A novel neural network model for rolling linear guide pair optimization design

This paper addresses the inefficiency of conventional design methods for rolling linear guide pairs, which rely on finite-element analysis and large data demands. This study pioneers a few-shot learning framework based on meta-learning, employing a model-agnostic meta-learning strategy to train an inverted-bottleneck residual fully connected network. The network achieves high prediction accuracy ( R 2 >0.91) with only 126 samples. An integrated parametric platform reduces modelling time from 50 to 2–3 min, significantly improving efficiency. Optimization via sequential least squares quadratic programming demonstrates a 57 % reduction in vertical deformation. However, the current work focuses on static performance optimization, leaving dynamic aspects for future research. This approach offers an efficient and data-effective paradigm for precision mechanical component design.