Machine learning-driven empirical modeling and optimization in rotary tool micro-ultrasonic machining

Rotary Tool Micro-Ultrasonic Machining (RT-MUSM), a variant of conventional micro-USM, is well-regarded for its capability to machine hard and brittle materials regardless of their electrical conductivity. It offers enhanced machining rates and better form accuracy compared to conventional methods. The interaction between the tool, abrasive particles, and workpiece plays a critical role in energy penetration, which directly influences performance parameters. Hence, understanding the effect of input parameters governing energy transfer is essential. In this study, a data-driven approach was adopted to model the RT-MUSM process using machine learning techniques. An empirical relationship between key input parameters i.e. tools rotational speed, workpiece feed rate, power rating, and abrasive size, and output responses was established through dimensional analysis using Buckingham’s Pi theorem. To enhance the model’s accuracy, the Levenberg–Marquardt algorithm, a machine learning-based nonlinear optimization method, was employed to estimate constants and variables within the model. The depth of channel (DOC) and tool wear rate (TWR) were used to quantify productivity and machining accuracy, respectively. Model predictions showed strong agreement with experimental results, demonstrating the effectiveness of the ML-driven approach. Furthermore, a genetic algorithm-based multi-response optimization was applied to identify the optimal combination of input parameters for minimizing TWR while maximizing DOC. The mechanisms of material removal and tool wear were also examined through microscopic analysis. This integrated use of machine learning significantly improved the modeling precision and optimization efficiency of the RT-MUSM process.