Form-finding with AI: from catenary to catenet
For the last decade, the bulk of my practice has been building shell structures — thin, form-active surfaces where geometry and structure are the same thing. A shell that looks right is a shell that is right. The trouble is finding it.
The old way was catenary form-finding: hang chains in tension, invert, cast in compression. Heinz Isler's studio was full of these. It works, but it's slow, and it only gives you one surface per physical rig.
A shell that looks right is a shell that is right.
What changed
Two things. First, differentiable geometry frameworks — you can now define a shell and its loads in code and let the gradients tell you where to push it. Second, diffusion models trained on millions of structural simulations. We've been running a small model internally that takes a footprint and a load brief, and returns a mesh that's within 8% of optimal, in about 4 seconds.
The catenet
We call the internal tool catenet. It's not a product — it's a design partner. I still do the final work in Rhino and Grasshopper. But catenet lets us explore a hundred shells over coffee, instead of one shell over a week.
What it means for Shwaas
Shwaas, our 2023 project near Mysore, is five shells on five different footprints. In the old workflow, that would have been five full form-finding studies — months of work. With catenet, the initial exploration took an afternoon. We then hand-tuned the final surfaces over two weeks. The savings are not in the final work. They're in letting us try more ideas before we commit.
What's next
I'm interested in closing the loop with the digital twin data we're now collecting at Shwaas. If the shells are sensing themselves — strain, temperature, rainfall — can the next shell be form-found against real-world behavior, not just simulated loads? I think so. More on that soon.
— Utsav