Generative Design by Reinforcement Learning: Enhancing the Diversity of Topology Optimization Designs

Generative design refers to computational methods that can automatically conduct exploration under constraints defined by designers. Among many approaches, topology optimization-based generative designs aim explore diverse designs, which cannot be represented conventional parametric approaches. Recently, data-driven optimization research has started exploit artificial intelligence, such as deep learning or machine learning, improve the capability of exploration. This study proposes a reinforcement (RL) based process, with reward functions maximizing diversity designs. We formulate sequential problem finding optimal parameter combinations in accordance given reference design. Proximal Policy Optimization is used framework, demonstrated case an automotive wheel problem. To reduce heavy burden process required our RL formulation, we approximate neural networks. With efficient data preprocessing/augmentation and architecture, networks achieve generalized performance symmetricity-reserving characteristics. show RL-based produces large number within short inference time exploiting GPU fully automated manner. It different from previous approach using CPU takes much more processing involving human intervention.