughness of machined parts accurately. It has high generalization ability when applied to unseen data. For instance, this research endeavor has formulated a super learner machine learning model aimed at predicting surface roughness by leveraging a diverse array of machine learning techniques, including decision trees, random forests, gradient boosting, and extreme gradient boosting. The optimization of these models was achieved through the application of grid search hyperparameter tuning and K-fold cross-validation methodologies. The predictive efficacy of the proposed super learner model is compared with that of all alternative models. Achieving a coefficient of determination (R²) of 99.8% between the experimental and predicted values for surface roughness in the test dataset, the suggested super learner model exhibited superior predictive capabilities relative to its counterparts. This model is identified as the most accurate, distinguished by the highest coefficient of determination (R²), the lowest mean absolute error (1.92%), the lowest mean absolute percentage error (1.76%), and the lowest root mean square error (2.29%). In addition, the interpretations of the model's predictions are clarified using the Shapley additive explanations (SHAP) technique, thereby shedding light on the significant variables that affect the surface roughness of tempered steel AISI 1060.

Decision tree

random forest

adaptive boosting

gradient boosting

extreme gradient boosting

At present, hardened steels are widely employed across a myriad of engineering applications, including the production of railway components, gears, bearings, forging and extrusion dies, punches, shafts, and nozzles. Traditionally, hard products are manufactured through a series of processes, including forming, annealing, cutting, heat treatment, and surface finishing. Notably, the implementation of hard turning for the fabrication of various hard components eliminates the necessity for annealing and external finishing from the conventional operational sequence [1]. Consequently, the reduction in the number of operational stages contributes to the attainment of sustainable production, as numerous resources are conserved. Nevertheless, the machining of hardened materials inherently induces elevated temperatures [2], which adversely affects machinability and, in turn, deteriorates the surface roughness of the machined components.

A plenty of investigations have underscored the critical role of surface roughness in ascertaining the overall quality of a product. The arithmetic average of the surface roughness (Ra), articulated in absolute terms, provides a standard metric for characterizing surface roughness over a designated length [3]. Additionally, the hardness and composition of the machined materials exert an influence on the resulting surface roughness. A considerable reduction in resource consumption (including human labor, machinery, time, and material) would be achievable if the surface roughness of machined components could be accurately predicted based on key influencing factors.

Through the application of various traditional and artificial intelligence-based methodologies, researchers have developed predictive models for surface roughness. For instance, in the context of hard turning, Hessainia et al. [4] employed Response Surface Methodology (RSM) to anticipate surface roughness. The parameters manipulated in the study included cutting speed, feed rate, cut depth, and tool vibrations. The machining experiments were conducted under dry conditions. The efficacy of the predictive model was substantiated through RSM, ANOVA, surface plots, and regression plots. A regression coefficient of 99.9% was determined for the surface roughness. Furthermore, the optimal surface roughness was achieved at the minimum feed rate, maximum cutting speed, and the least depth of cut. In the hard turning process of AISI 52100 steels, Bouacha et al. [5] conducted a statistical analysis of surface roughness and cutting force. The cutting tool utilized was Cubic Boron Nitride (CBN), and the machining was performed in a dry environment. The statistical analysis was facilitated by RSM, complemented by ANOVA. ANOVA was employed to evaluate the influence of the factors involved, while the desirability function was utilized to optimize the response variables. The feed rate and cutting speed were found to exert significant effects on surface roughness. The simultaneous optimization of surface roughness and cutting force resulted in final parameters of 246 m/min for cutting speed, 0.08 mm/rev for feed rate, and 0.15 mm for depth of cut.

To make the machining system financially viable, process enhancement is required prior to machining in addition to response prediction. Mia et al. [3] employed both single- and multiple-objective optimization techniques while turning Ti alloys. The process of optimization showed that compared to all other random experimental runs, the optimal experimental run utilizes a significant number of high-quality resources less. Artificial intelligence-based techniques for performance index optimization include genetic algorithms, which can deal with problems including nonlinear complex interactions between the response and inputs. It is clear that Bouacha et al. [6] built a prediction and optimization model of the cutting force, tool wear, volume of metal removed, and surface roughness using multi-objective optimization techniques like the genetic algorithm, the Gray-Taguchi method, and the composite desirability function. The optimization model generated from evolutionary algorithms works better than other methods because it can handle complex and nonlinear interactions and deal with both discrete and continuous variables. A genetic algorithm was also employed by Rabiei et al. [7] to lower surface roughness and establish the parameter values for this surface finish during the grinding procedure. The studies were conducted using traditional dry flood cooling and minimal quantity lubricant. The second-order surface roughness model, which functioned as the GA objective function, was constructed using RSM. Support vector machines (SVM) and support vector regression (SVR) have been used recently to predict quality aspects in multivariate domains. A support vector machine (SVM) was used to simulate surface roughness, and its performance was compared to models made with artificial neural networks (ANNs) [8]. Using a genetic algorithm (GA) with an ANN and SVR, Gupta et al. [9] optimized the parameters while turning. Additionally, they contrasted the results with those of techniques based on non-artificial intelligence, like response surface methods (RSMs) and regression. The variables were cutting speed, feed rate, and machining time; the quality features were surface finish, tool wear, and power consumption. The best results have been obtained by combining artificial intelligence-based methods, such ANN and SVR, with genetic algorithm-based optimization.

