Linking 3D surface texture to frictional performance in steel-rubber contact using a machine learning approach

Abstract Understanding and predicting friction in rubber–steel sliding contacts remains a major challenge due to the complex interplay between surface topography and contact mechanics. Areal texture parameters defined by ISO 25178 offer a comprehensive description of surface geometry, yet their individual and combined effects on frictional behaviour are not fully understood. Herein, we calculate fifty-six texture parameters for seventeen steel samples with varying surface topographies. The coefficient of friction (COF) is evaluated through a pin-on-disc test, using a rubber pin as the contact material. Measurements were conducted under three contact protocols: single wetting, periodic wetting and greasing and wetting. For each protocol, the dataset is randomly divided into 60% for training, 20% for validation, and 20% for testing, and four Machine Learning (ML) algorithms (SVM, RF, GBM, and ANN) are trained to predict the COF. A final step of validation is conducted on the mean value of three friction measurement repetition, where generated models are used to predict this mean value of COF. Under the three protocols, an excellent agreement between measured and predicted friction was founded, with Root Mean Square Errors (RMSE) as low as 0.035 and coefficients of determination R 2 reaching 0.9. Moreover, the use of texture parameters to train predictive ML models yields robust accuracy, even across different contact conditions. This paves the way for predicting frictional performance based solely on surface topography measurements and tracking the temporal evolution of texture parameters. This approach is especially relevant when an intermediate body is present in the contact interface, where physical modelling becomes both complex and costly.