Enhancing practical modeling: A neural network approach for locally-resolved prediction of elastohydrodynamic line contacts

Abstract

When modeling bearings in the context of entire transmissions or drivetrains, there are practical limits to the calculation resources available to calculate single bearings or even contacts. In settings such as these, curve-fitting methods have historically been deployed to estimate the elastohydrodynamic lubrication conditions. Machine learning methods have the potential to enable more sophisticated physical modeling in the context of larger computation environments, as the evaluation time of a trained model is typically negligible. We present a neural network that accurately evaluates the locally variable elastohydrodynamic film pressure and film thickness distributions and explore its application to (e.g.) cylindrical roller bearings. Employing a neural network for the EHL film thickness calculations rather than the curve-fitted, simplified methods that are today’s standard can enable a more physically precise modeling strategy at almost no additional computational cost.