While numerous publications exist regarding the substantial application of both conventional and artificial intelligence in surface roughness prediction, the majority of the models now in use are empirically based on small-scale experimental data, which frequently restricts the models' accuracy. It is therefore advised to choose a different and more potent modeling approach.

Machine learning (ML) algorithms have the ability to predict the correlations between the input parameters and the response(s) without requiring prior assumptions of the underlying mathematical and physical models, which sets them apart from most empirical approaches [10]. A wide range of engineering problems have been successfully modeled in recent years using ML-based models [11–19]. To this cause, B. Rolf et al. [20] employed a genetic algorithm (GA) to solve a hybrid flow scheduling program in the manufacturing sector and found that using a GA to assign dispatching rules yielded better results than using traditional industrial dispatching techniques was effective. In the context of construction and building materials, M.S. Ahmed et al. [21] projected the splice strength of an unconfined beam sample using support vector regression. O. Addin et al. [22] used the naive Bayes classifier to detect deterioration in engineering materials, emphasizing the importance of machine learning in material and design.

J. Peters et al. [23] modeled the ecohydrological distribution in ecological modeling using the random forest approach. K. L. McFadden [24] employed logistic regression in the field of ergonomics to forecast the probability of an error being made by US airline pilots. L. Zhou et al. [25] used the KNN machine learning technique in the field of computers and security. Machine learning techniques have great promise for use in nuclear engineering, as demonstrated by B.P. Dubey et al.'s [26] use of ANNs to approximate power distribution in pressurized heavy water reactors (PHWRs) in a timely and precise manner. [27] employed ensemble machine learning to ascertain the material and design of rectangular RC columns' plastic hinge length.

Similar to this, numerous machine learning techniques have been applied in machining [28–33] to address difficult situations. The adoption of more effective machine learning techniques, such ensemble learning, has not gotten enough attention in the past, despite the fact that machine learning is extensively employed in the manufacturing and construction sectors.

For instance, the goal of this research is to develop and apply superior machine learning for the purpose of predicting the surface roughness of tempered steel (AISI 1060) under effective cooling. Ensemble learners are robust nonparametric algorithms capable of handling complex relationships with a great deal of interaction that are highly nonlinear. Ensemble models combine various machine learning methods into a single predictive model in order to improve predictions and decrease bias and variance errors. Bootstrap aggregating, or "bagging," is a technique that uses highly randomized trees and random forests. Adaptive, gradient, and strong gradient boosting are examples of boosting. Another method of using ensemble learning is called "stacking," which combines several base learners using a meta-model and aggregates their output to get a final prediction that is more accurate than a single model. [34, 35]

As per Wolpert [34], stacking can be seen as an advanced form of cross-validation. Regression problems have seen successful use of it [36]. The literature does not provide machine learning applications for evaluating the surface roughness of tempered AISI 1060 steel, despite the encouraging outcomes of applying numerous algorithms in this area. Moreover, while machine learning models are typically regarded as "black boxes," it is essential to make sure an ML model can be explained in order to verify its results. Lundberg and Lee [37] established a uniform Shapley additive explanations (SHAP) technique to explain the output of any machine learning model. Only a few research have examined the interpretability of the machine learning (ML)-based models employed in the field of manufacturing engineering [38–40]. To the best of the author's knowledge, no work has documented how ML approaches can evaluate tempered AISI 1060 steel's surface roughness.

Hence, the goal of this work is to establish super learner machine learning for prediction of surface roughness of tempered steel AISI 1060 under effective cooling. To overcome the shortcomings of previous efforts, tree based and boosting ensemble models trained on a dataset have been combined to create a novel super learner machine learning model. The inability of ML models to be explained is a major drawback of current research. A unified Shapley additive explanations (SHAP) technique has been used to explain the machine learning mode's output in order to fill this gap.

The work's novelty

The buildup of tree based and ensemble machine learning models and their comparison
Combining the GB, XGB, and DT models to develop a super learner model
Using the unified SHAP technique; learning mode, examine the ML model's output and rank the input features and their interactions that affect the surface roughness of tempered steel 1060.

This research developed a fast, accurate, and intelligent framework for predicting the surface roughness of tempered steel AISI 1060 by utilizing the potential of the super learner machine learning model.

2.1 Research data

A database used in this investigation obtained from [41]. Table 1 lists the cutting speed (Vc), feed rate (F), workpiece hardness (H), machining environment (ME), and surface roughness (Ra) as well as the statistical information pertaining to each of these variables. Table 1 lists the ranges in which the input variables fall: 58–161 mm/min for Vc, 0.1–0.16 mm/rev for F, and 40–56 HRC for H, dry–HPC for ME. A training set and a test set, comprising 80% and 20% of the entire dataset, respectively, are then created by randomly dividing the data.

