Machine Learning for Solving Elastohydrodynamic Lubrication Problems for Fluid Power Applications

Design of efficient, reliable fluid power pumps and motors requires accurate modeling of lubricating interfaces in these devices, as lubricating interfaces contribute significantly to the efficiency and performance of hydraulic pumps and motors. Analysis of these lubricating interfaces, however, is complicated due to the complex, coupled physics governing thin films, including elastohydrodynamic effects, cavitation, mixed lubrication, and other factors. Furthermore, for industrial applications, models must be computationally inexpensive enough to be useful for design iteration and optimization. Machine learning is a technique that is rapidly growing in the scientific community for its ability to model complex, nonlinear relationships. In this research, two use cases of machine learning for fluid power modeling were studied. The first use case developed a convolutional neural network to predict the pressure distribution in a journal bearing as a function of the undeformed film thickness distribution. This analysis demonstrated that the machine learning model was capable of incorporating coupled elastohydrodynamic effects and evaluating the pressure distribution significantly more quickly than a comparable numerical model. The model also showed potential for analyzing a grooved journal bearing outside the training data domain. The second use case developed a numerical model of a novel design of bent axis piston pump and integrated a recurrent neural network into the model to emulate a journal bearing in the pump. This hybrid physics-machine learning model demonstrated the feasibility of coupling numerical and machine learning models and achieving comparable results for other lubricating interfaces in the pump; however, differences in micromotion between the numerical and hybrid physics-machine learning models were observed, and further research is needed to understand how to optimize the neural network architecture and training to reduce error. This research highlights several potential use cases for machine learning in fluid power research and demonstrates the feasibility of replacing or augmenting certain numerical analyses with machine learning.