Toward Explicability of Machine Learning Models Applied to Tribology

Abstract This work proposes to use machine learning to predict and interpret the instantaneous coefficient of friction from pin-on-disc tribological tests. A database was constructed from 14 experiments to link third-body morphology to the coefficient of friction. Morphological descriptors were extracted from the scanning electron microscope (SEM) image of the friction track, including wear particle and texture descriptors. A random forest algorithm was trained to predict the coefficient of friction with high accuracy. Emphasis is placed on model interpretability using the SHAP tool (SHapley Additive exPlanations) to understand the relative influence of morphological features. This approach aims to provide tribological insights into the structural, mechanical, and physical phenomena governing instantaneous friction, opening new perspectives for understanding tribological processes.