Deep learning-based prediction of optimal surface texturing parameters for water-lubricated SiC

Surface texturing has emerged as a crucial strategy for enhancing tribological performance at lubricated mechanical interfaces. However, current texture design primarily relies on heuristic approaches, lacking systematic guidelines and exhibiting randomness and subjective dependencies. This study presents a Residual Attention Network (RAN) methodology for optimizing composite textures on water-lubricated silicon carbide (SiC) surfaces. The framework synergistically integrates advanced deep learning architecture with high-fidelity computational fluid dynamics simulations to accurately predict key tribological performance indicators. Model validation demonstrates that our RAN-based approach exhibits superior prediction accuracy and stability compared to traditional Multi-Objective Bayesian Optimization (MOBO). Friction experiments confirm that RAN textures approach super-lubricity levels, achieving a 75.7 % reduction compared to MOBO textures.