Data-driven prediction of cylinder-induced unsteady wake flow
Shicheng Li, James Yang, Penghua Teng
Abstract
Understanding cylinder-induced wake is pivotal in fluid dynamics, providing essential insights for the design and analysis of various structures, including offshore platforms, bridges, and buildings. To achieve fast and accurate modeling, this study introduces a novel reduced-order model (ROM) utilizing dynamic mode decomposition (DMD) and an advanced deep learning framework, specifically an attention-enhanced convolutional neural network-long short-term memory networks model (CNN-LSTM), for predicting cylinder-induced unsteady wake flows. The DMD efficiently simplifies complex fluid systems while retaining key dynamics, thus significantly saving computational costs. By leveraging its combined strengths, the CNN-LSTM with an attention mechanism effectively captures complex spatiotemporal features. The resulting ROM accurately reproduces the wake processes around a cylinder (group), demonstrating high consistency with computational fluid dynamics (CFD) solutions (coefficient of determination > 0.98), and showcases satisfactory resilience to a (Gaussian) noise level of up to 25 %. This study contributes a robust ROM capable of handling spatiotemporal dynamics, facilitating swift prediction of future outcomes using historical data, which is particularly critical for efficient real-time analysis and informed decision-making in dynamic settings, e.g., digital twins and predictive maintenance.