Predicting Tire Contact Pressure Using Neural Networks and a Physics-Based Prior: An Application to Tread Wear

Footprint mechanics, which govern most of the tire properties, are complex to model and expensive to measure. A physics-based numerical model requires extensive parameterization for modeling each subcomponent, while data-driven models are limited by the cost of extensive experimentation and may be unreliable beyond their data envelope. This work develops a hybrid predictor in which a physics-based tire contact model serves as an analytical prior. A small neural network is implemented that learns the cell/sensel level residual between the prior and the experimental data, collected using a Tekscan pressure pad. The prior consists of a flexible belt formulation with a lateral belt cross section that builds a contact patch shape and pressure distribution. The hybrid model is compared with a pure neural network model for baseline comparison. The designed validation protocol tests the prediction over sparse sections of the dataset, comprising leave-one-condition-out (LOCO), performance evaluation with a data holdout, extrapolation to the corner of the input parameter envelope, and data efficiency curves. It is observed that the hybrid matches or exceeds the performance of pure neural networks using a fraction of data, for the majority of the evaluated criteria, while producing physically constrained predictions. As an application, a simple Archard’s wear model is implemented to develop a proof-of-concept algorithm for estimating instantaneous local wear rates from predicted pressure fields over the vehicle duty cycles. This application reproduces camber-, load-, and pressure-dependent wear patterns while inheriting the data efficiency of the hybrid model.