Reliability assessment of surface roughness prediction models through dual external validation in turning of SUS 304

Surface roughness (Ra) is a key indicator of machined surface quality and influences product performance. This study investigates the effects of cutting speed (v), feed rate (f), tool nose radius (r), and machining diameter (d) on Ra during external turning of SUS 304 stainless steel. A Taguchi L27 experimental design was employed, and four predictive approaches, namely Polynomial Regression, Random Forest Regression (RFR), Artificial Neural Networks (ANN), and Extreme Learning Machine (ELM), were evaluated. To assess model reliability, 12 additional experiments were conducted and divided into two external validation groups: interpolation validation and unseen-parameter-combination validation. Taguchi, ANOVA, and RFR analyses identified f and r as dominant factors affecting Ra, whereas v and d showed minor effects. Although ANN achieved high interpolation accuracy, ELM achieved the most favorable numerical performance according to R2, MAE, and RMSE in the six Group B experiments. However, the differences among the four models were not statistically significant. The results demonstrate that model ranking can change when independent external validation is introduced, and that reliance solely on internal validation can overestimate predictive performance. The proposed dual external validation strategy provides a more realistic basis for evaluating predictive models in machining applications.