A data-driven computational framework for non-intrusive reduced-order modelling of turbulent flows passing around bridge piers

Repetitively conducting high-fidelity numerical simulations under varying conditions is often a crucial requirement in the optimisation design of offshore bridges and structures. Reduced-order modelling (ROM) provides an efficient approach to quickly and reliably obtain solutions by extracting low-dimensional representations from full-order numerical systems. This paper presents a novel data-driven computational framework for non- intrusive ROM of turbulent/unsteady flows passing around bridge piers, consisting of two interconnected components: the Stacked Autoencoder (SAE) and the Dynamic Mode Decomposition (DMD). The novelty lies in utilising SAE to achieve nonlinear dimensionality reduction by projecting the full-order dynamical system onto a low-dimensional latent space, followed by constructing reduced-order models through data-driven DMD to represent fluid dynamics in the latent feature space. This new SAE-DMD-based method is applied to develop reduced-order models for two unsteady flow problems, and it is also compared with classical DMD and highfidelity numerical simulations in terms of modelling accuracy, forecasting efficiency and memory requirements. The results demonstrate that the proposed method can rapidly offer reliable predictions while significantly reducing memory usage and it exhibits excellent extrapolation capability by accurately preserving primary nonlinear characteristics of fluid dynamics. This new method shows potential to overcome computational challenges associated with high-resolution numerical modelling for complex large-scale flow problems.