Architectural Shape Generation Driven By Deep Learning In Conceptual Design Stage

Schematic design is the soul of architectural design.Building mass is an important content of the conceptual design,which have a great influence on the deepening of future architectural solutions and even the entire architectural design process.Along with the continuous improvement of the requirements for the efficiency and effect of architectural design,the demand for new architectural design tools is also increasing.This research,based on the latest progress of artificial intelligence technology,introduces deep learning to drive architectural mass generation in the architectural conceptual design stage to better assist architects in scheme creation.This research aims to realize architectural design automation with intelligent technology means;assist architectural design with the human-computer cooperative design mode,bring diversity to architectural mass exploration and improve design efficiency and design effect.The research summarizes the contents and research gap of building mass design and architectural generative design and analyzes the technical characteristics and technical advantages of deep learning.By studying the mathematical logic in the mass design of conceptual schemes,the paper introduces deep learning to drive the building mass generation,and constructes the corresponding deep neural network model.Also,the research summarizes and proposes the design process and three technical paths to help the architects deal with the key problems of the conceptual design stage.The research integrates the main software and develops the QDG-system to assist the architect according to the proposed system framework based on the deep neural network model.Besides,this paper also focuses on establishment process,data classification basis and selection method of the building mass database for training the mass generation platform developed.This research focuses on the three paths proposed: the basic building mass generation design based on image translation,the deepening building mass generation design based on image completion and the deepening building mass generation design based on style transfer.This paper analyzes the principle of every technical path to realize the design function,develops the corresponding deep neural network as the driver to realize the specific functional requirements,introduces the algorithm principle,program framework and condition parameters of each deep neural network in detail,elaborates the training and employed method of each program,and summarizes the design process of each technical path realizing building mass generation.In this paper,the application and verification of the method proposed and the system developed are carried out in combination with the specific design tasks.The design process and generated results proves that the design method and the platform system have good design effect and design efficiency.The usability of the system and method was verified by SUS questionnaire survey.It is hoped that this research can provide useful enlightenment and reference for the field of intelligent architectrual design,computational design theory and technical tool development.