다중 스케일 위상최적화를 위한 조건부 적대적 생성 신경망을 통한 재료 표현

Abstract

In this paper, a novel material representation method for multi-scale topology optimization is proposed. The number of design variables of every microstructure reduces by the generator network. The generator is trained together with the discriminator simultaneously in an adversarial way. Some of the condensed design variables are applied as conditions of the generative networks to control the microstructure much easier than without any condition. These conditions also make the generated samples be uniformly distributed without augmentation of the training data. The isotropic microstructure is tested, and the result shows the effectiveness of the proposed method. By this method, geometric constraints are not necessary in the optimization phase.