Using artificial neural networks to accelerate thermo-elastohydrodynamic lubrication simulations

Thermo-elastohydrodynamic lubrication (TEHL) simulations have long been used to predict the lubrication performance of various engineering systems. This study evaluates the use of artificial neural networks (ANNs) for predicting key parameters in tribological systems and investigates the speed and accuracy of three hybrid simulation frameworks that combine ANNs with classical finite volume method (FVM) approaches to solve TEHL problems for point contacts. Six ANNs were trained to predict rigid body separation, maximum fluid temperature, mid-film friction coefficient and spatial profiles of hydrodynamic pressure, liquid film friction and fluid temperature. A total of 600 classical TEHL simulations were used for training and testing. The ANN predictions showed excellent agreement with FVM results for all outputs except for three-dimensional temperature profiles with small variations. The hybrid frameworks achieved simulation speeds up to three times faster than classical FVM simulations without loss of accuracy. With a minor sacrifice in accuracy, speed-ups of up to nine times were observed. To evaluate generalisation across spatial discretisations, the frameworks were also tested on coarser (65 × 65) and finer (193 × 193) meshes, despite being trained only on a 129 × 129 mesh. The results showed that both the ANN predictions and hybrid solutions remained accurate across all mesh resolutions with comparable acceleration effects. This highlights the scalability and robustness of the hybrid frameworks, as well as the ability of ANNs to generalise across different resolutions, significantly reducing the need for retraining and expanding the practical applicability of the approach.