EN / TR
← Back to Research Lab
02 — BUILDING Active 2024-2025

AI Form Finding

Machine Learning for Architectural Morphogenesis

AI Form Finding Research
GAN-generated architectural forms interpolating between structural typologies

Form-finding has traditionally been the domain of intuition, precedent, and physical modeling. Frei Otto spent years with soap films and hanging chains. We asked: can we compress that process into hours?

This research trains Generative Adversarial Networks (GANs) and diffusion models on 50,847 architectural images, 3D models, and structural analysis outputs. The goal is not to replace the architect but to expand the design space—to generate forms that a human might never conceive but that physics will accept.

Our key finding: AI-generated forms that pass structural validation (FEA analysis) occur at a 68% rate after fine-tuning, vs. 12% for naive generation. The remaining 32% fail due to cantilever instability, torsion, or material stress exceedance. These failures are themselves informative—they reveal the boundaries of structural feasibility that the model is learning.

Research Process

01

Dataset Curation

50,847 images and 3D models scraped from architectural databases, categorized by typology (tower, bridge, shell, frame), period, and performance metrics. Each entry tagged with structural system, span, and material.

Python Scrapy Blender
50,847 samples
02

GAN Training

StyleGAN3 trained for 72 GPU-hours on NVIDIA A100s. We tested 3 architectures: vanilla StyleGAN3, custom structural-conditioned variant, and a diffusion model (Stable Diffusion fine-tuned). The structural-conditioned GAN outperformed others on FEA pass rate.

PyTorch NVIDIA A100 Weights & Biases
72 GPU-hours
03

Structural Validation

Generated forms converted to mesh → voxel → FEA model pipeline. Karamba3D evaluates for displacement, stress, and buckling. Pass criteria: max displacement < L/500, max stress < 80% yield. Current pass rate: 68% (up from 12% before fine-tuning).

Karamba3D Grasshopper Python
68% pass rate
04

Human-AI Collaboration

Architects guide generation via sketch input (edge detection → latent space projection), constraint sliders (symmetry, cantilever ratio, aperture density), and iterative feedback loops. Average iteration time: 4.2 seconds per variant.

Archly.ai Grasshopper Rhino
4.2s per variant
AI-generated biomimicry architectural form

Latent Space Navigation: GAN-generated organic form emerging from neural network interpolation. Multiple ghosted iterations show the evolution from structural typology to biomimicry-inspired architecture.

What We Discovered

01

GANs can interpolate between Gothic tracery and Zaha Hadid-style organicism, finding structurally valid hybrid forms in the latent space

23 novel typologies identified
02

AI-generated forms often discover non-intuitive load paths—diagonal bracing patterns that human designers rarely propose but that reduce steel by 12-18%

12-18% material reduction
03

Latent space navigation enables real-time 'what-if' exploration: architects can morph between tower → bridge → shell in continuous space

~4s interpolation time
04

Combining AI generation with physics simulation produces structures that are 30% more efficient than GAN-only outputs (measured by stress/mass ratio)

30% efficiency gain

Current Limitations

Mode collapse: the model occasionally fixates on specific typologies (especially towers), requiring manual intervention

Interpretability: we cannot fully explain why certain latent vectors produce structurally superior forms

Scale blindness: the model struggles to maintain consistent scale relationships without explicit conditioning

Fabrication gap: generated forms are optimized for structural performance, not for manufacturing feasibility

Explore AI-Assisted Design?

Join our beta program for Archly.ai or discuss how AI can accelerate your design process.