• Novel LSTM-MLP framework predicts EHL friction using full roughness profiles. • Sequential processing captures topographical features missed by scalar metrics. • Physics-informed design models viscous and boundary friction components separately. • Deep learning approach significantly outperforms standard regression baselines. This study proposes a novel physics-informed deep learning surrogate framework for the prediction of mixed elastohydrodynamic lubrication (EHL) friction. It implements parallel Long Short-Term Memory (LSTM) networks followed by a Multi-Layer-Perceptron (MLP) module, capable of extracting key features from entire surface roughness profiles. Unlike scalar parameters that lose spatial details, the LSTM components directly encode the full spatial sequences of the roughness profiles, extracting comprehensive topographical features. These features are integrated with operating conditions to predict viscous and boundary friction components separately. Trained on a comprehensive dataset of 888,720 mixed EHL simulation points, the LSTM-MLP model achieves exceptional predictive accuracy, with a Mean Squared Error (MSE) of 1.91 × 10 –7 , significantly outperforming General Linear Regression (MSE 1.56 × 10 –5 ) and standard MLP (MSE 6.72 × 10 –7 ) baselines. Crucially, the model successfully distinguishes between surfaces with identical statistical parameters (such as RMS amplitude) but distinct textures, proving its ability to generalize and capture the true physical governing mechanisms of rough surface friction.
Sheng Li (Wed,) studied this question.