Reduced-order modeling of hydrofoil cavitating flows via proper orthogonal decomposition and attentive bidirectional long short-term memory

Unsteady cavitating flows are characterized by strong nonlinearity, multiphase interactions, and a broad range of temporal and spatial scales, posing major challenges for accurate and computationally efficient modeling. In this study, a hybrid reduced-order modeling framework is developed by integrating Proper Orthogonal Decomposition (POD) with a Bidirectional Long Short-Term Memory (BiLSTM) neural network augmented by an attention mechanism. POD is applied to extract dominant spatial modes from high-fidelity simulation data, while the BiLSTM-Attention network is trained to predict the temporal evolution of modal coefficients. The proposed model is validated on a canonical unsteady cavitation case involving a hydrofoil, where the first ten POD modes account for 73.5% of the total kinetic energy. The reconstructed velocity fields show strong agreement with the original Large Eddy Simulation (LES) data, accurately capturing key flow features such as cavity growth, shedding cycles, and wake dynamics, with low prediction error across multiple spatial locations and time steps. These results demonstrate that the hybrid POD-BiLSTM-Attention framework offers a robust and data-efficient approach for reduced-order modeling of cavitating flows, with potential applications in flow control, real-time simulation, and cavitation-induced vibration prediction.