How well can modern AI agents produce engineering designs for optics? Well – I wanted to find out.
To test this, I’ll walk through my goal. I need to get a lens made and I need it fast. I also need it to work properly. Lastly, I need it to be manufacturable at scale which means using glass substitutes often found in China.
So – using Claude Chat, I took performance specs from several existing lenses, handed them over and added:
“Make me a lens design that meets these general performance criteria.”
I also asked :
- Use Chinese available glass (e.g. K series)
- Validate the design using a raytrace
- produce estimated performance data
- produce design drawings
- produce a prescription
- produce a ZMX file
Here is the design it produced, it sure doesn’t look bad so far!
After a while it decided to build it’s own python analysis tool using multiple metrics, it did the job and generally things look ok. However – is this actually OK? Let’s drop the lens design into Zemax.
Using a COTS 200mm doublet as a comparison, I pulled a image, plot, ray fan and Huygens spot to inspect. All look quite similar, the spot size can be realized a bit better, but consider the aperture on the reference file was slightly larger, so that means difference in NA.
AI Doublet Left — COTS Design at right
Ray Fan Diagram
Note there is a difference here, this design is not optimized – with an optimized second surface, it’s obvious a lot of improvement can be made. Yet for a commercial type lens, I think this is a good example of a sufficient result – it’s less than the diffraction limit, which is a suitable bar for pass failure.
Huygens spot size is below, again note the difference in the psf due to optimization missed.
Finally, we can compare the simulated image formation – here I don’t think there is a clear winner, which is what we are shooting for in a commercially produced result. Very cool.
