Efficient Identification and High-Fidelity Extrapolation of the Parameters in LuGre Frictional Model at Rubber-Metal Interface Utilizing Machine Learning and Physics-Informed Prior Model with Limited Experimental Dataset

Abstract

Accurate and efficient prediction of friction behavior at rubber-metal interfaces is essential for advanced robotic and mechanical systems, yet remains challenging due to nonlinear friction characteristics and limited experimental data. A framework integrating machine learning, optimization algorithms, and physics-prior model was proposed for efficient identification and reliable extrapolation of LuGre model parameters within limited experimental dataset, thereby saving time and experimental cost. A hybrid Gaussian Process Regression-Particle Swarm Optimization (GPR-PSO) method was developed to identify static and dynamic parameters within a limited normal load range (2∼10 N). Furthermore, a physics-prior extrapolation model based on Greenwood-Williamson contact theory combined with Gaussian Process Regression was developed to ensure physically consistent parameter evolution under unmeasured conditions. The rubber-aluminum tribo-pair was employed as a representative case featured as styrene-butadiene rubber/natural rubber (SBR/NR) block sliding against aluminum plate was investigated on a pin-on-disk test rig to verify the proposed framework, where the rubber specimen and the aluminum counterpart had surface roughness of approximately 4.6 µm and 0.27 µm, respectively. The results demonstrated that the identified friction responses agreed well with measurements, achieving R2 values above 0.91 and NRMSE values below 12%. Even with a 50% extrapolation beyond the training load range, the proposed framework maintained high predictive accuracy (R2 > 0.93, NRMSE < 10%). This study provides an efficient and physically interpretable methodology for friction parameter identification and extrapolation under limited-data conditions.