A differentiable physics-embedded framework for thermally informed Neural-UniTire modeling and identification

Abstract

To address the need for thermally informed tire modeling for handling-oriented force prediction under limited thermal sensing, this paper presents a differentiable physics-embedded sequential framework for thermally informed Neural-UniTire modeling and identification. First, a differentiable radial thermal model is developed from a continuous radial thermal formulation through finite-volume Crank–Nicolson discretization, with hybrid global, physics-parameterized, and neural thermal coefficients used to reconstruct radial multilayer thermal states from sparse boundary measurements. Then, a Physics-Embedded Neural-UniTire model is developed by embedding neural networks into the UniTire framework and extending it to unsteady conditions through a second-order hysteresis element. Finally, a thermally informed correction mechanism is introduced for the Physics-Embedded Neural-UniTire model, in which reconstructed radial multilayer thermal states with operating-condition information are used to generate neural thermo-correction factors for stiffness- and friction-related parameters. Experiments on an MTS flat-track test rig with a passenger-car tire and a semi-slick tire under pure lateral-slip and longitudinal-slip conditions show good agreement with the measured temperature and force responses. The radial thermal model achieves accurate temperature estimation from sparse boundary measurements. The unsteady Physics-Embedded Neural-UniTire model improves force prediction over the steady-state formulation, especially in the nonlinear slip region and phase lag. The thermally informed model improves force prediction over the baseline and sparse-input ablation models under pulse-excitation conditions, while the learned thermal correction remains physically interpretable.