Physics-informed graph transformer network for predicting cable-stayed bridge structural deflection response

doi:10.1016/j.aei.2025.103897

Physics-informed graph transformer network for predicting cable-stayed bridge structural deflection response

2025-09-25

0

PRE

AI

DOI：10.1016/j.aei.2025.103897

Original

Original request for help

Save

Abstract

En 中文

Accurately predicting the deflection response of complex structures such as cable-stayed bridges remains a significant challenge. To address this, this study proposes a physics-informed graph transformer network (PiGTN). The PiGTN leverages a physics-informed adjacency matrix derived from structural mechanics to encode sensor dependencies. An enhanced graph convolutional network (GCN) with second-order convolution captures spatial correlations, while a transformer encoder models temporal dynamic. The training process employs a mixed loss function that integrates data-driven loss with physical “soft constraints” from the Euler-Bernoulli beam solution, ensuring predictions conform to physical laws.

Physics-informed graph transformer network for predicting cable-stayed bridge structural deflection response

Abstract

Journal

Organization