Ensemble Machine Learning for Predicting Machining Responses of LB-PBF AlSi10Mg Across Distinct Cutting Environments with CVD Cutter

Abstract

The efficiencies of additive manufacturing (AM) over conventional processes have enabled the rapid production of aluminum (Al) alloys with AM. Because laser beam powder bed fusion (LB-PBF) parts do not offer the surface quality and geometrical accuracy for direct use, the functional surfaces of LB-PBF parts are usually machined by subtractive machining. The machinability of LB-PBF AlSi10Mg was studied in dry, MQL (used corn oil), and cryo-LN₂ cutting environments across distinct speed–feed combinations using CVD-AlTiN-coated carbide inserts, and surface integrity and tool life were quantified in terms of surface roughness (Ra) and flank wear (Vb), respectively. The lowest Ra (0.98–1.107 μm) was obtained with cryo-LN₂, followed by MQL and dry cutting environments, because the trends observed were consistent with the surface mechanisms observed in 3D topography and bearing curves. Similarly, the tool wear results mirrored the Ra results, lowest with LN₂ (0.087–0.110 mm), due to improved thermal management, reduced adhesion and abrasion, and shorter contact length. Cryo-LN₂ provided the best surface finish and tool life among all tested environments. To enable data-driven prediction, the limited dataset was augmented using SMOTE, and machine learning (ML) models were trained to predict Ra and Vb. CatBoost was found to yield the best Ra predictions (R² = 0.9090), while Random Forest and XGBoost yielded the best Vb predictions (R² ≈ 0.878).

1. Introduction

A wide range of industries, including automotive, aerospace, construction, and packaging, have made extensive use of Al alloys because of their favorable material qualities and cost advantages over steel and titanium alloys. Cast and wrought Al alloys are distinguished by their microstructure, molecular composition, and processing features [1]. The advantages of AM technologies over conventional manufacturing methods like casting, forging, and extrusion have led to a recent rise in efforts to manufacture Al alloys using AM technologies [2]. These advantages include the ability to fabricate lightweight parts with complex geometries. The most advanced of the several AM techniques is LB-PBF, in which metallic powder particles are melted and fused by a high-power-density laser after the powder has been spread out over the powder bed [3]. There are difficulties in fabricating parts with AM techniques; the parts undergo complex thermal processing (i.e., rapid melting and solidification), which can alter their microstructure, create residual stresses, and result in process-induced flaws like pores, lack of fusion, Ra, etc. [4].

The surface quality and geometric accuracy of as-built LB-PBF parts are generally insufficient for direct use, necessitating subtractive machining of their functional surfaces to enhance part quality [5]. Conventional subtractive manufacturing (SM) is a widely established set of techniques employed by the manufacturing sector to achieve desired part features [6]. However, the LB-PBF process enhances toughness and strength by creating a distinctive, fine-grained, and anisotropic microstructure [7]. Consequently, the machinability of LB-PBF parts can differ significantly from their wrought or cast counterparts, even with identical chemical compositions [8]. This is because machining removes material by applying mechanical, chemical, and thermodynamic loads that alter the workpiece (W/p) surface and subsurface [9,10]. To investigate this, Zimmerman et al. investigated the machining characteristics of LPBF-produced AlSi10Mg by examining process forces, surface morphology, micro-hardness, and burr development relative to build direction. The study, which compared the AM part to a cast equivalent under dry machining, confirmed that the manufacturing method impacts surface properties [11]. Even though a number of studies have been conducted on the directionality of AM micro-structural features based on AM variables, the machining behavior of AM parts with anisotropic microstructures has not yet been thoroughly examined [12]. This gap is critical as manufacturers seek alternatives to conventional cutting fluids (CFs) to manage machining heat, which pose significant environmental, health, and financial drawbacks [13,14]. The MQL technique has gained popularity because of its affordability and eco-friendliness, in which compressed air atomizes a minimal quantity of oil (50–500 mL/h) onto the cutting zone [15]. Additionally, the MQL method significantly impacts surface properties, shear strength, and cutter life [16]. Its effectiveness was demonstrated by Ramoni et al., who found that when machining AM-AlSi10Mg, MQL improved surface roughness (Ra) by up to 63% and reduced tool wear (Vb) by up to 45% compared to dry and flood conditions [17]. Cryogenic (cryo) machining offers another environmentally friendly alternative, using evaporative liquid gases like liquid nitrogen (LN₂) to cool the cutting zone without leaving contaminating residue [18]. Ross et al. studied the effects of different cutting environments on tool wear and surface integrity during the milling of LB-PBF AlSi10Mg. Adhesion wear was identified as the main failure mode seen when the AM AlSi10Mg alloy was milled. The hardness of the machined surface was increased under cryo conditions compared to other cutting environments [19].

