This paper presents the development of a novel Physics-Informed Neural Networks (PINN) method to predict lubrication film characteristics at the slipper-swashplate interface of axial piston pumps. It embeds the Reynolds equation and continuity conditions into the neural network’s (NN) loss function, bridging conventional numerical simulations and data-driven methods. A numerical model for the slipper-swashplate interface was established using the Finite Difference Method (FDM) to generate datasets for training two PINN model variants that incorporate physical constraints, as well as a counterpart pure-data-driven Neural Network (NN) model. The validation shows excellent accuracy of the PINN models (the relative error below 5%) compared with the counterpart NN model (the relative error about 8%). Regarding prediction runtime, the trained PINN and NN models reduce computational time by 99.4% relative to the FDM approach. The study highlights the capabilities and limitations of PINN in tribology, stressing explicit geometric constraints for physical consistency. This study presents an efficient surrogate model for real-time lubrication assessment, establishes a framework for using PINNs in complex tribological systems, and contributes to hydraulic system design optimization and condition monitoring to enable more efficient operation of axial piston pumps. • A surrogate PINN model is established to predict the lubrication film thickness in axial piston pump slipper pairs. • Hard geometric constraints ensure physical consistency of film thickness. • Reynolds equation and mass continuity are embedded in the loss function. • The surrogate model achieves high accuracy matching FDM results. • Computational speed increases by 99.7%, enabling near real-time analysis.
Qiu et al. (Tue,) studied this question.