Table 1

Description of the data used in this research
parameter	Mean	STD	Min	Max	Q1	Q2	Q3
Vc	103.73	39.981	58	161	63.75	98	149.5
F	0.13	0.022	0.1	0.16	0.105	0.13	0.155
H	48	6.566	40	56	40	48	56
Ra	1.289	0.419	0.62	2.35	0.897	1.11	1.768
Cooling	Dry, HPC
Min = minimum, Max = maximum, STD = standard deviation, Q1, Q2, Q3: 25th, 50^th, and 75th percentiles

2.2 Overview of machine learning models

2.2.1 Super learner model

To build a single, more powerful model, the super learner model approach combines multiple machine learning models, or "base learners" [42]. Figure 1 depicts the super learner model that is proposed. Using K-fold (K = 10 in this case) cross-validation, the best combination of base learners is found by giving each base learner a distinct weight, as decided by the algorithm. The super learner model method was applied in this investigation (Fig. 1) using the subsequent steps:

The training data must first be divided into 10-folds. The second phase involves choosing a group of strong models to serve as candidate base learners. These models are then individually optimized before being used in the super learner architecture. The model is improved by fitting it to the complete training dataset in the third step, which involves evaluating each candidate base learner's performance using 10-fold cross-validation and creating a matrix of predictions during the validation process. Lastly, a meta-model that predicts the values of the target variables is fitted using the prediction matrix as input. In this study, the potential base models for the super learner method were decision tree (DT), gradient boost (GB), and extreme gradient boost (XGB), with linear regression (LR) acting as the meta-model, as given in Fig. 1.

2.2.2 Decision Tree

Among type of supervised machine learning method that resembles a flow chart in structure is the decision tree (DT) algorithm [43]. The interior nodes define the attribute tests, whereas the top node in the tree, referred to as the root node, stands for the most important node. Predictions are made using the leaf nodes at the base of the tree, which contain the value of the target variable.

The algorithm creates the tree by recursively applying a series of predetermined rules to divide the N-dimensional input space R^N into J disjoint regions {R1,... Rj}, where each area has the same prediction for all inputs within it. As a function h(x), the tree can be represented as follows using Eq. (1):

$$\:\text{h}\left(x\right)={\sum\:}_{j=0}^{J}{b}_{j}{I}_{\left(x{\epsilon\:}{R}_{j}\right)}$$

1

where $\:{b}_{j}$denotes the prediction value inside region $\:{R}_{j}$ and the disjoint region$\:{R}_{j}$ is intended for each leaf of the tree. When $\:x{\epsilon\:}{R}_{j}$, the indicator function $\:{I}_{\left(x{\epsilon\:}{R}_{j}\right)}$equals one.

2.2.3 Ensemble models: RF, GB, and XGB

Among ensemble model, the boosting ensemble is a highly popular and effective type of ensemble model, where the base learners, especially DTs, are arranged sequentially to form a strong final model [44].

a) Random forest

A collection of several randomly generated decision tree predictors is called a random forest (RF). In the random forest technique, bootstrap samples randomly chosen samples from the initial training dataset with replacement are used by each decision tree predictor. Furthermore, when dividing nodes in the decision tree based on the optimal split among a random subset of the characteristics chosen at each node, random subsets of input features are taken into account [45].

There are two processes involved in splitting at each node. Initially, from the bootstrap sample, a random subset of input features is chosen [45]. At each decision tree node, the optimal subset feature is then chosen to carry out the decision split [45. The average of the forecasts from each decision tree predictor determines the final prediction of the RF. [46]. See Fig. 2.

b) Gradient boosting

Gradient boosting (GB), builds a strong model by combining several base learners. To get beyond the drawbacks of earlier models, the GB, uses a gradient descent method. Eq. (2) [48] illustrates how the GB algorithm initializes the model with a constant value:

$$\:{F}_{o}\left(\text{x}\right)=\frac{\text{a}\text{r}\text{g}\:\text{m}\text{i}\text{n}}{\gamma\:}{\sum\:}_{i=1}^{N}{L(Y}_{i,})$$

2

In the t^th step, the GB updates the model according to Eq. (4), fitting the base learner to$\:\:{h}_{t}\left(X\right),\:\:$ the pseudo residual (negative gradient descent) $\:{r}_{i,t}$of the preceding learner, Eq. (3).

$\:{r}_{i,t}=-{\left[\frac{\partial\:L\left({Y}_{i\:},F\left({X}_{i}\right)\right)}{\partial\:F\left({X}_{i}\right)}\:\right]}_{F\left(X\right)={F}_{t-1\left(X\right)}}$ , for i = 1,……,N (3)

$$\:{F}_{t}\left(X\right)={F}_{t-1}\left(X\right)+{\gamma\:}_{t}{h}_{t}\left(X\right),\:\:fort=1,\dots\:\dots\:\dots\:,T$$

4

where T is the total number of base learners, and Eq. (5) illustrates how to find the best solution to a single variable optimization to get the multiplier $\:{\gamma\:}_{t}$.

$\:{\gamma\:}_{t}={arg}_{\gamma\:}\:min{\sum\:}_{i=1}^{N}L({Y}_{i\:},{F}_{t-1}\left({X}_{i}\right)\:+\:\gamma\:{h}_{t}({X}_{i}$ )) (5)

where $\:\gamma\:$ ϵ [0, 1] is the learning rate [48].

c) Extreme gradient boosting

A variation of the boosting model called extreme gradient boosting (XGB) was created to offer a more effective way to execute GB [49]. The inclusion of a regularization component in the objective functions of XGB is one of its primary characteristics; this reduces the complexity of the model and helps to mitigate overfitting [49]. This is shown by Eq. (6). By adding this regularization term, XGB is able to produce models that are more reliable, more generalizable, and more predictive [49].

