The durability evaluation of modern shoe material foaming media has long relied on physical testing and faces technical bottlenecks such as long cycles and difficulty in revealing microscopic failure mechanisms. In response to this problem, the research aims to explore an evaluation path that takes into account local morphological quantification and long-range force dependence analysis and constructs an intelligent durability prediction model that integrates the window transformer attention mechanism and micro topological dimensionality reduction representation. The results show that in the application of anti-fatigue topology optimization, the model successfully suppressed the median stress of the foam matrix from 8.7MPa to 3.4MPa and strictly controlled the local extreme value at 6.1 MPa, reducing simulated stress concentration. Facing a severely distorted scene with a porosity of 70%, the mean absolute error was stably suppressed at 5.1k cycles, and network-only inference required 32.6 ms, demonstrating excellent fault tolerance and computational economy. In addition, in the long-term dynamic variable load verification, the predicted deformation rate at the end of life (9.80%) was highly consistent with the FEA reference level (9.90%), showing close agreement with the simulated loading trajectory. The research has opened up the cross-scale entity mapping from porous physical network to macroscopic mechanics and established an intelligent digital template with great transformation value for high-throughput virtual concurrent screening of shoe material formulas and anti-fatigue reverse auxiliary engineering.
