Generative design has been instrumental in expanding designers' ability to create diverse alternatives. However, the current generative building morphology design presents two broad weaknesses. Firstly, it fails to consider the interaction between a design and its backdrop context, particularly in high-density metropolitan areas. Secondly, it fails to harness existing design knowledge embedded in existing designs. This paper aims to develop a data-driven generative design model: VmRF, which can learn from existing designs and generate plausible and contextual building morphologies. The model consists of a variational autoencoder (VAE) to compress high-dimensional building morphology datasets into low-dimensional building morphology datasets and a multivariate random forest (mRF) to identify explainable relationships between design parameters and morphology patterns. Performance evaluation shows the superiority of the VmRF model in terms of training speed and prediction fitness. Consequently, the proposed model promotes enhanced design efficiency, innovation in contextual awareness, and evidence-based decision-making in building morphology design.

中文翻译：

---

用于密集都市地区建筑形态的数据驱动的生成式上下文设计模型


生成式设计有助于扩展设计师创建各种替代方案的能力。然而，目前的生成式建筑形态设计存在两大弱点。首先，它没有考虑设计与其背景环境之间的互动，尤其是在高密度的大都市地区。其次，它未能利用嵌入在现有设计中的现有设计知识。本文旨在开发一种数据驱动的生成式设计模型：VmRF，它可以从现有设计中学习并生成合理且符合上下文的建筑形态。该模型由一个变分自动编码器 （VAE） 和一个多元随机森林 （mRF） 组成，前者用于将高维建筑形态数据集压缩为低维建筑形态数据集，后者用于识别设计参数和形态模式之间的可解释关系。性能评估显示了 VmRF 模型在训练速度和预测适应性方面的优越性。因此，所提出的模型促进了设计效率的提高、情境感知的创新以及建筑形态设计中的循证决策。