Predicting ship hull flow-field distributions using a soft-constrained ANN model

Accurate determination of shear and pressure force distributions around a ship's hull is paramount for hydrodynamic optimization tasks, as integrating these fields across the hull's surface provides the total drag force applied on the hull. While Computational Fluid Dynamics (CFD) provides this capability, it is often limited by high computational cost and time-consuming pre-processing, post-processing, and simulation times. The challenge is further amplified during design exploration studies, where simulations are performed across multiple operational conditions. To address these limitations, we propose a soft-constrained Multitask deep neural network, named HydroForceNet, which serves as a surrogate model for CFD simulations on marine vessel hulls. Our proposed architecture can accurately predict pressure and shear distributions on various Wigley-based geometries and calculates the resistance components, using three-dimensional geometric and operational inputs, at a fraction of the computational cost of a traditional CFD evaluation. Finally, to further illustrate its applicability, the proposed artificial neural network is integrated into a genetic algorithm-based optimization task, producing a new hull geometry with a 15.77 % reduction of hydrodynamic resistance compared to a reference hull geometry, after evaluating over 2500 designs within 2 min, while faithfully reproducing the flow field.