Data-Driven Modelling of TEHL Contacts: ANN-Based Prediction of Friction, Film Thickness, and Temperature

Abstract

Conventional Thermo-Elastohydrodynamic Lubrication (TEHL) simulations, while accurate, are computationally expensive and impractical for system-level analyses. In this study, Artificial Neural Network (ANN)-based surrogate models were developed for predicting key TEHL outputs, including Coefficient of Friction (CoF), rolling friction, maximum temperature, and film thickness metrics. A comprehensive dataset of approximately 36,000 samples was generated by solving the generalised Reynolds equation and Boussinesq integral under non-Newtonian and non-isothermal conditions, capturing thermal effects that are often neglected in previous studies, particularly for friction prediction. Separate ANN models were trained for each output while hyperparameters were optimized, achieving high predictive accuracy with R² values exceeding 0.99 for most outputs. To address data sparsity and improve, particularly, the CoF prediction, a novel adaptive sampling strategy was implemented, enriching low-density regions of the output space and enhancing generalisation and accuracy. Sensitivity analysis revealed that CoF is primarily influenced by the piezo-viscous coefficient and Slide-to-Roll Ratio (SRR), while film thickness metrics depend on entrainment speed and viscosity. Moreover, maximum temperature rise is dominated by SRR and pressure-viscosity effects, with thermal properties playing a minor role for most of the TEHL outputs. The proposed models demonstrate good agreement with reference TEHL solutions across diverse operating conditions, offering a computationally efficient alternative for real-time calculations of lubricated contacts.