Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Using Generative AI to Explore Termite-Inspired Architectural Forms.

Daneshvar Kakhki, Shadi LU (2025) ASEM01 20251
Department of Architecture and Built Environment
Abstract
This thesis explores the potential of generative AI to recreate biological spatial systems, using termite mound structures as inspiration. Termite mounds, known for their unique internal networks and passive climate regulation, feature complex spatial logics that challenge traditional modeling techniques. This project investigates whether diffusion-based generative models can be trained to mimic such structures, and how their outputs might be reconstructed into meaningful 3D geometries.

Multiple custom datasets were created from a single termite mound model, capturing sectional views in three formats: open-view, capped surface, and cavities-focused. Each dataset was paired with both simple and descriptive captions to test the impact of... (More)
This thesis explores the potential of generative AI to recreate biological spatial systems, using termite mound structures as inspiration. Termite mounds, known for their unique internal networks and passive climate regulation, feature complex spatial logics that challenge traditional modeling techniques. This project investigates whether diffusion-based generative models can be trained to mimic such structures, and how their outputs might be reconstructed into meaningful 3D geometries.

Multiple custom datasets were created from a single termite mound model, capturing sectional views in three formats: open-view, capped surface, and cavities-focused. Each dataset was paired with both simple and descriptive captions to test the impact of semantic detail on training outcomes. LoRA models were trained using Stable Diffusion 1.5 and Hunyuan Video, and different generation workflows were tested using ComfyUI. Outputs were evaluated based on spatial continuity, form coherence, and biomimetic resemblance.

The results show that clean, abstract datasets paired with well-structured captions improve model performance in producing consistent, spatially connected sequences. Using a reverse-slicing workflow and CT-scan software, selected outputs were reconstructed into 3D meshes and fabricated as physical prototypes. A final comparison of nine trained models revealed key differences in tunnel connectivity, mass-void distribution, and overall architectural logic.

This research demonstrates that AI-driven workflows, when carefully structured, can bridge the gap between 2D generative outputs and 3D architectural constructs. It also highlights current limitations in fully mimicking termite-inspired spatial systems, offering insights for future explorations in generative design and digital fabrication. (Less)
Please use this url to cite or link to this publication:
author
Daneshvar Kakhki, Shadi LU
supervisor
organization
course
ASEM01 20251
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Generative AI, Diffusion Models, LoRA(Low-Rank Adaptation), LoRA Fine-Tunning, biomimicry in Architecture, Termite Mound Architecture, AI-Driven Design, AI-based Form Finding, Image-to-3D Reconstruction
language
English
id
9196460
date added to LUP
2025-06-11 08:56:56
date last changed
2025-06-11 08:56:56
@misc{9196460,
  abstract     = {{This thesis explores the potential of generative AI to recreate biological spatial systems, using termite mound structures as inspiration. Termite mounds, known for their unique internal networks and passive climate regulation, feature complex spatial logics that challenge traditional modeling techniques. This project investigates whether diffusion-based generative models can be trained to mimic such structures, and how their outputs might be reconstructed into meaningful 3D geometries.

Multiple custom datasets were created from a single termite mound model, capturing sectional views in three formats: open-view, capped surface, and cavities-focused. Each dataset was paired with both simple and descriptive captions to test the impact of semantic detail on training outcomes. LoRA models were trained using Stable Diffusion 1.5 and Hunyuan Video, and different generation workflows were tested using ComfyUI. Outputs were evaluated based on spatial continuity, form coherence, and biomimetic resemblance.

The results show that clean, abstract datasets paired with well-structured captions improve model performance in producing consistent, spatially connected sequences. Using a reverse-slicing workflow and CT-scan software, selected outputs were reconstructed into 3D meshes and fabricated as physical prototypes. A final comparison of nine trained models revealed key differences in tunnel connectivity, mass-void distribution, and overall architectural logic.

This research demonstrates that AI-driven workflows, when carefully structured, can bridge the gap between 2D generative outputs and 3D architectural constructs. It also highlights current limitations in fully mimicking termite-inspired spatial systems, offering insights for future explorations in generative design and digital fabrication.}},
  author       = {{Daneshvar Kakhki, Shadi}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Using Generative AI to Explore Termite-Inspired Architectural Forms.}},
  year         = {{2025}},
}