Efficient Propeller Wake Prediction Model: Transformer-SMFANet and Its Application

Accurate prediction of propeller wake is essential for understanding fluid mechanics related to propeller performance. This paper introduces the Transformer-SMFANet neural network (TSNN) to predict the spatiotemporal evolution of propeller wake. The TSNN model consists of two modules: Transformer, which captures temporal relationships, and SMFANet, which captures local and global spatial dependencies. A gating mechanism integrates these features into a unified representation. High-fidelity E779A propeller wake flowfields, simulated using OpenFOAM, serve as training data. Results show that TSNN exhibits an “error step phenomenon,” where errors within the same batch remain similar, while those between batches form a step-like pattern. Despite this, TSNN effectively predicts propeller wake evolution with strong consistency to CFD results. The optimal mean-squared error for instantaneous wake prediction is on the order of [Formula: see text], while for the average wake, it is on the order of [Formula: see text], demonstrating excellent predictive performance. The proposed deep-learning-based model provides a novel approach for propeller wake prediction and has potential applications in vibration reduction and noise suppression via flow control technology.