This paper presents a machine learning neural network capable of approximating pressure as the distributive result of the Reynolds equation and discusses results for a journal bearing at steady state.&nbsp; Design of efficient fluid power pumps and motors requires accurate models of lubricating interfaces, as significant frictional losses, leakages, and wear occur at these locations.&nbsp; However, most existing simulation models are structured around numerical solutions to the Reynolds equation.&nbsp; Inherent in these solutions are iterative loops, and when combined with fluid-structure interaction and thermal solvers, the resulting models contain many nested iterative calculations, leading to long simulation duration and limiting the ability to use such models in optimization studies.&nbsp; A new approach offering faster simulation times without neglecting complex physical effects is required.In recent years, machine learning has gained popularity for its ability to learn and approximate complex mathematical operations.&nbsp; Past studies have successfully applied machine learning to fluid and solid body finite element modeling; however, no studies encountered to date have attempted to apply machine learning to predict full, two-dimensional pressure distributions for tribological interfaces.This paper introduces a proof-of-concept neural network capable of approximating the pressure distribution as the result of the Reynolds equation for given distributive geometrical boundary conditions and steady-state operating conditions, and results are presented for the test case of a journal bearing. The machine learning architecture selected for this study was a convolutional neural network; specifically, an 8-layer U-Net.&nbsp; This model was trained to predict the two-dimensional pressure distribution over the journal bearing film from a given film thickness distribution.&nbsp; An example neural network was trained on 1026 samples of film thickness/pressure distributions for a misaligned journal bearing, varying eccentricity and misalignment angle.&nbsp; Training and testing data were generated using an in-house multiphysics simulator, and included the effects of elastic deformation and cavitation. After tuning hyperparameters and training, the neural network was able to predict the pressure distribution with great accuracy on a held-out test set, with more than an order of magnitude decrease in prediction time relative to the numerical solution.&nbsp; A physics-informed loss function calculating residuals from the Reynolds equation is compared to standard L1 and L2 loss functions for training the neural network.&nbsp; This article also reports different neural network outputs and data normalization strategies and their influence on model accuracy.&nbsp; These results demonstrate the potential of a physics-informed machine learning model to be integrated into fluid power pump and motor simulations for faster performance evaluation and optimization.

Machine Learning Prediction of Journal Bearing Pressure Distributions, Considering Elastic Deformation and Cavitation

Paper Proceedings: https://asmedigitalcollection.asme.org/FPST/FPMC2021/volume/85239

2021 ASME/BATH Symposium on Fluid Power and Motion Control (FPMC2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.