Machine-learning-based prediction of cut quality parameters of plasma arc cutting

Abstract

Stainless Steel 304 is known for its excellent mechanical properties; however, it is difficult to machine using conventional methods. Plasma arc cutting, a non-conventional machining technique, offers an efficient alternative; however, the quality of the cut is highly sensitive to process parameters. In this study, a design-of-experiment-based Central Composite Design was used to design experiments and analyze the effects of key input parameters such as Current, Cutting Speed and Pressure on cut quality indicators such as Kerf Width and Heat-Affected Zone. The experimental data were further utilized to train and evaluate machine learning regression models, including Linear Regression, Random Forest, Gradient Boosting, eXtreme Gradient Boosting, and Adaptive Boosting. Pearson's Correlation Heatmap was employed to validate the relationships between input parameters and response variables. Comparative analysis showed that the Gradient Boosting model provided the most accurate predictions, closely matching the experimental results for both the Kerf Width and Heat-Affected Zone. Specifically, the Gradient Boosting model achieved R² values of 0.983 for the Kerf Width and 0.968 for the Heat-Affected Zone. In contrast, the traditional model based on Response Surface Methodology yielded lower R² values of 0.878 for the Kerf Width and 0.886 for the Heat-Affected Zone. These results clearly indicate that the machine learning models, particularly Gradient Boosting Regression, outperform the traditional statistical model based on Response Surface Methodology in terms of accuracy and prediction efficiency for plasma arc cutting of Stainless Steel 304.