Multi-scale friction coefficient: From roughness to system computation using deep learning
Victor Lalleman, Pierre Gosselet, Cédric Hubert, Stéphane Salengro, Vincent Magnier
TL;DR
The paper tackles how surface defects influence friction in railway wheel-axle interfaces and the ensuing fretting-fatigue risk. It introduces a three-layer workflow: a macro-scale FEM with a baseline CoF, a meso-scale DEM to generate a friction database for rough contacts, and a neural-network regression to predict mesoscopic CoF from surface topology and microscopic parameters for FEM enrichment, using a macro coefficient of $\mu_{macro}=0.11$ and exploring $\mu_{micro} \in [0.01,0.09]$. The approach is validated by DEM-derived data, achieving a prediction error of about $6.31\%$ on unseen configurations and revealing a potential 25% reduction in peak stresses when enrichment is applied, depending on micro-roughness. The framework provides a path to more accurate, topology-aware stress and fatigue predictions for wheel-axle systems and can be extended with wear evolution and experimental data to strengthen reliability assessments.
Abstract
The presence of surface defects (roughness, surface imperfections, profiles, etc.) in a contact inevitably leads to the modification of its local properties, such as the coefficient of friction. In railway wheelsets, this surface condition is crucial as it dictates appropriate fatigue design for the final use. However, these local phenomena are not well understood and require a real step back. Therefore, the aim of this paper is to propose a multiscale numerical strategy to better understand these phenomena. The multiscale strategy is divided into two steps. Initially, an analysis by the Discrete Element Method (DEM) modelling the interaction of generated rough surfaces is carried out to determine the coefficient of friction. In a second step, the results of DEM are introduced into a structural calculation where the enrichment of the coefficient of friction is done on each finite element contact. Given the wide variety of potential surface defects (size, distribution, height, etc.), a large number of DEM simulations is performed. A specially developed deep learning program is then used to account for these dispersions. The application targeted in this paper is the fitting of a wheel on a railway axle.