Hydrodynamic Characterization of Bodies of Revolution through Statistical-Empirical Prediction Modeling Using Machine Learning

Abstract

Machine learning algorithms, namely artificial neural network modeling, were used to create prediction models for force and moment coefficients of axisymmetric bodies of revolution. These prediction models had highly nonlinear functional relationships to both geometric parameters and inflow conditions, totaling five input factors. A uniform experimental design was created consisting of 50 design points in these five factors and dictated which test points to simulate. Data was generated using computational fluid dynamic simulations, which were performed on all geometries using NavyFOAM at the experimental conditions prescribed by the designed experiment. The prediction models were validated by comparing behavioral trends in responses to previous research conducted by the author on a similar geometry. A test data sets was also created and used to ensure that the prediction models were not overfit to the training data and that they could accurately predict arbitrary geometries and inflow conditions within the experimental design region. Once the prediction models were validated, they were used to study the effects of varying the geometric parameters, inherent to the experiment, on each of the force and moment coefficients. Multidisciplinary optimization (MDO) schemes used in the early concept design phases for aero/hydrodynamic vehicles often use simplified planar maneuvering characteristics based on empirical or analytical relations in order to limit the computational cost of maneuverability prediction. This method leaves a more detailed analysis of the maneuvering behavior of a design to later in the process, where improvement or correction of an adverse behavior may be difficult to implement. The analysis of out-of-plane conditions or combined pitch-yaw conditions especially, are usually relegated to the detail analysis phase as empirical/ analytical descriptions of these conditions are lacking in the literature. It is therefore desired to develop a method to move these more detailed maneuvering analyses forward in the design phase.