Physics-informed gradient boosting regression tree for tool wear prediction in milling

Abstract Accurate prediction of tool flank wear is critical for maintaining machining quality, minimizing unplanned downtime, and enabling condition-based tool replacement. Data-driven models may violate physical wear laws under limited or varying data, whereas mechanistic models often lack the flexibility to capture complex, nonlinear wear evolution. To bridge this gap, this study proposes a physics-informed gradient boosting regression tree (PIGBRT) approach that integrates the generalized tool wear degradation law into the GBRT learning framework. In this method, the physical degradation law is formulated as a regularization term and integrated with the conventional data loss to construct a composite objective function, thereby jointly enforcing data fidelity and physical consistency during model training. The regression trees are then iteratively constructed along the negative gradient of the composite loss, enabling the physical prior to continuously constrain the learning trajectory rather than being imposed only during post-processing. Experimental validation on milling datasets demonstrates that the proposed PIGBRT achieves accurate tool wear prediction, with superior generalization and physical consistency compared with purely data-driven, mechanistic, and other advanced methods. These results highlight the effectiveness of embedding explicit physical knowledge into ensemble learning and establish PIGBRT as a promising framework for physics-consistent and generalizable tool monitoring in milling processes.