The digitization of the industry and the development of computer processing capabilities have led to a constant increase in pertinent information in the manufacturing sector in recent years. More notable advances in computational science and artificial intelligence (AI) architecture have led to the growing acceptance and use of AI, making learning algorithms more effective. ML applications in the manufacturing sector are highly popular [20]. To forecast, ML models, including Random Forest (RF) [21], XGBoost (XGB) [22], and (CatB) [23], were employed in previous research works. Danish et al. used a test dataset consisting of predicted Vb to assess tool wear and class separation. Many ML techniques were employed to generate predictions. With an average prediction accuracy of over 95%, the study demonstrates that multi-layer perceptron (MLP) outperforms other models used in the investigation [24].

Ultimately, the machinability of the materials that have been examined thus far is significantly impacted by LB-PBF. The machining procedure and the production techniques, in addition to the LPBF manufacturing parameter settings, seem to influence the nature of this impact. Maintaining part quality during finishing when employing LB-PBF AlSi10Mg components requires a knowledge of how LB-PBF impacts the resultant machining process. In this investigation, dry, MQL, and cryo-LN₂ environments were used under various speed–feed combinations to mill LB-PBF-AlSi10Mg using a CVD-coated cutter. The process’s sustainability was assessed through the use of surface analysis and tool wear. In addition, ML was introduced for the prediction.

2. Experimental Setup

2.1. Fabrication and Machining

An EOS M280 system (Krailling, Germany) was used to fabricate AlSi10Mg specimens via LPBF. The process utilized a laser power of 370 W, a scan speed of 1300 mm/s, a hatch spacing of 190 μm, and a layer thickness of 30 μm. To ensure consistency, the laser was calibrated before each build cycle, during which the powder was selectively melted according to the CAD geometry. The machinability of the fabricated AlSi10Mg specimens was assessed through a milling operation. The tests were conducted on a Feeler FV 1000 (Taichung, China) vertical CNC machining center. A BAP 300R tool holder fixed with APMT 1135 (rake angle = 6° and clearance angle = 11°) carbide inserts was employed; CVD-coated inserts are widely used due to their high cutting performance, wear resistance, and long life in machining operations. Based on a comprehensive literature review to ensure optimal cutting conditions, the cutting speed (Vc), feed rate (fr), and depth of cut (DOC) were selected as detailed in

Table 1. Experimental details.

Milling Setup	Feeler FV 1000
Material (W/p)	AlSi10Mg
Measurement (cm)	10 × 5 × 1
Insert (tool) coating	CVD-AlTiN
Model	APMT
Vc (m/min)	45–75
fr (mm/rev)	0.08–0.10
Radial DOC (mm)	12
Axial DOC (mm)	1
Length of cut (mm)	20 (4 passes)

. A schematic representation of the investigation flow is presented in Figure 1.

2.2. Cooling Conditions

The machinability and overall performance of the LPBF-made AlSi10Mg alloy were tested in three different cutting environments: dry cutting, MQL with used corn oil, and cryogenic cooling with LN₂. Dry machining is better for the environment because it does not use CFs, but it usually makes the cutting temperatures higher, and the tools wear out faster. In the MQL condition, the used corn oil was supplied to the cutting zone as fine spray through a dedicated system at 4 bar supply pressure using a 2 mm nozzle, while the droplets were directed to the tool–workpiece interface to reduce friction and cutting temperature, thus consuming significantly less fluid compared with conventional flood cooling and allowing for stable operation for longer machining runs. The kinematic viscosity of used (for cooking) corn oil after heating (175–185 °C) is significantly higher than that of virgin corn oil at approximately 50–60 mm²/s at 40 °C, whereas virgin corn oil is typically around 35–45 mm²/s at 40 °C. In the cryogenic condition, LN₂ was applied close to the cutting edge at 3 bar through a 5 mm nozzle with the objective of maximizing the heat extraction from the shear and contact zones to reduce the cutting temperature, mitigate the adverse thermal effects, and enhance the surface integrity. For both the MQL and LN₂ conditions, the nozzle was maintained at 45° for all the trials.

