Abstract
The forecasting of high-fidelity, transient fluid flows remains a formidable computational challenge due to the high-dimensional and non-linear nature of the data produced by numerical simulations. The study presents a novel hybrid deep-learning framework to address this challenge by combining a Variational Autoencoder (VAE) with a Long Short-Term Memory (LSTM) network for the efficient spatiotemporal prediction of flow fields around a NACA 0012 airfoil. The VAE first learns a compressed, low-dimensional latent representation of the high-dimensional flow data, which includes pressure and velocity components, achieving effective dimensionality reduction while preserving essential spatial features. An LSTM network then learns the temporal dynamics within this latent space to forecast future states. The model was trained and validated on a dataset of 300 time-step snapshots, successfully generating a 15-step-ahead forecast. Results demonstrate the model’s high accuracy, with a key metric – the maximum velocity magnitude – predicted with only a 3.8 % error compared to the ground truth simulation data. This study validates the VAE-LSTM architecture as a powerful and computationally efficient tool for forecasting complex fluid dynamics, offering significant potential for applications in real-time control and design optimisation where rapid prediction is critical.