Physics-Informed Machine Learning for Predicting Rubber–Surface Friction from Dynamic Friction Tester Measurements: Load–Speed Interaction, Load, Roughness, and Material Response

Loading...
Thumbnail Image

TR Number

Date

2026-09-11

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

The interaction between tires and the road is one of the main factors determining how safely and effectively a vehicle can operate. Predicting this friction, however, is far from straightforward because it depends on several variables, such as structural and material properties of tire, road roughness, tire slip, applied load, temperature, and velocity, that interact in complex and nonlinear ways. Existing models can handle simplified conditions but often fall short when applied to real driving situations. This thesis addresses the problem through physics-informed machine learning. Approximately 600 steady-state friction observations produced by a Dynamic Friction Tester (DFT), collected by Su [8] at the CenTiRe Laboratory at Virginia Tech, were assembled and used to evaluate a range of machine learning techniques. Eight rubber compounds were tested across three standardized abrasive surfaces spanning coarse to fine texture. Random Forest, Support Vector Regression, Artificial Neural Network, and a Hybrid Stacking Ensemble were trained and compared against the measured friction coefficients. Because the measurements were made on abrasive papers rather than on pavement, the friction coefficients span a wider and higher range than a tire encounters in service; the models are therefore calibrated to a DFT-on-abrasive regime rather than to a road regime. The central contribution of this work lies in how the learning problem is posed rather than in the volume of data analyzed. Raw experimental inputs such as sliding velocity and viscoelastic loss modulus are coupled through the excitation frequency relationship 𝜔 ≈ 𝑞𝑉s described in Persson’s viscoelastic friction theory [2,59–61]. The feature basis was therefore re-expressed through two mechanism-guided interaction terms: Fn×Vs, a power-dimensional load–speed interaction, and Ts×E″, a strength–damping coupling formed from the compound’s maximum tensile strength and viscoelastic loss modulus. The physical interpretation of the first term is deliberately limited: Fn×Vs is not the actual frictional power delivered at the interface, which is 𝐹f𝑉s= 𝜇𝐹n𝑉s. Su [8] likewise expresses sliding work using friction force, 𝑊 = 𝐹f𝑠. The second term is also an engineered interaction rather than a derived physical quantity; its units, Pa², are not those of energy or dissipation. Models trained on this reparameterized feature set achieved high accuracy while producing feature-importance rankings that are interpretable in tribological terms. Overall, the study shows that a feature basis built from tribological reasoning yields accurate friction predictions whose importance structure is interpretable in mechanistic terms. A matched raw-feature control was subsequently trained on {Dp, Ts, Fn, Vs, E″, grit} under the same split, preprocessing and hyperparameters, and all four architectures were re-evaluated under leave-one-compound-out validation. The control shows that the composite terms are not universally more accurate: across twenty repeated splits they give the Random Forest a small but reproducible gain (ΔR² = +0.005, p = 0.024), while the raw basis is better for SVR and ANN and indistinguishable for the Hybrid Ensemble. Compound-holdout validation shows that the Random Forest predicts entirely unseen formulations with pooled R² = 0.921, including R² = 0.858 when trained on compounds A–G and tested on the withheld compound H. The claim advanced here is therefore one of interpretability, physical consistency and demonstrated material generalization rather than of universal predictive superiority. The framework is a step toward friction models that can be audited against theory as well as against data.

Description

Keywords

Rubber-surface friction, Dynamic Friction Tester, Physics-informed machine learning, Feature engineering, Load-speed interaction, Normal load, Surface roughness, Viscoelastic response, Friction coefficient prediction, Tribology

Citation

Collections