A Physics-Embedded convolutional neural network for Tire-Pavement dynamic contact characteristic and friction prediction

Tire–pavement friction is critical for driving safety yet direct measurement across a road network is operationally expensive. Existing data-driven predictors either treat texture-to-friction as an opaque regression that demands large datasets and offers no mechanistic insight or rely on physics-informed soft penalties. This work proposes an innovative Physics-Embedded Convolutional Neural Network (PE-CNN) that embeds tire rubber–texture contact mechanics directly into the network architecture as a fully differentiable and latent physics layer to structurally guide the network. The network predicts microscale deformation ratio to construct a dynamic contact envelope that represents the effective rubber-pavement contact. The effective texture is then processed through a fully tensorized two-dimensional spectral analysis and hysteresis integral to predict friction. A ResNet-18 backbone augmented with a Convolutional Block Attention Module (CBAM), and friction-loss gradient propagates uninterrupted through the physics layer back into the CNN weights. On a comprehensive laboratory dataset of texture–friction measurements spanning different asphalt mixture types, five-fold cross-validation within the development set was used to select model parameters. The PE-CNN achieves R 2 of 0.822 on independent testing data, compared to R 2 of 0.75 for the direct CNN and R 2 of 0.744 for a global-cut-plane method with the same physics-embedded pipeline, emphasizing the contribution of dynamic contact envelope. A trajectory-level evaluation across progressive tire polishing further demonstrates that the predicted effective-contact field evolves coherently with surface wear. The framework provides an inspectable connection between evolving pavement texture, effective tire-pavement contact, and friction deterioration.