| title: VAE-FDM Explorer | |
| emoji: 🏗️ | |
| colorFrom: blue | |
| colorTo: red | |
| sdk: docker | |
| app_port: 7860 | |
| pinned: true | |
| license: mit | |
| tags: | |
| - structural-engineering | |
| - form-finding | |
| - jax | |
| - three-js | |
| - force-density-method | |
| short_description: Structural form-finding with differentiable mechanics | |
| # VAE-FDM Explorer | |
| 3D interactive tool for structural form-finding. A neural encoder predicts force densities and a mechanical decoder (Force Density Method) produces the equilibrium shape. | |
| Drag control points or adjust sliders to explore shell geometries. All shapes satisfy mechanical equilibrium — the FDM decoder solves it directly. | |
| Built on top of the work by Pastrana et al., ICLR 2025. | |
| ## How it works | |
| - **Backend**: FastAPI serving a JAX neural network (~2ms inference on CPU) | |
| - **Frontend**: Three.js with real-time 3D interaction | |
| - **Physics**: Force Density Method guarantees equilibrium for every prediction | |
| ## Links | |
| - [Paper (arXiv)](https://arxiv.org/abs/2409.02606) | |
| - [Source code](https://github.com/efradeca/vae-fdm) | |
| ## Configuration | |
| Check the [Spaces Docker SDK reference](https://huggingface.co/docs/hub/spaces-sdks-docker) for more details. | |