Capturing the dichotomous behaviors of asperity wear with neural-network surrogate modeling

We explore the use of neural networks as surrogate models for atomistic simulations of rough surface sliding. While molecular dynamics (MD) simulations provide valuable insights into the tribological behavior of rough surfaces at the asperity level, their high computational cost limits their scalability, making it challenging to analyze macroscale phenomena arising from microscale interactions. To address this limitation, a neural network-based surrogate model is developed to replace the physics solver in atomistic simulations. Our approach involves representing the state of MD simulations as grid-based images containing occupancy and virial stress information, which serve as input to a U-Net-type neural network trained to predict the next simulation state. The performance of models trained with different loss functions is compared, highlighting trade-offs between sharpness and shape conservation. A key finding is that training with the mean absolute error (MAE) loss produces more defined but smoother boundaries for wear particles and cracks, while training with the mean squared error (MSE) loss results in predictions that are more accurate with angularity but blurrier. Post-processing techniques are tested to enhance prediction sharpness. While the surrogate models manage to learn some of the complex interactions between asperities and to predict wear particle formation, recursive application of the models reveals challenges due to accumulated errors, suggesting the need for a more complete state representation and different neural network structure. The findings provide a foundation for integrating data-driven surrogate models into multiscale tribological simulations. • Neural networks are explored as surrogate models for computationally expensive molecular dynamics simulations of rough surface sliding contact. • A U-Net architecture is trained on grid-based images of occupancy and virial stress to predict simulation states, with comparisons of MAE and MSE loss functions revealing trade-offs in prediction sharpness and shape accuracy. • MAE-trained models produce sharper boundaries for wear particles and cracks but smoother shapes, while MSE-trained models offer better angularity at the cost of blurriness; post-processing enhances sharpness. • Surrogate models capture complex asperity interactions and wear particle formation, but recursive application exposes accumulated errors, indicating needs for improved state representation and network design. • The work establishes a foundation for integrating data-driven surrogates into multiscale tribological simulations, bridging microscale and macroscale phenomena.