Teaching LLMs how to solid model
a year ago
- #AI
- #CAD
- #LLMs
- LLMs can create CAD models for simple 3D mechanical parts and are expected to improve rapidly.
- An AI mechanical engineer could handle tasks like material selection, design for manufacturing, and CAD model improvement.
- Training generative models on existing CAD files is a promising approach, with research focusing on diffusion and transformer architectures.
- LLMs can generate solid models using programmatic interfaces like OpenSCAD, leveraging their code generation capabilities.
- OpenSCAD allows users to write scripts that render into solid CAD models, making it suitable for LLM-driven modeling.
- LLMs can utilize general mechanical engineering knowledge, such as correctly sizing holes for bolts.
- An evaluation pipeline was developed to test LLMs' ability to generate CAD models via OpenSCAD, comparing generated STL files to reference models.
- Geometric checks in the evaluation include volume difference, bounding box alignment, chamfer distance, and Hausdorff distance.
- Frontier models like Sonnet 3.5 and Sonnet 3.7 show significant performance improvements in CAD modeling tasks.
- Start-ups like AdamCad and Zoo.dev are developing text-to-CAD products, though initial evaluations show LLMs generating OpenSCAD code outperform Zoo.dev's API.
- Future improvements for LLM-driven CAD modeling include better spatial reasoning, iterative debugging tools, and parameter adjustment UIs.
- GenCAD is expected to become a common tool for mechanical engineers within 6-24 months, with long-term potential to revolutionize CAD suites.