Grey-box modelling for tool wear prediction in milling: Fusion of finite element insights, time-resolved cutting signals and metaheuristic feature selection

Reliable prediction of tool wear is essential for ensuring productivity, quality, and cost-efficiency in modern machining operations. This study presents a hybrid Grey-Box machine learning framework that combines White-Box finite element simulation outputs (interface temperature, relative sliding velocity), process parameters (feed speed, cutting velocity, depth of cut), dynamic time-series features from cutting force measurements, with a Black-Box machine learning model to predict tool wear in high-speed milling. The experimental campaign involved TiN-coated and uncoated carbide tools under dry machining conditions, with flank and rake wear measured after each cutting pass. Finite element simulations were conducted to extract localized thermomechanical features, such as interface temperature and relative sliding velocity, which were used as physically meaningful inputs. A two-step feature selection method—based on analysis of variance (ANOVA) and the whale optimisation algorithm (WOA)—was employed to identify the most relevant input features. Among the machine learning models tested, the gradient boosting regressor (GBR) showed the highest accuracy, achieving coefficient of determination (R 2 ) scores of 0.953 for rake wear and 0.920 for flank wear. Omitting White-Box features resulted in significantly lower performance, confirming their critical role. The model’s predictions were also in line with the unseen test cases. These results highlight the effectiveness of combining simulation-informed features with empirical data for tool condition monitoring, offering a scalable and interpretable approach for predictive maintenance in smart manufacturing. • Developed a novel hybrid Grey-Box machine learning framework for accurate tool wear prediction in high-speed milling operations. • The framework synergistically combines physics-based (White-Box) finite element simulation outputs, process parameters, and empirical data-driven (Black-Box) features. • Utilized finite element simulations to extract localized thermomechanical features, such as interface temperature and relative sliding velocity, as physically meaningful model inputs. • Conducted extensive experiments using both TiN-coated and uncoated carbide tools under dry machining conditions and measured the flank and rake wear. • Employed a two-step feature selection strategy involving analysis of variance (ANOVA) and the whale optimization algorithm (WOA) to identify the most relevant and informative input features for the predictive model. • Among various machine learning models tested, the gradient boosting regressor (GBR) achieved the highest accuracy, with R 2 scores of 0.953 for rake wear and 0.920 for flank wear.