Obj(ϴ) = $\:{\sum\:}_{i=1}^{N}L({Y}_{i\:},\:\:{Ŷ}_{i})\:+{\sum\:}_{t=1}^{T}\varOmega\:({ƒ}_{t\:})\:$ (6)

where $\:\varOmega\:$(⋅) is the regularization term indicated in Eq. (7) and L (⋅) is the training loss between the predicted (Ŷ) and actual (Y) values of the response:

$\:\varOmega\:\left({ƒ}_{t\:}\right)$ = γ M + $\:\frac{1}{2}$λ$\:{‖{\omega\:}_{t}‖}^{2}$ (7)

where M is the number of leave nodes in the tree, $\:{\omega\:}_{t}$is the leaf weights, and γ and λ are the penalty coefficients.

2.4 Tuning hyperparameters and cross-validation

The predictive power and generalization capacity of a particular model are determined by the values of its hyperparameters. Hyperparameter optimization or tuning helps determine the optimal settings for the hyperparameters. Hyperparameter optimizing is commonly achieved through the use of grid search. A K-fold cross-validation technique is used in the hyperparameter tuning procedure to avoid overfitting issues. Initially, 80% and 20% of the final dataset are divided into training and testing datasets, respectively, from the original dataset. Next, the following three steps are used to carry out K-fold cross-validation:

a) To ensure that every observation is utilized exactly once for training and validation, the training dataset is divided into independent K groups or folds of equal size without replacement.

b) The model is fit using K − 1 folds, and the remaining one-fold is used for validation.

c) Step (b) should be repeated K times to achieve K performance indices.

The mean of K performance indices is used to determine the model's final performance. In this work, the hyperparameters are optimized by combining a grid search technique with a 10-fold cross-validation (K = 10). 90% of the training dataset is used to train the model in 10-fold cross-validation, with 10% of the training dataset being used as a validation dataset in each iteration. Refer to Fig. 3. [50].

2.5 ML performance measures

Four different statistical measures are used to assess the correctness of the ML models: mean absolute percentage error (MAPE), coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error (MAE). The following is the mathematical expression for these measures (see Eqs. (13) through (16)).:

$$\:MAPE=\frac{1}{N}{\sum\:}_{i=1}^{N}\left|\frac{{Y}_{i-}{Ŷ}_{i}}{{Y}_{i}}\right|$$

8

$$\:RMSE=\sqrt{\frac{1}{N}{\sum\:}_{i=1}^{N}{({Y}_{i}-{Ŷ}_{i})}^{2}}$$

9

$$\:MAE=\frac{1}{N}{\sum\:}_{i=1}^{N}\left|{Y}_{i-}{Ŷ}_{i}\right|$$

10

$$\:{R}^{2}=1-\frac{{\sum\:}_{i=1}^{N}({Y}_{i}-{Ŷ}_{i})}{{\sum\:}_{i=1}^{N}({Y}_{i}-{\stackrel{-}{Y}}_{i})}$$

11

where N is the number of data points, Ȳ is the mean of the Y values, and Y and Ŷ are the target and predicted values, respectively.

3.1 Hyperparameter optimization results

This section looks into the developed machine learning models' ability to predict the responses. It also evaluates the model's performance using a variety of performance indices, such as coefficient of determination (R²), mean absolute error (MAPE), root mean squared error (RMSE), and mean absolute percentage error (MAPE). Using a grid search and 10-fold cross-validation, the hyperparameters of the DT, RF, GB, and XGB algorithms are optimized and validated.

Table 2 provides a summary of the findings. The optimum maximum depth of the tree for the DT model was determined to be 40. The optimal maximum depth of the tree for the RF model was 10, the optimal minimum sample leaf was 1, the optimal minimum sample split was 2, and the optimal number of estimators was 100. The optimal number of estimators for the GB was 100, with a learning rate of 0.1 and a maximum tree depth of 1. The optimal number of estimators for XGB was 100, with a maximum depth of tree of 10 and a learning rate of 0.1 The optimized DT, GB, and XGB are integrated by the super learner model using linear regression (LR) as a meta-model.

Table 2

Optimized hyperparameters
Model	Hyperparameter	Optimal value
Decision tree	Maximum depth of tree	40
Random forest	Number of estimators Maximum depth of the tree Minimum samples leaf Minimum samples split	100 10 1 2
Gradient boosting	Number of estimators Maximum depth of tree Learning rate	100 1 0.1
Extreme gradient boosting	Number of estimators Maximum depth of the tree Learning rate	100 10 0.1

3.2 Model performance

Scatter plots of the experimental surface roughness versus the predicted surface roughness using the DT, GB, XGB, and super learner models, respectively, are presented in Figs. 4a–4e. The results show that all the models have a high degree of surface roughness prediction accuracy on both the test (R²$\:\ge\:$96.9%) and training (R²$\:\ge\:$98.5%) datasets. This highlights the benefits of employing ensemble and sophisticated machine learning models to precisely forecast the surface roughness (Ra) of tempered steel AISI 1060. As shown in Figs. 4a–4e, the predicted surface roughness for each model is concentrated around the 45-degree diagonal line, indicating a precise correlation between the predicted surface roughness and the corresponding experimental values.

