<p>Rubber materials are widely employed in sealing, vibration damping, and dynamic contact engineering applications due to their exceptional elasticity and damping properties. The surface morphology roughness of rubber directly influences interfacial tribological behavior, sealing reliability, and fatigue life. Currently, characterization of surface roughness primarily relies on physical testing methods, which lack effective theoretical prediction approaches. However, physical testing significantly increases both temporal cycles and economic costs. To address the demand for efficient prediction of rubber surface roughness parameters, this study innovatively proposes a machine learning-based numerical prediction method. A comparative investigation was conducted on three machine learning models—artificial neural network (ANN), convolutional neural network (CNN), and recurrent neural network (RNN)—focusing on their core performance metrics and predictive capabilities. The results demonstrate that RNN exhibits significant superiority over ANN and CNN in core performance indicators. Regarding final prediction performance, RNN achieved high prediction accuracy rates of 93.6%, 92.9%, 91.88%, and 93.06%, outperforming ANN and CNN models with more precise and stable predictive accuracy. This model effectively resolves the issues of low efficiency and poor repeatability inherent in traditional contact measurement methods, thereby providing a novel approach for intelligent assessment of surface quality in rubber products.</p> Graphical abstract