A U-net based reconstruction of high-fidelity simulation results for flow around a ship hull based on low-fidelity inviscid flow simulation

Abstract

In this study, neural networks are trained to transform inviscid simulation data for flow around a ship hull into data representative of viscous flow simulations. The objective is to provide high-fidelity viscous flow simulation data using machine learning in conjunction with inviscid flow simulation results, which are significantly less time-consuming to generate. This approach has the potential to accelerate high-fidelity flow simulations by a factor of more than 100, enabling simulation-based design for ship hulls with numerous repetitive cases. To create the training dataset, a variety of hull forms are generated from six baseline hull forms using parametric modification function techniques. Inviscid and viscous flow data for each hull are obtained through potential flow analysis and computational fluid dynamics - simulations, respectively. The neural network structure and hyperparameters are subsequently optimized through parametric studies. The trained neural networks are then employed to predict viscous flow simulation data based on inputs comprising inviscid flow data and hull form geometry. The results demonstrate that the neural networks successfully predicted both the pressure distribution around the hull and the free surface elevation. Notably, the ability to predict the free surface elevation is significant, given that inviscid flow simulations inherently lack this capability. Additionally, the neural network's dimensionality reduction feature is utilized to visualize how the flow and hull form data were clustered within the latent space based on baseline hull forms and ship speed.