2.3. Measurement

Ra, as a direct indicator of the machined surface quality, was measured using a SE3120 (Kawasaki, Japan) roughness tester, which can provide accurate and repeatable data to evaluate the effect of machining condition. The tester accuracy was calibrated with slip gauges before measurement, and five sampling lengths with a cut-off length of 0.8 mm were taken for each machining condition; the average of the five segment values was reported as the final Ra value for the condition. Tool Vbmax was measured as an indicator of tool life and cutting performance by using an OPTIV optical microscope to observe and quantify in detail the wear on the flank face, and the measurement of Vb, together with Ra, allowed for a more comprehensive evaluation of the influence of how the machining environment and cutting conditions influence the final surface finish.

3. Tool Material

In this study, machining trials were conducted using carbide inserts with AlTiN deposited by CVD. Figure 2 presents the schematic of coating thickness and properties of the AlTiN coating. The CVD-AlTiN coating was selected to combine the high hardness and oxidation resistance of AlTiN with the superior coating adhesion and thickness obtained by the CVD process. The uniform coating of 2–5 µm acts as an effective barrier against abrasive, adhesive, and crater wear under severe cutting conditions. The AlTiN layer exhibits a high hardness and excellent thermal and oxidation stability up to 800–900 °C, which helps to maintain cutting edge integrity at elevated temperatures. Compared with PVD coating, CVD coating provides stronger metallurgical bonding to the substrate. The relatively smooth and chemically stable AlTiN promotes chip flow and reduces BUE formation.

4. Results and Discussions

4.1. Surface Roughness

The Ra outcomes demonstrate a strong influence of the machining environment and cutting parameters on the milled LB-PBF AlSi10Mg. Figure 3 illustrates the effect of different feed–speed and environment combinations on the resulting surface roughness (Ra). For the used speed–feed combinations, cryo-LN₂ produced the lowest Ra values (0.98–1.107 μm), followed by MQL (1.448–1.637 μm), and the dry condition exhibited the highest Ra (2.028–2.417 μm). The non-presence of coolant leads to elevated temperatures and higher friction in the cutting area. LB-PBF AlSi10Mg exhibits a heterogeneous microstructure with melt pools, fine Si network, and residual porosity that promotes less stable chip formation and greater plastic deformation of the surface asperities, resulting in a rough machined surface. The introduction of MQL lessens the friction and moderates the cutting temperature through the formation of a thin lubricating film and localized cooling, thereby stabilizing chip flow and ploughing effects on the surface, which leads to an appreciable reduction in Ra relative to dry cutting. The LN₂ condition suppresses excessive thermal softening, which promotes a controlled shearing mechanism. As a result, the LN₂ machined surface exhibits the lowest Ra. Across all environments, lower fr values (0.08 mm/rev) are associated with lower Ra in relation to high fr values (0.10 mm/rev).

Figure 4 compares the 3D surface topography and associated Abbott–Firestone bearing curves of the machined under distinct cutting environments. Under the dry cutting strategy, the surface exhibits pronounced peaks and valleys distributed along the feed direction [25]. The height difference between the highest peaks and lowest valleys is relatively large, indicating a rough surface profile. The bearing curve on the right side shows larger core roughness (Sk). This shows more aggressive cutting, higher friction, and more tool–W/p interaction. When MQL is employed, the surface becomes noticeably smoother compared to dry cutting. The 3D topography shows peaks and valleys, but their amplitudes are reduced. The corresponding bearing curve shifts towards lower roughness values, indicating reduced peak height and lower Sk. The presence of a small amount of lubricant lowers the cutting temperature. With cryo-LN₂ cooling, the machined surface exhibits more uniform topography among the three conditions. The bearing curve is much closer to the horizontal axis, indicating minimal peak height and the lowest Sk of all cases.

