Under low-speed and heavy-load conditions, the liquid film thickness in journal bearings is often on the same order as the surface roughness. Conventional analytical methods for bearings considering rough surfaces require high computational costs. A rapid prediction method for lubrication characteristics based on physics-informed neural networks (PINNs) is presented in this paper to improve the efficiency of rough bearing analysis in engineering applications. For conventional eccentricity conditions, an extended PINN architecture is used to solve the average Reynolds equation, with the eccentricity ratio as an input parameter to predict bearing characteristics under rough-surface conditions. For extremely high eccentricity-ratio conditions, based on a residual-based adaptive sampling strategy, physical constraints are combined with data fitting loss to improve the prediction accuracy and stability using a hybrid-driven PINN method. The extended PINN and the hybrid-driven PINN are combined to achieve lubrication performance prediction of rough bearings over a wide eccentricity ratio range. The model's validity is verified by comparing its results with experimental data and with those obtained using the finite difference method. Finally, the influence of surface roughness on the bearing lubrication characteristics is analyzed. The results show that the proposed PINNs-based bearing performance prediction method considering rough surfaces has high prediction accuracy and computational efficiency, providing an effective engineering analysis tool for the rapid prediction of rough bearing performance.
Ding et al. (Sun,) studied this question.