Deviated Elastohydrodynamic Line Contacts: Prediction Modeling using Machine Learning

Abstract

As manufacturing related surface deviations considerably influence film thickness and pressure distributions in EHL contacts, their incorporation into numerical modeling is crucial. However, deviated EHL simulations are computationally demanding and complex. Against this background, this study aims to enable real-time prediction modeling of film thickness as well as pressure parameters in EHL line contacts underlying surface waviness by means of developing prediction models by machine learning approaches, namely artificial neural networks (ANN) and GAUSSian process regression (GPR). Following latin hypercube sampling, more than 1800 transient EHL simulations, accounting for non-NEWTONian lubricant behavior and double-sided deterministic waviness, served as a database for prediction modeling. After hyperparameter optimization, ANN as well as GPR models revealed comparably high prediction accuracies with adjusted coefficients of determination greater than 0.998 and 0.975 with respect to film thickness and maximum pressure parameters, respectively. The results obtained contribute to the fast and precise consideration of surface waviness in EHL contacts in further research and early phases of product development.