A hybrid Conv-LSTM network with skip connections for nonlinear reduced-order modeling of spatiotemporal flow fields

High-fidelity numerical simulation of complex unsteady fluid dynamics remains computationally intensive, limiting its practicality for real-time prediction and repeated evaluation scenarios. To address this challenge, this paper proposes a hybrid deep learning framework for nonlinear reduced-order modeling of spatiotemporal fluid dynamics, where a skip-connected convolutional autoencoder is integrated with a long short-term memory (LSTM) network, referred to as Conv - LSTM . The convolutional encoder extracts fine-scale spatial features, while the LSTM module captures long-term temporal dependencies within a low-dimensional latent space. Skip connections enhance feature reuse and help retain multiscale flow structures, improving both model convergence and predictive accuracy. The proposed Conv-LSTM framework is validated on three representative flow problems: flow past a cylinder, global sea surface temperature dynamics, and isotropic turbulence. Across all numerical cases, Conv-LSTM outperforms both the traditional dynamic mode decomposition (DMD) and a baseline Conv-LSTM without skip connections, yielding lower reconstruction errors and higher prediction fidelity. Moreover, Conv-LSTM exhibits strong robustness against noise and maintains reliable performance under various perturbation levels. These results demonstrate the Conv-LSTM framework's ability to generalize across diverse flow regimes and its potential as a scalable, data-driven solution for efficient and accurate modeling of complex spatiotemporal flow systems. • Hybrid Conv-LSTM framework: A novel reduced-order model combines skip-connected Conv-AE with LSTM for spatiotemporal flow prediction. • Enhanced feature retention: Skip connections improve spatial detail preservation and training stability in complex flow modeling. • Superior accuracy and robustness: Outperforms DMD and baseline models across multiple spatiotemporal flow cases, even under noisy conditions. • Broad applicability: Successfully predicts vorticity, velocity, and temperature in periodic, climatic, and turbulent flows.