Exploration on Machine Learning Layout Generation of Chinese Private Garden in Southern Yangtze

Abstract

Machine learning has been proved to be feasible and reasonable in architectural field by extensive researches recently, whereas its potential is far from being tapped. Previous studies show that the training of GAN by labelling can enable a computer to grasp interrelationship of spatial elements and logical relationship between spatial elements and boundary. This study set the learning object as layout of private gardens in southern Yangtze with higher complexity. Chinese scholars usually analyse private garden layout based on their observation and experience. In this paper, based on Pix2Pix model, we enable a computer to generate private garden layout plan for given site conditions by learning classic cases of traditional Chinese private gardens. Through the experiment, taking Lingering garden as example, we continuously adjust the labelling method to improve learning effect. The finally trained model can quickly generate private garden layout and aid designers to complete scheme design with private garden element corpus. In addition, the working process of training GAN enables us to discover and verify some private garden layout rules that have not been paid attention to.

A Discussion on an Urban Layout Workflow Utilizing Generative Adversarial Network (GAN) - With a focus on automatized labeling and dataset acquisition

“…Previous research has shown that training a GAN with labeling allows the ML algorithm to learn the interrelationships of urban spatial elements (Liu et al, 2021). The DLS methods with GAN offer a new codesign possibility.…”

“…The architect inputs design decisions (labels) to meet the urban design task, and the GAN quickly generates urban design proposals based on the chosen labels. However, in the current DLS approach, most of the inputs to the GAN are existing physical conditions of the site, such as existing boundaries (Tian, 2020), entrances (Liu et al, 2021), and roadway networks (Shen et al, 2020). These DLS methods lack the subjective speculative labeling of architects as input, in addition, there is a need for a systematic and automated method to replace manual labeling.…”

“…The Generative Adversarial Network (GAN), a DL model, demonstrates huge potential for automated design and planning processes. GAN enables machines to learn the interrelationships of urban morphological features and urban layout logic by labeling (Liu et al, 2021). Although GAN models are adapted to quickly produce results similar to historical urban samples, it is still a challenge to integrate architects' decisions for producing solutions that meet urban future development requirements.…”

“…DLS is an important GAN data-processing process that can influence the way people collaborate with DL models. For example, architects can control the urban layout of GAN by inputting different labels, such as the road network (Shen, 2020) or the shape of urban elements, like ponds, paths, and site boundaries (Liu et al, 2021). However, these labelling rules, consisting only of simple physical constraints, make it difficult for the GAN to produce layout results that incorporate the architect's subjective speculations.…”

Exploration on Diversity Generation of Campus Layout Based on GAN

“…In recent years, machine learning technology has made some progress in the field of building layout generation [2,4,5,7]. With more complex research objects and ever larger scales, some scholars begin to explore the possibility of combining machine learning with campus planning and layout design.…”