This observation demonstrated how well the ensemble models that were suggested work to forecast the surface roughness of tempered steel (AISI 1060). With the highest R² of 100% for the training dataset, the DT and XGB outperformed the other models in terms of prediction ability when compared to RF, and GB, as seen in Figs. 4a–4e. On the test dataset, however, R² of DT dropped to 96.9% and R² of XGB is 99.7%. RF has R² of 99.7% for training and 98.6% for test dataset GB has R² of 98.5% on the training dataset and R² of 99.0% on test dataset.

By using a combination of DT, GB, and XGB, to optimize the predictions, the super learner model further improves the models' predictive power. R² scores for this model's training and test datasets were 99.7% and 99.8%, respectively. When comparing generalizability of the proposed models, super learner (99.8%) > XGB (99.7%) > GB (99.0%) > RF (98.6%) > DT (96.9%) see Fig. 4a-4e.

Table 3

Machine learning performance
Training					Test
	RMSE (%)	MAE (%)	MAPE (%)	R² (%)	RMSE (%)	MAE (%)	MAPE (%)	R² (%)
DT	0.00	0.00	0.00	100.00	8.72	8.30	7.15	96.90
RF	2.65	2.12	1.72	99.70	5.85	4.80	4.18	98.60
GB	5.90	4.50	3.71	98.50	5.03	4.25	3.84	99.00
XGB	0.14	0.00	0.09	100.00	2.80	5.98	2.26	99.70
Super learner	2.77	2.30	1.82	99.70	2.29	1.92	1.76	99.8

3.3 Comparison of predictive models

Four key performance metrics were used to thoroughly assess the effectiveness of the optimized machine learning models: the MAE, RMSE, MAPE, and R². The statistical outcomes of the ML models' performance are shown in Table 3. Additionally, the performance metrics are presented in Figs. 5 and 6 for the training and test datasets, respectively, to further illustrate the effectiveness of the generated models.

when looking at the DT it achieved R² of 100% for the training dataset and 96.9% for test dataset respectively. the RSME and MAPE for the training datasets were 0.00% and 0.00%, respectively. But for test dataset, it has RMSE and MAPE of 8.72% and 7.15% respectively. The DT model suffers from overfitting and is unable to generalize effectively to new data, while having a highest R² of 100% in the training datasets. On the test datasets, however, it has a lowest R² of 96.9% and a higher RMSE and MAPE. Among all proposed model the DT exabit the least prediction capability on new or unseen data.

for the RF model, it has R² value of 99.7% for training dataset and 98.6% for test dataset. On the test dataset, The R² of RF is greater than that of DT. The RF model has the RMSE of 2.65% and MAPE of 1.72% for the training dataset. But for test dataset, it has RMSE and MAPE of 5.85% and 4.18% respectively. For test dataset, the RMSE and MAPE of RF is less than that of DT. This finding shows that, in comparison to DT, RF exhibits a higher capacity for generalization.

On the training datasets, GB obtained an R² of 98.5%, whereas on the test datasets, it achieved an R² of 99.0%. The GB has the lowest R² of 98.5% on the training datasets of all the models examined, but it has a high R² of 99.0% on the test dataset. When comparing with DT and RF, it has greatest R² on test dataset indicating that it has the greatest generalization capacity. The GB model has the RMSE of 5.90% and MAPE of 3.71% for the training dataset. But for test dataset, it has RMSE and MAPE of 5.03% and 4.25% respectively. For test dataset, the RMSE and MAPE of GB is less than that of DT and RF. This finding shows that, in comparison to DT and RF, GB exhibits a higher for generalization capability.

With the highest R² value of 100% and 99.7% for training and test dataset respectively, the XGB is the most successful at handling new data because it has a highest generalization capacity when comparing with DT, RF, and GB. The XGB model has the RMSE of 0.14% and MAPE of 0.09% for the training dataset. But for test dataset, it has RMSE and MAPE of 2.80% and 2.26% respectively. XGB has higher R² and lower RMSE and MAPE when comparing with GB because of inclusion of a regularization component in the target functions of XGB. This, reduces the complexity of the model and helps to mitigate overfitting [49]. By adding this regularization term, XGB is able to produce models that are more reliable, more generalizable, and more predictive.

Finally, the super learner model exhibits significant improvements in performance, achieving the lowest errors (MAPE and RMSE) and the highest R² on both test datasets. For training datasets, the super learner model's R² is 99.7%, and for test datasets, it is 99.8%. The super learner model has the RMSE of 2.77% and MAPE of 1.82% for the training dataset. But for test dataset, it has RMSE and MAPE of 1.92% and 1.76% respectively This suggests that the super learner model is more successful at handling new data because it has a higher generalization capacity when comparing with XGB.

3.4 Model explainability via the SHAP approach

The SHAP technique provides an explanation of the Ra regression process for every observation in the database. By dividing each prediction into the total effects of all the input features, the SHAP value enables an explanation of each prediction. The model that performed the best in this study's predictions was explained by a tree-based SHAP. A typical Ra prediction plot based on SHAP values is displayed in Fig. 7.

