EN / TR
AI Form Finding Research
04 - BUILDINGActive2024-2025

AI Form Finding

What if machines could dream buildings?

Scale 04 Building
Samples 50.8k images
Training 72 GPU-hrs
Team 3 researchers
Tools PyTorch A100
Status Active Research
01

Frei Otto spent years with soap bubbles. Gaudi hung chains upside down. Form-finding has always been slow.

We asked: what if we could compress that process? What if a machine could explore thousands of forms in the time it takes an architect to sketch one?

So we trained neural networks on 50,847 architectural images and 3D models. Not to copy existing buildings, but to learn the underlying logic of structure. Then we asked the networks to generate new forms.

The results surprised us. The AI didn't just remix what it had seen. It found forms we hadn't imagined. Branching structures that look like coral. Shells that twist in ways no human would draw. And 68% of them actually work. They pass structural analysis. Physics accepts them.

AI-generated architectural forms with structural validation

400 Forms in 2 Hours: Green overlay indicates structural validation. The AI generates faster than we can evaluate.

02

Theoretical Framework

01

Training Data

50,847 images and 3D models from ArchDaily, university archives, and our own projects. Tagged by structure type, material, and span.

02

Structural Feedback

Every generated form goes through Karamba FEA. If it fails, that failure teaches the network. Over time, the AI learns physics.

03

Speed

200 forms per hour on a single GPU. That's a design space no human team could explore manually.

04

Material Efficiency

Because the AI optimizes for structure, generated forms often use 20-30% less material than conventional designs.

03

Research Process

01

Curate Data

50,847 images and 3D models, tagged by typology, structure, and material

02

Train Network

StyleGAN3 for 72 GPU-hours on NVIDIA A100s with architectural conditioning

03

Validate Structure

Every generated mesh goes through Karamba FEA. Failures become training signal.

04

Human Selection

Architects guide the process with sketches, sliders, and iteration

04

Research Phases

01

Dataset Curation

Six months collecting, cleaning, and tagging 50,847 architectural samples. Most AI projects fail here. We didn't.

02

Network Training

StyleGAN3 on four A100 GPUs for 72 hours. We tried diffusion models too, but GANs were faster for iteration.

03

Physics Integration

Connecting the latent space to Karamba FEA. Now the network gets feedback on structural validity in real time.

04

Product Deployment

Wrapping this into Archly.ai so architects can use it without touching Python.

05

Key Metrics

50,847
Training Samples
Largest architectural dataset
68%
Pass Rate
Structurally valid forms
200
Forms/Hour
Generation speed
72 hrs
Training Time
Four A100 GPUs
06

Key Thinkers

01

Frei Otto

German Architect, 1925-2015

Otto spent decades with physical models, soap films, and hanging chains. He proved that optimal forms emerge from material behavior, not drawing. Our AI compresses his life's work into hours.

02

Mario Carpo

Architectural Historian

Carpo distinguishes between 'digital' and 'computational' design. Digital means drawing on a computer. Computational means letting the computer design. Our work is computational.

03

Zaha Hadid Architects

Pioneering Parametric Practice

ZHA showed that curved, flowing forms could be built at scale. Our AI extends their parametric language into territory even they haven't explored.

04

Ian Goodfellow

ML Researcher, GAN Inventor

Goodfellow invented GANs in 2014. Without his breakthrough, none of this would be possible. We adapted his framework for structural constraint satisfaction.

07

Case Studies

ITU Biomimetic Pavilion

Istanbul Technical University

The first built structure from our AI pipeline. A coral-inspired branching canopy that passed structural analysis on the first try. Built in 2024 as a proof of concept.

AI-generated Type
89% Material Efficiency

Dubai Bridge Competition

Dubai Design Week 2024

We generated 1,247 pedestrian bridge variants in one afternoon. The jury selected a form no human would have drawn. Second place.

2nd place Result
1,247 Variants

Archly.ai Form Engine

Commercial Product

The research became a product. Architects type constraints in plain English. The AI generates validated options. Currently in beta with 340 users.

Beta Stage
340 Users

Comparative Analysis

GANs

Fast but Risky

Generator versus discriminator. Produces forms quickly, but can get stuck repeating itself. Needs careful tuning.

FastAdversarialMode Collapse Risk

Diffusion Models

Slow but Reliable

Builds forms by gradually removing noise. Higher quality, more diversity, but takes longer to generate.

High QualitySlowDiverse

Neural Radiance

Not Really Generative

NeRF reconstructs existing spaces from photos. Great for documentation, but doesn't invent new forms.

ReconstructionDocumentationNot Creative

Our Approach

GAN + Physics

We combine fast generation with real-time structural feedback. The physics engine rejects bad forms before you see them.

Physics-AwareValidatedFast
05

Optimization Results

100% 75% 50% 25% 0%
68%
55%
12%
3%
Fine-Tuned GAN
Diffusion Model
Naive GAN
Random Noise

What percentage of generated forms can actually be built?

08

Key Findings

01

The AI finds forms between styles. Feed it Gothic cathedrals and Zaha Hadid towers, and it interpolates. The results are structurally valid hybrids that exist in no historical category.

23 novel typologies
02

AI-generated forms often discover non-intuitive load paths. Diagonal bracing patterns that reduce steel by 12-18%. The network optimizes for physics, not aesthetics.

12-18% material saved
03

Latent space navigation feels like time travel. You can morph continuously from tower to bridge to shell. The journey itself suggests forms.

4 seconds to interpolate
04

Physics feedback makes everything better. GANs with structural validation produce forms 30% more efficient than GANs alone.

30% efficiency gain
09

Honest Limitations

Computational Cost

Mode collapse is real. The network sometimes fixates on towers. We don't fully understand why.

Computational Cost

Interpretability is limited. We can't always explain why a particular latent vector produces a good form.

Data Dependency

Scale blindness. The model struggles with consistent scale unless we explicitly condition it.

Behavioral Assumption

Fabrication gap. Forms are optimized for structure, not for how you'd actually build them.

10

Conclusion

Machines can dream buildings. Not copies of what they've seen, but genuinely new forms that physics accepts. With 68% structural validity and 12-18% material savings, AI-assisted design isn't a future prospect. It's working now.

Limitations

  • Mode collapse requires intervention
  • Interpretability remains limited

Future Directions

  • Real-time physics-aware generation
  • Direct-to-fabrication pipeline