Machine learning-based surface roughness prediction in magnetorheological finishing of polyamide influenced by initial conditions

Abstract

Surface roughness prediction enhances manufacturing efficiency and reduces costs by minimizing trial-and-error testing. Machine learning can address the uncertainty and nonlinear relationships in magnetorheological finishing (MRF), providing a reliable alternative. However, its application in this area remains underexplored. Therefore, this paper proposes a machine learning-based model for predicting the final surface roughness Rₐd of polyamide 6 based on initial conditions and several other variables. Experiments were conducted with varying process parameters consisting of durations, rotational speeds, and gaps between the tool and workpiece to generate training data. Statistical analysis was performed to assess correlations, trends, and model complexities. Four output-input schemes are formulated to identify the best configurations. The deployed machine learning models are Feedforward Neural Networks (FFNN) trained using the Levenberg-Marquardt (LM) and Extreme Learning Machines (ELM). The LM base-FFNN model accurately predicted the outputs with fewer hidden nodes, while ELM offered comparable accuracy with faster training, albeit requiring more parameters. The models were evaluated based on R2 and RMSE values, achieving R2 values of >0.90 in training and testing cases. Among the proposed schemes, the one predicting the difference between final and initial surface roughness (ΔRad) while considering all inputs using ELM provided the best accuracy compared to the other schemes. Direct prediction of ΔRad shows potential, but the data is more concentrated toward the half range, reducing the generalization capability. The gap parameter can affect the ΔRad prediction accuracies slightly as it affects the weakening or strengthening of the magnetic fields. Meanwhile, the elimination of the initial surface roughness condition as one of the inputs can severely degrade accuracy, resulting in an R2 value of <0.40. In conclusion, our findings emphasize the promise of machine learning-based predictive models and the importance of incorporating initial conditions for assisting MRF-based polishing processes.