AIRSHIP HULL OPTIMIZATION USING ANN & CFD SIMULATIONS

This project focused on airship hull shaping for minimum drag.  A method employing the artificial neural network (ANN) to speed up numerical optimization was investigated and applied to computation fluid dynamic (CFD) simulation for drag minimization.  The aerodynamic drag of the airship hull directly impacts the engine power requirement and is an important aspect of airship design.  Three-dimensional analysis using CFD simulations produces the best estimation of the drag.   However, direct application of CFD to hull shape optimization may require formidable computation time since a single CFD iteration can take hours and thousands of iterations may be required to reach a satisfactory solution.  To reduce the number of optimization iterations, the ANN was used to predict the best input for the next CFD iteration based on previous CFD outputs.  The open-source CFD program OpenFOAM was used for the drag simulations.  The simulations encompassed a three-dimensional bare hull in a free stream of incompressible air flow.  The hull diameters at various points along the axis from the nose to the tail were used as optimization variables. The hull shape contour was obtained by connecting the discrete diameter values using cubic spline interpolation.  The CFD program came with large library of Reynolds-averaged simulation (RAS) turbulence models and the standard high-Re k-w model was included in the current study.  A number of airship hull shapes were generated and their drag coefficients simulated to provide an initial training data set for the ANN.  A nonlinear optimization program searched the input parameter space of the ANN for a hull shape that has the lowest drag as predicted by the ANN.  The new hull shape parameters were again fed to the CFD program to calculate the drag.  The shape parameters and the drag were then added to the training data set to retrain the ANN.  This process was repeated until a minimum drag was reached.  The project was carried out in various stages.  In the first stage, CFD simulations on NACA Model 111 airship hull was performed and results compared to literature reports to validate the CFD model.  In the second stage, ANN topologies with various hidden layer and node configurations were studied and the best one was selected based on small training and validation data sets.  In the third stage, the ANN and the CFD were used in conjunction to optimization the hull shape.   Initial results indicated that the ANN could reduce the number of computation iterations used for hull shape optimization using CFD.