4.2. Tool Wear

Figure 5 shows the resulting Vb measured across the different combinations of feed–speed and environment. The results of Vb are also consistent and show clearly the effect of the cutting environment on the tool life, and the highest Vb (0.171–0.204 mm) was obtained for dry cutting. Here, no cooling or lubrication is provided, and, as a consequence, the heat is continuously accumulated at the tool–W/p interface, which accelerates the degradation of the cutting edge and the fast growth of the wear land on the flank face. As for the MQL case with used corn oil, the Vb was significantly reduced to 0.131–0.152 mm, because in this condition, the fine mist is delivered directly into the cutting area and reduces friction and takes part of the generated heat, which diminishes both thermal and mechanical loads on the tool, keeping the cutting edge active for longer. The lowest Vb was obtained when using cryo-LN₂ (0.087–0.11 mm). In this condition, the strong cooling at the cutting zone maintains the tool’s hardness and strength, slowing the degradation of the cutting edge and leading to a substantial increase in tool life.

The appearance of the cutting edge in different environments and speed–feed conditions can be seen in Figure 6, and it helps in explaining the dominant wear mechanisms. The dry cutting resulted in significant BUE along with attrition and abrasion marks, because without any coolant, the temperature at the cutting area is high, the W/p material is softened, and it tends to adhere to the tool to form BUE, which is more common in machining Al. As the adhered layer forms and breaks repeatedly, it pulls off small particles from the tool. Meanwhile, the hard particles from the W/p or the fractured BUE are trapped between the tool and chip and scratch the surface, leading to abrasive wear. Under MQL, the BUE size and the severity of attrition and abrasion were decreased because the lubricant mist leads to a thin film and some cooling at the contact area, which decreases the friction and temperature, reduces the tendency of the material to adhere, and limits the tool–chip contact; it also helps to take away the debris. Cryo-LN₂ resulted in the strongest suppression of the wear mechanisms [26], with very little BUE, abrasion, and attrition, because the very low temperature of cryo-LN₂ reduces the thermal softening of the W/p, decreases the adhesion, and suppresses the chemical and diffusion-related interactions at the interface. Additionally, cooler chips are more likely to break, which shortens the contact length and thereby reduces the mechanical stresses that drive attrition and abrasion.

5. Machine Learning

ML has emerged as a revolutionary technology in industrial research in recent years. It enables researchers to gain knowledge based on data of complex and nonlinear process behaviors that are challenging to represent using conventional analytical or empirical models [27]. Conventional mathematical models frequently depend on simplifying assumptions and exhibit constraints in their adaptability to stochastic fluctuations in machining parameters, material characteristics, and environmental conditions. ML algorithms can find hidden links between input variables and performance outcomes by learning from experimental data. As sustainability, process optimization, and precision manufacturing become more important, the need for smart prediction models that cut down on the need for a lot of trial-and-error testing has grown. Adding ML to machining research makes it possible to analyze data in real time, plan maintenance, and improve cutting conditions, all of which boost productivity. The workflow of ML is presented in Figure 7. RF is an ensemble learning algorithm that uses the idea of combining several decision trees to make predictions more accurate and avoid overfitting [28]. Each tree in the forest learns from a random sample of the training data that was made through bootstrapping. It considers only a random sampling of features at each node split. By ensuring poor correlation, randomization of the trees enhances the ensemble’s capacity for averaging and generalization. XGB trains an ensemble of decision trees on each of the residual errors, one at a time, and uses a second-order Taylor expansion of the loss to approximate the gradient descent update, which leads to faster training and more accurate updates [22]. XGB also uses shrinkage, subsampling, and L1 and L2 regularization to improve generalization. CatB is a gradient boosting method that usually involves ordered boosting, which is based on the boosting algorithm to reduce overfitting and is also the reason why CatB is able to handle categorical features effectively [29]. Furthermore, sequential averaging and ordered target encoding are applied to CatB in order to avoid overfitting, and balanced trees are used to improve computational efficiency and to make the model more generalizable across different feature sets.

