3D Ship Hull Design Direct Optimization Using Generative Adversarial Network

Abstract

The direct optimization of ship hull designs using deep learning algorithms is increasingly expected, as it proposes optimization directions for designers almost instantaneously, without relying on complex, time-consuming, and expensive hydrodynamic simulations. In this study, we proposed a GAN-based 3D ship hull design optimization method. We eliminated the dependence on hydrodynamic simulations by training a separate model to predict ship performance indicators. Instead of a standard discriminator, we applied a relativistic average discriminator to obtain better feedback regarding the anomalous designs. We add two new loss functions for the generator: one restricts design variability, and the other sets improvement targets using feedback from the performance estimation model. In addition, we propose a new training strategy to improve learning effectiveness and avoid instability during training. The experimental results show that our model can optimize the form factor by 5.251% while limiting the deterioration of other indicators and the variability of the ship hull design.