Exploring the correlation between design of experiments and machine learning prediction accuracy in ultra-precision hard turning of AISI D2 with CBN insert: a comparative study of Taguchi and full factorial designs

Abstract Efficient optimization of processes involves excellent knowledge of the influence cum contributing effect on product quality, reliability, and improved productivity via precision machining in modern manufacturing technology. Thus, this research explores the correlation between design of experiment, process optimization, and predictive accuracy of machine learning models. An ultraprecision hard turning finishing experiment was carried out on an AISI D2 of 62 HRCs via a CBN insert with two different experimental designs, namely, the Taguchi design and the full factorial design. The process parameters, cutting speed ( v c ), feed ( f ), and depth of cut ( ɑ p ) were investigated with signal‒to‒noise ratios ( S/N ) and the response surface method ( RSM ), which are suitable for each experimental design for response parameters: surface roughness ( R ɑ ) and the material removal rate ( MRR ), respectively. The Bayesian regularization neural network (BRNN)–based machine learning model was implemented to estimate the surface roughness with data from each experimental design. The results show that surface roughness was strongly influenced by the feed, whereas the material removal rate was affected by all process parameters. The model performance significantly improved as additional process parameters were introduced in the full factorial design, with an R 2 of 0.99% and a MAPE of 8.14%. An empirical equation for estimating R ɑ is expressed in matrix form using the weights and biases from the BRNN. For the integration of the proposed model in real-time manufacturing and decision-making, an additional experimental test was performed to validate the models with a new dataset. The results show that the full factorial design has an improvement of 36% in predictive accuracy with minimum error over the Taguchi design and provides excellent interpretability of the process parameters. A dual assessment metric criterion was employed to ascertain the credibility of the models with corresponding designs.