To improve the stability of our model training process, a data augmentation technique was applied, and due to the limited size of the dataset, the Synthetic Minority Over-sampling Technique (SMOTE) [30] was employed to balance the distribution by generating synthetic instances for the minority classes. Following the application of SMOTE, the dataset consisted of approximately 300 samples, and therefore, it is able to proceed with the development of predictive models. Predictive models were developed for two responses (Ra and Vb) because these were the key variables of interest in the study, and three distinct algorithms were utilized to create these models (RF, XGB, and CatB). The algorithms were selected because they were suitable for the dataset in this investigation, thus allowing us to leverage their strengths and improve the accuracy of the predictions. Three inputs were considered for prediction: Vc, fr, and cutting environment. The hyperparameters for each model were tuned separately for Ra and Vb to optimize performance based on the distinct characteristics of each target variable. In this study, 80% of the responses were used for training, and 20% were used for testing. RF with 120 estimators, a maximum tree depth of 8, a minimum of five samples per split, and a minimum of two samples per leaf was employed for Ra prediction. XGB for Ra used 40 estimators with a shallow depth of 3, a conservative learning rate of 0.03, and strong regularization (L1 = 0.5; L2 = 2.0) to prevent overfitting on the synthetic data. The learning rate for CatB was 0.07, the depth was 5, and the number of iterations was 80. To predict Vb, RF used 70 estimators with a depth of 6, as well as higher minimum sample limits (10 per split, with 5 per leaf). XGB was set up with 100 estimators, a depth of 6, a learning rate of 0.08, and lighter regularization (L1 = 0.05; L2 = 0.5). CatB for Vb had a simpler structure with 40 iterations, a depth of 3, and a low learning rate of 0.04. It also had higher L2 regularization (4.0) to keep the model from overfitting.

Figure 8 and Figure 9 present the prediction comparison of ML models for responses and the performance metrics of responses with models, respectively. For Ra, CatB performed well with a test R² of 0.9090, an RMSE of 0.1462, and an MAE of 0.1187. RF was very close behind with an R² of 0.9025. In contrast, XGB performed weakly for Ra, with an R² of 0.7419 and RMSE and MAE values of 0.2462 and 0.2156, respectively. There is a difference in the performance measure for Vb when compared to Ra. RF had the best accuracy, with an R² of 0.8784, an RMSE of 0.0136, and an MAE of 0.0107. With an R² of 0.8781, RMSE of 0.0136, and MAE of 0.0108, XGB performed similarly to RF. But CatB did not do as well as the other two algorithms, with an R² of 0.8189, an RMSE of 0.0166 mm, and an MAE of 0.0132 mm. The difference in model rankings between Ra and Vb shows how important it is to choose the right algorithms. These results showed that ensemble methods are great for modeling machining responses.

6. Conclusions

In this investigation, three environments, including dry, MQL (used oil), and cryo-LN₂, were employed across a range of speed–feed combinations in the machining of LB-PBF-AlSi10Mg using a CVD-AlTiN-coated cutter. Moreover, data-driven modeling was used to predict Ra and Vb. The inferences are given below.

• The Ra of the milled LB-PBF AlSi10Mg under all tested speed–feed conditions was found to be lowest for cryo-LN₂ (0.98–1.107 μm), followed by MQL (1.448–1.637 μm), and highest for dry cutting (2.028–2.417 μm). Cryo-LN₂ was most effective in suppressing thermal softening, allowing for controlled shearing, and hence, the lowest Ra and Sk were obtained.
• The tool wear results showed a significant environmental effect on the Vb, and the dry cutting resulted in the highest wear (0.171–0.204 mm), while the MQL significantly reduced the wear (0.131–0.152 mm), and the cryo-LN₂ had the lowest wear (0.087–0.110 mm). The cryo-LN₂ gave the most favorable tool life performance, followed by the MQL, while the dry cutting gave the least favorable tool life performance.
• Data augmentation is applied in this work to ensure the stability of the model training due to the small number of samples, and the dataset is augmented to about 300 observations by applying SMOTE, which generates synthetic samples in order to balance under-represented regions of the feature space. Predictive models for Ra and Vb are developed by utilizing three ensemble learning algorithms, RF, XGB, and CatB, which have shown great performance in the modeling of nonlinear relationships in manufacturing processes.
• CatB yielded the best predictions for Ra, and RF performed closely, while XGB was not the best for Ra. RF and XGB performed very similarly at the top for Vb, and CatB performed poorly. The change of rank between Ra and Vb suggests that the algorithm should be response specific, and therefore, the ensemble methods are very successful in modeling machining responses.