Machine Learning for Prediction of Surface Roughness in Stellite 6 Turning: A Comparative Study of ANN, Random Forest and Support Vector Regression Models

Abstract

In recent years, the landscape of predictive modeling has been significantly transformed by the adoption of sophisticated methodologies, collectively known as soft computing techniques. The surface condition of a machined part plays a crucial role in its overall performance. It can significantly affect properties such as wear behavior, resistance to corrosion, and fatigue strength. Among the various indicators used to evaluate this finish, surface roughness (Ra) remains one of the most important criteria for assessing the quality of a machined surface.

Stellite 6 is a cobalt-based alloy widely employed in applications that demand high wear resistance and withstand elevated temperatures. While its mechanical strength makes it highly reliable in harsh environments, it also makes the alloy difficult to machine, particularly when aiming to achieve a good surface finish. The problem is framed using the Design of Experiments (DOE) to measure the Ra of Stellite 6 turning: a total of 27 experiments were performed.To better anticipate surface roughness during turning, this study investigates the performance of several machine learning (ML) models developed in Python. Three algorithms: Artificial Neural Networks (ANN), Random Forests (RF), and Support Vector Regression (SVR) were trained using an experimental dataset that includes tool noise radius, cutting speed, feed rate, and depth of cut.

Model performances were assessed using coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error (MAE) metrics. Among the evaluated models, the RF achieved the highest prediction accuracy, followed by ANN and then SVR. The results highlight the capability of Python-based machine learning approaches to capture the nonlinear relationships between cutting parameters and surface roughness. The Random Forest Regressor (RFR) demonstrates the best ability to predict surface roughness (Ra), closely followed by the Epsilon-SVR using the RBF kernel. RFR is less sensitive to data scaling and effectively fits the complex non-linear relationships in the synthetic dataset.

This work provides a foundation for future strategies that integrate both advanced prediction to enhance the machinability of hard-to-cut materials such as Stellite 6.