This paper presents a machine learning neural network capable of approximating pressure as the distributive result of elastohydrodynamic effects for a journal bearing at steady state. Design of efficient, reliable fluid power pumps and motors requires accurate models of lubricating interfaces; however, many state-of-the-art simulation models are structured around numerical solutions to the Reynolds equation which involve nested iterative loops, leading to long simulation durations and limiting the ability to use such models in optimization studies. This study presents a machine learning model capable of approximating the pressure solution of the Reynolds equation for a journal bearing with given distributive geometric boundary conditions and considering cavitation and elastic deformation at steady state operating conditions. A 1024-sample training set was generated using an in-house multiphysics simulator. A hyperparameter optimization study was conducted, leading to the six-layer U-Net convolutional neural network architecture proposed. After training, the neural network accurately predicted pressure distributions for test samples with different geometric inputs from the training data, and accurately estimated resultant journal bearing loads, showing the feasibility of post-processing the machine learning output for integration into other fluid power models. Additionally, the neural network showed promise in analyzing geometric inputs outside the space of the training data, approximating the pressure in a grooved journal bearing with reasonable accuracy. These results demonstrate the potential of a machine learning model to be integrated into fluid power pump and motor simulations for faster performance evaluation and optimization.

This paper presents a machine learning neural network capable of approximating pressure as the distributive result of elastohydrodynamic effects and discusses results for a journal bearing at steady state. Design of efficient, reliable fluid power pumps and motors requires accurate models of lubricating interfaces; however, most existing simulation models are structured around numerical solutions to the Reynolds equation which involve nested iterative loops, leading to long simulation durations and limiting the ability to use such models in optimization studies. This study presents the development of a machine learning model capable of approximating the pressure solution of the Reynolds equation for given distributive geometric boundary conditions and considering cavitation and elastic deformation at steady-state operating conditions. The architecture selected for this study was an 8-layer U-Net convolutional neural network. A case study of a journal bearing was considered, and a 438-sample training set was generated using an in-house multiphysics simulator. After training, the neural network predicted pressure distributions for test samples with great accuracy, and accurately estimated resultant loads on the journal bearing shaft. Additionally, the neural network showed promise in analyzing geometric inputs outside the space of the training data, approximating the pressure in a grooved journal bearing with reasonable accuracy. These results demonstrate the potential to integrate a machine learning model into fluid power pump and motor simulations for faster performance during evaluation and optimization.