A machine-learning assisted topology optimization approach is proposed for architectural design with artistic flavor. This work establishes a novel framework to systematically integrate structural topology optimization with subjective human design preferences. To embed artistic flavor into the design, neural style transfer technique is adopted for measuring and generating the prior knowledge from a reference image with concerned artistic flavor. With the use of different convolutional layers in the VGG-19 (Visual Geometry Group) model-based CNN (Convolutional Neural Network), both style and content of the artistic flavor from low to high levels of abstraction can be constructed. Then, the measured knowledge can be integrated into pixel-based topology optimization as a formal similarity constraint. Both 2D and 3D problems are solved to illustrate the effectiveness of the proposed approach where inheritance of artistic heritage can be achieved in a systematic manner.

提出了一种机器学习辅助拓扑优化方法，用于具有艺术气息的建筑设计。这项工作建立了一个新颖的框架，系统地将结构拓扑优化与主观的人类设计偏好相结合。为了将艺术气息嵌入到设计中，采用神经风格迁移技术从具有相关艺术气息的参考图像中测量和生成先验知识。在基于VGG-19（Visual Geometry Group）模型的CNN（Convolutional Neural Network）中使用不同的卷积层，可以从低到高的抽象层次构建出既有风格又有艺术气息的内容。然后，可以将测量的知识作为形式相似性约束集成到基于像素的拓扑优化中。