Improving milling force predictions: A hybrid approach integrating physics-based simulation and machine learning for remarkable accuracy across diverse unseen materials and tool types

Abstract

The prediction of milling forces has been addressed using a range of methods, including physics-based models and data-driven approaches. Analytical predictions that rely on mathematical models may not always provide the desired level of accuracy, whereas data-based approaches require extensive testing. A hybrid approach, which combines physics-based simulation results with machine-learning algorithms that integrate measurement data from a limited number of tests, can be employed as an effective alternative to improve the accuracy of milling force predictions. Through the implementation of this novel milling hybrid model, the accuracy of the milling force predictions is significantly improved to levels that cannot be achieved with process models alone. In this approach, a trained machine learning algorithm using simulation results and a small set of test data is a valuable tool for predicting milling forces under various conditions with high accuracy. One of the greatest advantages of this method is that the ML model trained on Al7075-T6, Steel 1050, and Ti6Al4V materials also improved the prediction accuracy for completely different materials, such as Inconel 625. This is mainly due to the way materials are defined in the machine learning system, that is, by their thermomechanical properties, which allow different materials to be input without additional testing. Furthermore, this method can be used to predict the cutting forces of special milling tools (i.e., serrated edges with cylindrical and tapered end mills with flat, ball, and round noses) with a high level of accuracy. It is demonstrated that the accuracy of the cutting force prediction in various cases can be increased up to 98 % (R2) through the implementation of this method. According to statistical error analysis, the majority of deviations between the improved model predictions and measured results fall within a narrow band of −5 % to 5 %, encompassing 90 % of the observations. It is important to note that this high prediction accuracy was achieved with very limited test data and simulation results.