Intelligent design of steel-concrete composite box girder bridge cross-sections based on generative models

To enhance the efficiency and accuracy of composite box girder bridge design and achieve rapid and high-precision cross-section design, an effective intelligent algorithm is imperative. However, the development of intelligent design for steel-concrete composite box girder bridges is constrained by data scarcity and the performance of existing generative models. This paper introduces a pre-trained Vision Transformer as an Image Encoder (E_I) to enhance generative models for bridge design. Firstly, a dataset of 350 bridge designs is constructed for training and evaluation. Then, enhanced Condition-Feature models are developed and compared with fundamental generative models. The results show that the Condition-Feature Variational Autoencoder Generative Adversarial Network performs best, demonstrating the effectiveness of E_I in intelligent bridge design. This paper fills the gap in intelligent bridge design, offers valuable insights for future engineering research, and showcase the potential and application prospects of deep learning in bridge design.