search

tfp-photonics / neurop_invdes

Neural operator surrogates for electromagnetic inverse design
MIT License
31 stars 6 forks source link

gdsfactory integration? #1

Closed joamatab closed 1 year ago

joamatab commented 1 year ago

Great work!

how about integrating some of this work into gdsfactory?

we already support Meep's adjoint optimization and would be great to expand with your work

we also added your paper and github link to awesome photonics list

yaugenst commented 1 year ago

Hey, thanks for including it in the list, that's cool!

As for the gdsfactory integration - sounds interesting. Although I am a bit unsure what this would look like in practice? It seems to me like gdsfactory is geared towards on-chip devices and mode calculations, this is not really a setting that we were considering in this work. I guess it is straightforward to train some models for planar devices, but the mode calculations would need to be performed somewhere else.

If you are happy with just poking around a bit with free-space scatterers, then you are welcome to use the pre-trained models that we provide - the most accurate ones are the 2D nets trained on 16k samples. The inverse design procedure would need to be tweaked a bit to support 2D though.