Alec Hammond

One thing we should do, is survey and establish the most popular API for FMM/RCWA tools, and essentially "clone it," that way we can maximize traction and engagement. Then, we...

Got it, thanks! Is there a reason the `angular_frequency` needs to account for additional `k-points`, but `in_plane_wavevector` is agnostic to the number of wavelengths, and `permittivity` is agnostic to both?...

> Although simpler logic might be possible, I think it would not be as general. Imagine a case where a user wants to simulate certain wavelength/angle combinations. Ah yep this...

Here's some good info: https://github.com/facebookresearch/fmmax/blob/c6cd855fceb26fda5234ec704d1ea0cd0cbff6c9/src/fmmax/vector.py#L113-L130

Yep this sounds like a good plan to me. How hard do we anticipate the manual adjoint will be?

@mfschubert have you tried [profiling](https://jax.readthedocs.io/en/latest/profiling.html) eg using perfetto or tensorboard to see what functions are hanging? It does a good job even with really complicated stacks. I've profiled entire optimization...

> By the way, would you be willing to share some of that profiling info? Here's some profiling I did on a problem that essentially combined the `microlens.py` example with...

A simple example comparing `vmap` to `pmap` would be useful here.

@luochenghuang you've got some good insight here. Have you explored any of these other approaches? Would you be interested in exploring them?

We had some issues with this when wrapping Meep's adjoint solver with jax. Jax had trouble compiling the graph when the FOM was vector-valued and we had a custom vjp...