Attention-enhanced spatiotemporal surrogate modeling for rapid propeller wake prediction

High-fidelity numerical simulations of marine propeller wake inherently demand excessive computational resources due to complex nonlinear fluid dynamics. To address this severe bottleneck, a highly efficient spatiotemporal surrogate model integrating a three-dimensional convolutional neural network (CNN) with a standalone squeeze-excitation attention (SEA) mechanism is proposed. The convolutional encoder effectively reduces the dimensionality of the complex flow field data, while the attention block dynamically allocates weights across multiple physical channels to capture model temporal dependencies between historical snapshots. By excluding the recurrent module, the proposed CNN–SEA architecture achieves lower prediction errors for most flow variables than the evaluated recurrent configurations under the present settings. The training process is significantly accelerated utilizing an automatic mixed precision (AMP) framework combined with a cosine annealing with warm restarts (CAWR) strategy. Evaluated against large eddy simulation (LES) reference data, the proposed model accurately captures intricate wake topologies under extreme operating conditions. The proposed CNN–SEA architecture improves prediction accuracy for the highly dynamic flow component. Additionally, the neural network achieves an acceleration ratio exceeding four orders of magnitude during forward inference. This massive computational reduction enables real-time hydrodynamic evaluation and establishes a robust data-driven framework for marine propeller design optimization.