Recent AI-based prediction models have been developed to directly learn CFD simulation results, rather than being limited to predicting integrated hydrodynamic coefficients. Unlike conventional surrogate models commonly used in simulation-based design...
Recent AI-based prediction models have been developed to directly learn CFD simulation results, rather than being limited to predicting integrated hydrodynamic coefficients. Unlike conventional surrogate models commonly used in simulation-based design, these approaches enable real-time prediction of not only integral quantities such as resistance coefficients but also spatial distributions of pressure and wall shear stress. As a result, they are expected to be particularly useful in ship hydrodynamic design, where local pressure and flow distributions play a critical role.
In this study, a positional-encoding-based data processing method is proposed to address the reduced prediction accuracy of a conventional U-Net-based pressure distribution model in regions surrounding protruding appendages. The performance of the proposed approach is evaluated through comparative analysis. Simulation results show that the proposed U-Net with positional encoding (U-Net+PE) reduces both training and validation losses for surface pressure and wall shear stress on the sail compared to the CNN-only model, whereas losses increase for other parts.
To further reduce errors in the remaining parts, the projection grid seam is relocated from the 12 o’clock position to the 6 o’clock position during training.
With this modification, losses decrease for all components except for the wall shear stress on the main body, top rudder, and sail, where they remain slightly higher than those of the CNN-only model.