The mean of the experimental findings is shown by the base value in this figure, while the length of the bar indicates the SHAP value for each input parameter, indicating how important the parameter is in predicting surface roughness. Therefore, the explanation function's output is the model's prediction. Additionally, the direction (positive or negative) of the input factors' impact on the projected Ra is depicted in the figure.

Red bars indicate elements that have a tendency to raise the response, while blue bars indicate factors that have a tendency to decrease the response. Hardness has the biggest impact, as seen in Fig. 7, followed by feed, speed and coolant. In Fig. 7, the length and color of the bar correspond to the significance level and the direction (positive or negative) of each factor's effect.

The input components are ranked according to relevance in the summary plot of the SHAP values, which is displayed in Fig. 8. As indicated in the figure H has highest impact followed by F, Vc, and C respectively. Furthermore, as seen in Fig. 9, high levels of H and F increase surface roughness, while high level of Vc and C, reduce surface roughness.

Moreover, the significant interaction between the input elements can be found using the SHAP value. The most pertinent interaction between H and Vc is shown by the SHAP dependent graphs in Fig. 10a. For instance, Fig. 10a. illustrates a significant number of interactions between H and Vc. H and Vc are not interact linearly with one another. Figure 10b. shows the most pertinent interaction between H and F. as shown by the SHAP dependent graphs in Fig. 10b. For instance, Fig. 10ab. illustrates a significant number of interactions between H and F. H and F are interact linearly with one another.

Predicting the surface roughness of tempered steel AISI 1060 by conventional and artificial intelligence approaches has been the subject of extensive research; nevertheless, there aren't many advanced and ensemble ML studies obtainable. This research aims to forecast the surface roughness of tempered steel AISI 1060 for the first time by utilizing several machine learning methods. Advanced and ensemble machine learning models, including DT, GB, and XGB, are optimized and assessed before being combined into a super learner to predict surface roughness.

In order to rank the input features and their interactions for surface roughness prediction, and to explain the expected response, the application of a unified SHAP technique is explored. From this investigation, the following findings can be made:

The developed decision tree and boosting-based ML ensemble models demonstrate a high level of accuracy in predicting the surface roughness of tempered steel AISI 1060.
Among the models evaluated, the super learner, and the XGB model emerged as the most accurate model, closely followed by the GB, RF, and DT models,
The super learner model performed better than any of the other models in terms of generalizability and prediction accuracy. With the lowest mean absolute error (1.92%), lowest mean absolute percentage error (1.76%), lowest root mean square error (2.29%), and highest coefficient of determination (99.8) for the test dataset, this model is the most accurate.
The SHAP results show that the hardness of the workpiece has the greatest impact on the surface roughness, followed by feed rate, cutting speed, and coolants respectively.

ADB Adaptive boosting

AISI American Institute of Steel and Iron

ANN Artificial Neural Network

ANOVA Analysis of Variance

DT Decision trees

GB Gradient boosting

HPC High pressure coolant

RF Random Forest

RSM Response surface methodology

XGB Extreme gradient boosting

Funding: Theauthors declare no funding for this research

Competing interest: The authors declare that they have no competing interest

Code availability: the code used for the current study are available from the correspondent author upon reasonable request.

Authors contribution:

