Reconstruction of physical field characteristics of underwater vehicle wake based on data-driven approach

The characteristics of underwater vehicle wakes are intricately linked to multiple parameters, such as speed, depth, and environmental factors. Obtaining wake characteristic information for various operating conditions solely through numerical simulation methods can result in computational space explosion, rendering the time and computational resource costs prohibitive. This paper harnesses the robust image processing capabilities of convolutional neural networks and incorporates strategies such as attention mechanisms, dilated convolution techniques, and multi-scale feature fusion to design and construct a neural network architecture. Through a data-driven approach, it reconstructs multiple physical wake field characteristics resulting from underwater vehicle, including underwater velocity fields, surface divergence fields, surface kelvin wake, and surface thermal wakes. The study establishes a “black box” mapping between relevant parameters and the physical fields of wakes. The results demonstrate that the constructed network model achieves high accuracy in capturing both the macroscopic structures and pixel-level details of various physical fields. In comparison with the truth-values, the average normalized root mean square errors for the underwater velocity field, surface divergence field, surface kelvin wake, and surface thermal wakes are 6.10%, 3.40%, 8.21%, and 10.96%, respectively. The average structural similarity index values are 0.955, 0.966, 0.923, and 0.904, respectively. The predicted results closely match the truth-values for each physical field characteristic, effectively addressing the challenges of nonlinearity and multi-scale feature extraction in predicting complex flow fields, offering support for the rapid forecasting of multi-dimensional and multi-physical field characteristics of underwater vehicle wakes.