EN TR
AI Form Finding Research
Generative Systems04 - BUILDINGSimulation Study2024-2025

AI Form Finding

What if machines could dream buildings?

Scale 04 Building
Pipeline GAN+FEA physics loop
Speed 200 forms/hr
Tools PyTorch Karamba
Type Simulation Study
Updated 2026-07
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 assembled a simulation pipeline: generative networks fine-tuned on open architectural datasets and our own parametric studies, with a structural physics engine as referee. Not to copy existing buildings, but to learn the underlying logic of structure, and then to generate new forms.

The results surprised us. The networks didn't just remix what they had seen. They found forms we hadn't imagined. Branching structures that look like coral. Shells that twist in ways no human would draw. And in our runs, 68% of them pass structural analysis. To be clear about what this is: a simulation study. The numbers on this page describe our own test runs, not built work or an external benchmark.

AI-generated architectural forms with structural validation

400 Forms in 2 Hours: Green overlay indicates structural validation in our simulation runs. The pipeline generates faster than we can evaluate.

02

Theoretical Framework

01

Training Data

Open architectural datasets and our own parametric models, tagged by structure type, material, and span. Curation is the unglamorous half of the work.

02

Structural Feedback

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

03

Speed

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

04

Material Efficiency

Because generation optimizes for structure, forms in our runs often use 20-30% less material than baseline configurations.

03

Research Process

01

Curate Data

Open datasets and our own parametric models, tagged by typology, structure, and material

02

Train Network

Fine-tune generative backbones 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

Collecting, cleaning, and tagging open architectural datasets and our own parametric studies. Most generative projects fail here; it deserves the time it takes.

02

Network Training

Fine-tuning pretrained generative backbones on the curated set. We tried diffusion models too, but GANs were faster for iteration.

03

Physics Integration

Connecting the latent space to Karamba FEA. The network gets feedback on structural validity as it generates.

04

Product Deployment

Folding what holds up into Archly's generation workflows. Planned, not shipped; we say so honestly.

05

Key Metrics

68%
Pass Rate
In our simulation runs
12%
Naive Baseline
Before physics feedback
200
Forms/Hour
Single-GPU test setup
23
Novel Typologies
No historical category
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 pipeline is a digital descendant of his method.

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 aims at the second.

03

Zaha Hadid Architects

Pioneering Parametric Practice

ZHA showed that curved, flowing forms could be built at scale. Their built work sets the benchmark our generative experiments measure their variety against.

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

Where This Lives in Our Products

Constraint-driven generation in Archly

Planned

Archly today is an AI visualization workflow for architectural ideas. The form-generation pipeline in this study is the research track we intend to fold into it: describe a constraint, get structurally sane options back.

Archly →

Latent-space navigation in Grasshopper

Exploration

Falcon AI speaks Grasshopper. The latent-space experiments here inform how Falcon could propose parametric variations inside a live definition: morphing between options instead of rebuilding them.

Falcon AI →

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 structural feedback. The physics engine rejects bad forms before you see them.

Physics-AwareValidated in SimulationFast
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 passes FEA in our runs?

Scenario model: values from our own simulation runs

08

Key Findings

01

The networks find forms between styles. Feed them Gothic cathedrals and parametric towers, and they interpolate. The results are structurally valid hybrids that exist in no historical category.

23 novel typologies
02

Generated forms often discover non-intuitive load paths. In our runs, diagonal bracing patterns reduced 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. In our comparisons, 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. Passing FEA is not a construction document.

10

Conclusion

Machines can dream buildings. Not copies of what they've seen, but genuinely new forms that physics accepts. In our simulation runs: 68% structural validity, 12-18% material savings. The honest gap is between a form that passes FEA and a building you can detail; closing that gap is the work ahead.

Limitations

  • Mode collapse requires intervention
  • Interpretability remains limited

Future Directions

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