VIV-SDE-Net: A physics-informed neural framework for long-horizon prediction of bridge vortex-induced vibrations

Abstract

Long-horizon prediction of vortex-induced vibration (VIV) remains a major challenge in the safety assessment of long-span suspension bridges, as existing data-driven approaches often lack physical interpretability and provide unreliable uncertainty estimates. To address this gap, this study proposes a two-branch neural network physically guided by the stochastic differential equation (SDE) of the VIV dynamical system, which is named VIV-SDE-Net, to predict a VIV index and to quantify its predictive uncertainty. The VIV index is derived from the fundamental VIV equation and can be described by an SDE that consists of a deterministic component and a stochastic random-walk excitation to reflect the evolution of VIV events. Field validation using real monitoring data from a long-span suspension bridge shows that the proposed method accurately predicts moderate and severe VIV events up to 10 and 8 min in advance, respectively. Comparative experiments further demonstrate that the proposed approach outperforms a state-of-the-art Gaussian Process Regression model across different forecast horizons, providing more accurate and stable predictions as well as more reliable uncertainty intervals. These results highlight that the proposed VIV-SDE-Net has strong potential for real-time VIV early warning and broader applicability in physics-informed learning for wind–structure interaction problems.