First author: data collection, analysis of data, proposing first draft,

Second author: data collection, analysis of data, proposing first draft

Third author: data collection, analysis of data, proposing first draft

Forth author: data collection, analysis of data, proposing first draft

Grzesik W. Mechanics of Cutting and Chip Formation.
Mia M, Dhar NR. Optimization of surface roughness and cutting temperature in high-pressure coolant-assisted hard turning using Taguchi method. Int J Adv Manuf Technol [Internet]. 2016; Available from: http://dx.doi.org/10.1007/s00170-016-8810-2
Mia M, Khan A, Dhar NR. Study of surface roughness and cutting forces using ANN , RSM , and ANOVA in turning of Ti-6Al-4 V under cryogenic jets applied at flank and rake faces of coated WC tool. 2017;
Hessainia Z, Belbah A, Athmane M, Mabrouki T, Rigal J françois. On the prediction of surface roughness in the hard turning based on cutting parameters and tool vibrations. MEASUREMENT [Internet]. 2013;46(5):1671–81. Available from: http://dx.doi.org/10.1016/j.measurement.2012.12.016
Bouacha K, Athmane M, Mabrouki T, Rigal J françois. Int . Journal of Refractory Metals & Hard Materials Statistical analysis of surface roughness and cutting forces using response surface methodology in hard turning of AISI 52100 bearing steel with CBN tool. Int J Refract Met Hard Mater [Internet]. 2010;28(3):349–61. Available from: http://dx.doi.org/10.1016/j.ijrmhm.2009.11.011
Bouacha K, Athmane M, Khamel S, Belhadi S. Int . Journal of Refractory Metals and Hard Materials Analysis and optimization of hard turning operation using cubic boron nitride tool. RMHM [Internet]. 2014;45:160–78. Available from: http://dx.doi.org/10.1016/j.ijrmhm.2014.04.014
Rabiei F, Rahimi AR, Hadad MJ, Ashra M. Performance improvement of minimum quantity lubrication (MQL) technique in surface grinding by modeling and optimization. 2014;1–14.
Ekici S. Support vector machines models for surface roughness prediction in CNC turning of AISI 304 austenitic stainless steel. 2012;639–50.
Gupta AK, Guntuku SC. Optimization of turning parameters by integrating genetic algorithm with support vector regression and artificial neural networks. 2014;
Flood I. Toward the next generation of artificial neural networks for civil engineering. 2008;22:4–14.
Maheshwera U, Paturi R, Cheruku S. Materials Today : Proceedings Application and performance of machine learning techniques in manufacturing sector from the past two decades : A review. Mater Today, Proc [Internet]. 2020;(xxxx). Available from: https://doi.org/10.1016/j.matpr.2020.07.209
Bappy A. American Journal of Science and Learning for Development Exploring the Integration of Informed Machine Learning in Engineering Applications : A Comprehensive Review. 2024;3(2):11–21.
Sarker IH. Machine Learning : Algorithms , Real ‑ World Applications and Research Directions. SN Comput Sci [Internet]. 2021;2(3):1–21. Available from: https://doi.org/10.1007/s42979-021-00592-x
Alajmi MS, Almeshal AM. Prediction and Optimization of Surface Roughness in a Turning Process Using the ANFIS-QPSO Method. 2020;
Magalhães FC, Mg BH, Ventura CEH, Washington R, Sp SC, Abrão AM, et al. Prediction of surface residual stress and hardness induced by ball burnishing through neural networks Berend Denkena , Bernd Breidenstein and. 2019;14(3):295–310.
Arockiasamy FS, Suyambulingam I, Jenish I. A Comprehensive Review of Real-time Monitoring and Predictive Maintenance Techniques : Revolutionizing Natural Fiber Composite Materials Maintenance with IoT. 2023;31:87–110.
Sami W, Abd M. A new fine-tuned random vector functional link model using Hunger games search optimizer for modeling friction stir welding process of polymeric materials. J Mater Res Technol [Internet]. 2021;14:1482–93. Available from: https://doi.org/10.1016/j.jmrt.2021.07.031
Abd M, Senthilraja S, Zayed ME, Elsheikh AH, Mostafa RR, Lu S. A new random vector functional link integrated with mayfly optimization algorithm for performance prediction of solar photovoltaic thermal collector combined with electrolytic hydrogen production system. Appl Therm Eng [Internet]. 2021;193(March):117055. Available from: https://doi.org/10.1016/j.applthermaleng.2021.117055
Meanti G, Carratino L, Vito E De, Rosasco L. Efficient Hyperparameter Tuning for Large Scale Kernel Ridge Regression. 2022;1–24.
Rolf B, Reggelin T, Nahhas A, Lang S, Müller M. Assigning dispatching rules using a genetic algorithm to solve a hybrid flow shop scheduling problem. Procedia Manuf [Internet]. 2020;42:442–9. Available from: https://www.sciencedirect.com/science/article/pii/S2351978920306107
Ahmad MS, Adnan SM, Zaidi S, Bhargava P. A novel support vector regression (SVR) model for the prediction of splice strength of the unconfined beam specimens. Constr Build Mater [Internet]. 2020;248:118475. Available from: https://www.sciencedirect.com/science/article/pii/S0950061820304803
Addin O, Sapuan SM, Mahdi E, Othman M. A Naïve-Bayes classifier for damage detection in engineering materials. Mater Des [Internet]. 2007;28(8):2379–86. Available from: https://www.sciencedirect.com/science/article/pii/S0261306906002299
Peters J, Baets B De, Verhoest NEC, Samson R, Degroeve S, Becker P De, et al. Random forests as a tool for ecohydrological distribution modeling. Ecol Model [Internet]. 2007;207(2):304–18. Available from: https://www.sciencedirect.com/science/article/pii/S0304380007002931
McFadden KL. Predicting pilot-error incidents of US airline pilots using logistic regression. Appl Ergon [Internet]. 1997;28(3):209–12. Available from: https://www.sciencedirect.com/science/article/pii/S0003687096000622
