On the Prediction of Temporal Dynamics of Turbulent Flows using a Hybrid Temporal Architecture (TurboNet)

The accurate predictions of temporal dynamics in turbulent flow remains a formidable challenge in fluid mechanics, owing to the non-linear complexity of governing equations. Traditional numerical methods, while accurate, are computationally prohibitive for high-fidelity simulations over extended time horizons. This study presents TurboNet: a hybrid multi-layer temporal architecture combining Conv1D, LSTM, BiLSTM, that can model the temporal evolution of turbulent flows. To ensure robustness and generalizability, the proposed model is trained and validated on three distinct datasets comprising distinct flow regimes and physical characteristics: Periodic dynamics from in-house URANS simulations of Savonius wind turbine, wall bounded non-linear dynamics from DNS data of channel flow and open-boundary non-linear dynamics from LES data of backward-facing step flow. Evaluation metrics R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> ranging 0.97-0.99 and MSE peaking at 0.2463 demonstrate that the hybrid architecture effectively captures both the periodic and non-linear dynamics with high accuracy while significantly reducing computational cost. The results underscore the potential of ML based frameworks to complement the traditional CFD solvers, enabling efficient and scalable modelling of complex turbulent systems across diverse applications.