Optimization of process parameters for predicting surface roughness in the magnetorheological finishing of hollow cylindrical surfaces using a machine learning algorithm

Achieving high-precision surface quality in nano-finishing processes requires effective optimization of process parameters. This requires a large dataset for which experimentation is cumbersome, and the obtained experimental data is of a non-linear and contains outliers. Therefore, for effective prediction of surface roughness with non-linear experimental data, robust prediction techniques are required. Henceforth, this study focuses on optimizing parameters for the newly developed magnetorheological finishing process for internal cylindrical surfaces using a Gaussian Process Regression (GPR) model integrated with Bayesian optimization. The GPR model served as a probabilistic surrogate to predict surface roughness with exceptional accuracy, achieving an R² = 0.99 and an overall prediction accuracy of 99%. The proposed GPR model achieved superior prediction accuracy for surface roughness, with a Root Mean Squared Error (RMSE) of 3.3089, a Mean Absolute Error (MAE) of 2.6402, and a Relative RMSE (RRMSE) of 0.0222. It outperformed other machine learning models, SVM, DNN, and ANN, by achieving the lowest MSE (10.9487), MAPE (1.84%), and highest correlation coefficient (CC = 99.49%) in the testing phase, demonstrating its robustness and generalization capability. Bayesian optimization was employed to effectively search for the best combination of parameters., identifying the optimal conditions for achieving a nano-level finish. The robustness and generalization capability of GPR model were verified through K-fold cross-validation. Hyperparameter optimization further enhanced the model’s performance, reducing errors, enabling reliable prediction and achievement of ultra-smooth surface. Optimal results were obtained with a tool rotational speed of 420 rpm, working gap of 0.8 mm, and reciprocating speed of 3 cm/s. Under these conditions, the internal surface roughness was drastically reduced to an impressive Ra = $0.059\mu m$.