Deep autoencoder for off-line design-space dimensionality reduction in shape optimization

In shape optimization, design improvements significantly depend on the dimension and variability of the design space. High dimensional and variability spaces are more difficult to explore, but also usually allow for more significant improvements. The assessment and breakdown of design-space dimensionality and variability are therefore key elements to shape optimization. A linear method based on the principal component analysis has been developed in earlier research to build a reduced-dimensionality design-space, resolving the 95% of the original geometric variance. The paper presents an extension of the method to more efficient nonlinear approaches. Specifically, the use of a deep autoencoder is presented and discussed. The method is demonstrated for the design-space dimensionality reduction and hydrodynamic optimization of the hull form of a USS Arleigh Burke-class destroyer. A comparison with the linear method is finally shown and discussed.