@article{Xia2026, 
author = {Yichun Xia and Hui Cao and Yonggang Meng},
title = {A cU-Net-based surrogate model for predicting rough surface contact parameters in line-contact friction},
year = {2026},
journal = {Friction},
keywords = {line-contact friction, contact parameters, digital surrogate model, rough surface},
url = {https://www.sciopen.com/article/10.26599/FRICT.2026.9441313},
doi = {10.26599/FRICT.2026.9441313},
abstract = {Accurate and efficient prediction of contact parameters on rough surfaces is critical for the tribological design, condition monitoring, and failure analysis of line-contact components. Although physics-based models offer high prediction fidelity, their multi-field coupled iterative solvers are computationally expensive and complex, limiting practical engineering deployment. This work proposes a digital surrogate model based on the conditional U-Net (cU-Net) for predicting the spatial distributions of contact parameters, including contact temperature, pressure, and heat flux, on rough surfaces in line-contact friction. Real surface topographies are augmented via the Generative Patch Nearest-Neighbor (GPNN) model, and a wide-ranging training dataset is constructed through batch computation using a boundary lubrication physics-based model. The cU-Net incorporates an operating condition embedding mechanism to enable unified prediction across multiple operating conditions. Multi-scale spatial and channel attention modules are further integrated to enhance cross-scale feature extraction. A four-component progressive composite loss, assembled from established loss terms and tailored to the sparse extreme-value characteristics of rough surface contact parameters, is combined with Bayesian hyperparameter optimization and a three-stage progressive training strategy. On completely unseen rough surfaces the model achieves mean relative errors of 3.53%, 12.32%, and 13.66% for the temperature, pressure, and heat flux fields, respectively, with peak and top-5% errors that are lower still, and it is approximately 105 times faster than the physics-based model on the same workstation. Out-of-distribution tests quantify the degradation of accuracy under extrapolated operating conditions. Retraining with datasets generated from the corresponding physics-based models is expected to extend the framework to other lubrication regimes.}
}