Space Growth :A Multi-Step Neural Framework for 3D Architectural Completion using RAG and Voxel Transformers
(2026) ASEM01 20261Department of Architecture and Built Environment
- Abstract
- Architecture exists at the intersection of art and engineering, materially defined by the
spaces enclosed within physical boundaries. Just as natural language processing allows
machines to parse written text, artificial intelligence can comprehend architectural space
when it is translated into a structured, objective format. This thesis proposes a novel
machine learning framework for the automated understanding and generation of 3D
architectural environments. By utilizing voxel grids as a fundamental spatial syntax, we
introduce a bespoke Retrieval-Augmented Generation (RAG) system driven by a
Transformer architecture. The model generates structures iteratively through two
distinct stages: a "Bridge" phase, which logically connects... (More) - Architecture exists at the intersection of art and engineering, materially defined by the
spaces enclosed within physical boundaries. Just as natural language processing allows
machines to parse written text, artificial intelligence can comprehend architectural space
when it is translated into a structured, objective format. This thesis proposes a novel
machine learning framework for the automated understanding and generation of 3D
architectural environments. By utilizing voxel grids as a fundamental spatial syntax, we
introduce a bespoke Retrieval-Augmented Generation (RAG) system driven by a
Transformer architecture. The model generates structures iteratively through two
distinct stages: a "Bridge" phase, which logically connects existing architectural
contexts to retrieved pattern seeds, and a "Grow" phase, which expands the building
envelope based on spatial constraints. Evaluated against rigorous geometric and
topological metrics—including bounding box leakage, void ratio, and spatial
continuity—the results demonstrate the system's capability to successfully learn
architectural rules and autonomously roll-out structurally coherent building spaces. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9250501
- author
- Cai, Zhifeng LU
- supervisor
- organization
- alternative title
- Space Growth
- A Multi-Step Neural Framework for 3D Architectural Completion using RAG and Voxel Transformers
- course
- ASEM01 20261
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- AI, Architecture, Space Machine Learning, Space Reasoning, RAG, Transformer
- language
- English
- id
- 9250501
- date added to LUP
- 2026-09-15 10:21:18
- date last changed
- 2026-09-15 10:21:18
@misc{9250501,
abstract = {{Architecture exists at the intersection of art and engineering, materially defined by the
spaces enclosed within physical boundaries. Just as natural language processing allows
machines to parse written text, artificial intelligence can comprehend architectural space
when it is translated into a structured, objective format. This thesis proposes a novel
machine learning framework for the automated understanding and generation of 3D
architectural environments. By utilizing voxel grids as a fundamental spatial syntax, we
introduce a bespoke Retrieval-Augmented Generation (RAG) system driven by a
Transformer architecture. The model generates structures iteratively through two
distinct stages: a "Bridge" phase, which logically connects existing architectural
contexts to retrieved pattern seeds, and a "Grow" phase, which expands the building
envelope based on spatial constraints. Evaluated against rigorous geometric and
topological metrics—including bounding box leakage, void ratio, and spatial
continuity—the results demonstrate the system's capability to successfully learn
architectural rules and autonomously roll-out structurally coherent building spaces.}},
author = {{Cai, Zhifeng}},
language = {{eng}},
note = {{Student Paper}},
title = {{Space Growth :A Multi-Step Neural Framework for 3D Architectural Completion using RAG and Voxel Transformers}},
year = {{2026}},
}