Abstract: At the micro level, the friction phenomenon can be viewed as a collision process between asperities. A collision model for a pair of hemispherical asperities made of U71MnG was established based on the finite element method (FEM). This study extensively examined the impact of the lateral offset of asperities on the normal and tangential forces during the collision process. To accurately predict the relationship between the tangential and normal forces during the lateral collision of asperities, a nonlinear mapping model was constructed, relating normal and tangential forces to interference, offset, and the collision process. Given that the finite element computation is time-consuming, a subset of representative cases was analyzed using FEM, while the remaining cases were predicted by a trained Back Propagation Neural Network (BPNN). Additionally, hyperparameters of the BPNN were optimized using the Physics-Informed Nutcracker Optimization Algorithm (PINOA), effectively preventing overfitting and significantly improving the model's predictive accuracy. The results show that under five interference levels of 10%, 15%, 20%, 25%, and 30% R, the tangential force, normal force, and the corresponding friction coefficient for a single asperity increase with interference. At a fixed interference level, the friction coefficient decreases with increasing offset distance. Finally, the shoulder-contact collision model of the asperity was mapped to macroscopic surfaces using a statistical model, yielding a functional expression suitable for predicting real-world surface friction coefficients. The propose modeling framework coupled with machine learning achieves high-fidelity predictions of asperity contact responses within a relatively short computation time, providing a practical pathway for the design and prediction of friction behavior.