Automated hydrodynamic shape optimization using neural betworks

This paper presents an optimization method integrating automated CAD-based Computational Fluid Dynamics (CFD) and Neural Networks for hydrodynamic shape optimization, and its applications to underwater hull configurations. Unlike in classical optimization methods where direct CFD computations are used in the optimization loop, the present method uses trained Neural Networks to evaluate the hydrodynamic performance of the configuration throughout the optimization process. Here, the CFD tool is used to generate the data sets used for training the Network. Both the classical optimization and Neural Network approaches are presented and applied to the design/optimization of one of Pacific Marine's advanced lifting bodies. Results are compared and show that the NN approach can produce better designs at substantially lower computational costs than the classical approach. For the example treated with 28 design variables where total CPU time was the limiting factor, a 34% improvement with the NN approach is obtained when the classical approach only yields a 26% improvement using five times more CPU time. Areas of future research are outlined.