Physics-informed gradient boosting tree for vision-based prediction of gear flow field structures across wide temperature ranges

The quantification of the nonlinear evolution of gear flow field structures over wide temperature ranges is crucial for gearbox lubrication and reliability design. Conventional experimental methods are costly, numerical simulations demand substantial computational resources, and traditional data-driven models lack effective integration of physical constraints. To address these limitations, a physics-informed gradient boosting tree (PI-GBT) approach for fast and accurate prediction of gear flow field structures is developed. Data are acquired using a temperature-adjustable visualized test rig for churning oil flow, with six key flow field structure sizes summarized as prediction targets. The PI-GBT integrates fluid dynamics constraints into gradient boosting trees, and its performance is compared with Gaussian process regression, physics-informed neural networks, and gradient boosting tree. The proposed model achieved superior predictive performance with R ² values ranging from 0.843 to 0.991, and partial dependence analysis revealed the dominant physical mechanisms governing each flow field structure size. The model achieves rapid prediction of flow field structures within a temperature range of −30°C to 80°C. It provides methodological support for optimizing efficient lubrication in gearboxes.