Multi-objective Optimization of Railway Transition Zones with Machine Learning: Application to Prefabricated Epoxy Asphalt Cured Track Bed

Transition zones in high-speed railways suffer from abrupt stiffness variations that induce irregular dynamic responses and accelerate infrastructure deterioration. This study presents a surrogate-assisted multi-objective optimization framework that combines finite element (FE) simulations, a neural network-based surrogate model, and the NSGA-II algorithm to address this challenge. A validated 3D FE model of prefabricated epoxy asphalt cured track beds (PEACT) was used to generate 341 layout scenarios covering 13 response parameters. These data were used to train a neural network, which served as a static surrogate predictor for evaluating layout performance during the optimization process. The results show that module layout has a limited effect on peak responses but significantly improves smoothness, with three categories of optimal configurations identified. Compared with direct FE-based optimization, the proposed framework achieves substantial computational efficiency and provides data-driven design guidance for PEACT transition zones. This framework exemplifies the potential of hybrid data–simulation approaches to enhance adaptive and efficient railway infrastructure design.