Adjoint Solver Roadmap

[smartalecH]

The point of this post is to:

1. Consolidate all of the posted adjoint solver bugs. There are quite a few reported that are scattered in various issues and random PR comments. I was getting tired of searching through all of them. I plan to link to them here, mainly so we don't miss/duplicate any. Any progress on that bug should obviously be posted in its corresponding issue.
2. Provide a roadmap for future adjoint solver features. Many people have reached out, asking how they can contribute (woohoo!) unfortunately all of the planned "features" are scattered throughout the issues log and aren't really organized in any coherent way (e.g. some features must be implemented before others, some are significantly more complicated than others).

I'll continue to add to this, and we can use it as a discussion forum for abstract adjoint solver topics.

Current bugs

• Flaky DFT convergence - stop_when_dft_decayed in single precision #1831
• Incorrect gradients - Incorrect FourierFields gradients for D components #2090, Inconsistent finite-difference and adjoint gradients in EigenmodeCoefficient #2087, Finite-difference and adjoint gradients are inconsistent when the mode monitor is not in a waveguide #2060,Incorrect adjoint gradients with overlapping design variables #1984, incorrect adjoint gradient for point monitor of FourierFields objective function #1900, Simulation fields are NaN or Inf in the adjoint solver #1676, Backward propagating eigenmode adjoints are wrong #1585 Optimization failure with adjoint simulations #1471

Roadmap and planned features

• LDOS adjoints - LDOS adjoint #1401
• Poynting Flux adjoints - Adjoint optimization through Poynting Flux #1307
• Variable sidewall angles for material grids - variable sidewall material-grid #1636
• Pade extrapolation - Field extrapolation for faster adjoint calculations #1497