Zhou L, Zhu Y, Castiglione A. Efficient k-NN query over encrypted data in cloud with limited key-disclosure and offline data owner. Comput Secur [Internet]. 2017;69:84–96. Available from: https://www.sciencedirect.com/science/article/pii/S0167404816301663
Dubey BP, Jagannathan V, Kataria SK. Quick and reliable estimation of power distribution in a PHWR by ANN. Ann Nucl Energy [Internet]. 1998;25(8):567–79. Available from: https://www.sciencedirect.com/science/article/pii/S0306454997001060
Wakjira TG, Alam MS, Ebead U. Plastic hinge length of rectangular RC columns using ensemble machine learning model. Eng Struct [Internet]. 2021;244(March):112808. Available from: https://doi.org/10.1016/j.engstruct.2021.112808
Aggogeri F, Pellegrini N, Tagliani FL. applied sciences Recent Advances on Machine Learning Applications in Machining Processes. 2021;
Sizemore NE, Nogueira ML, Greis NP, Davies MA, Sizemore NE, Nogueira ML, et al. ScienceDirect ScienceDirect ScienceDirect Application of Machine Learning to the Prediction of Surface Roughness in Application of Machine Learning to the Prediction of Surface Roughness in Diamond Machining Diamond Machining. Procedia Manuf [Internet]. 2020;48(2019):1029–40. Available from: https://doi.org/10.1016/j.promfg.2020.05.142
Hahn T Von, Mechefske CK. Machine Learning in CNC Machining : Best Practices. 2022;1–27.
Mcdonnell MDT, Arnaldo D, Pelletier E, Matthew JAG jacob, Karnakis D, Eason RW, et al. Machine learning for multidimensional optimization and predictive visualization of laser machining. J Intell Manuf [Internet]. 2021;32(5):1471–83. Available from: https://doi.org/10.1007/s10845-020-01717-4
Farias A De, Luiz S, Almeida R De. Simple machine learning allied with data-driven methods for monitoring tool wear in machining processes. 2020;l.
Outeiro J, Cheng W, Chinesta F, Ammar A. Modeling and Optimization of Machining of Ti-6Al-4 V Titanium Alloy Using Machine Learning and Design of Experiments Methods. 2022;
Wolpert DH. Stacked generalization. Neural Networks [Internet]. 1992;5(2):241–59. Available from: https://www.sciencedirect.com/science/article/pii/S0893608005800231
Zhou H, Huang GB, Lin Z, Wang H, Soh YC. Stacked Extreme Learning Machines. IEEE Trans Cybern. 2015;45(9):2013–25.
A_1018046112532.pdf.
Mangalathu S, Hwang SH, Jeon JS. Failure mode and effects analysis of RC members based on machine-learning-based SHapley Additive exPlanations (SHAP) approach. Eng Struct [Internet]. 2020;219:110927. Available from: https://www.sciencedirect.com/science/article/pii/S0141029620307513
Brusa E, Cibrario L, Delprete C, Gianpio L, Maggio D. applied sciences Explainable AI for Machine Fault Diagnosis : Understanding Features ’ Contribution in Machine Learning Models for Industrial Condition Monitoring. 2023;
Garouani M, Ahmad A, Bouneffa M, Hamlich M, Bourguin G, Lewandowski A. Toward big industrial data mining through explainable automated machine learning. Int J Adv Manuf Technol [Internet]. 2022;120(1):1169–88. Available from: https://doi.org/10.1007/s00170-022-08761-9
Ukwaththa J, Herath S, Meddage DPP. A review of machine learning (ML) and explainable artificial intelligence (XAI) methods in additive manufacturing (3D Printing). Mater Today, Commun [Internet]. 2024;41:110294. Available from: https://www.sciencedirect.com/science/article/pii/S235249282402275X
Mia M, Dhar NR. Prediction and optimization by using SVR, RSM and GA in hard turning of tempered AISI 1060 steel under effective cooling condition. Neural Comput Appl. 2019;31(7):2349–70.
Yadav MP, Sharif T, Ashok S, Dhingra D, Abedin MZ. Investigating volatility spillover of energy commodities in the context of the Chinese and European stock markets. Res Int Bus Financ [Internet]. 2023;65:101948. Available from: https://www.sciencedirect.com/science/article/pii/S0275531923000740
Sutton CD. 11 - Classification and Regression Trees, Bagging, and Boosting. In: Rao CR, Wegman EJ, Solka JLBTH of S, editors. Data Mining and Data Visualization [Internet]. Elsevier; 2005. p. 303–29. Available from: https://www.sciencedirect.com/science/article/pii/S0169716104240111
Freund Y, Schapire RE. A Decision-Theoretic Generalization of On-Line Learning and an Application to Boosting. J Comput Syst Sci [Internet]. 1997;55(1):119–39. Available from: https://www.sciencedirect.com/science/article/pii/S002200009791504X
Fawagreh K, Gaber MM, Elyan E. Random forests: from early developments to recent advancements. Syst Sci Control Eng [Internet]. 2014 Dec 1;2(1):602–9. Available from: https://doi.org/10.1080/21642583.2014.956265
Wakjira TG, Ibrahim M, Ebead U, Alam MS. Explainable machine learning model and reliability analysis for flexural capacity prediction of RC beams strengthened in flexure with FRCM. Eng Struct [Internet]. 2022;255(August 2021):113903. Available from: https://doi.org/10.1016/j.engstruct.2022.113903
Shrestha DL, Solomatine DP. Experiments with AdaBoost.RT, an Improved Boosting Scheme for Regression. Neural Comput [Internet]. 2006 Jul 1;18(7):1678–710. Available from: https://doi.org/10.1162/neco.2006.18.7.1678
Statistics M. Institute of Mathematical Statistics. 2024;29(5):1189–232.
Chen T, Guestrin C. XGBoost : A Scalable Tree Boosting System. 2016;785–94.
Wakjira TG, Kutty AA, Alam MS. A novel framework for developing environmentally sustainable and cost-effective ultrahigh-performance concrete (UHPC) using advanced machine learning and multiobjective optimization techniques. Constr Build Mater [Internet]. 2024;416(January):135114. Available from: https://doi.org/10.1016/j.conbuildmat.2024.135114

Prediction of surface roughness of tempered steel AISI 1060 under effective cooling using super learner machine learning.

Status:

Version 1

Abstract

Figures

1 Introduction

2 Methodology