Differentiability

[jakubMitura14]

Hello is it possible to differentiate simulation for optimazation purposes ?

[rafaqz]

Its possible, there is even a differentiable branch where I used Zygote on small grids. I think there are better ways to do that though.

On master you can only use ReverseDiff.jl or similar.

But all options are pretty slow. We could work on making it faster I guess, but I havent had the time.

Raw simulations are fast enough to just run a million times and optimise without using gradients.

[jakubMitura14]

thank you !

[rafaqz]

With Enzyme.jl maturing we should be able to differentiate steps of simulations.

Rules should never allocate or have any type stability anyway so the only shortcomings of Enzyme do not apply in our case. Someone will just need to try it out, probably using step! manually rather than sim!.

[jakubMitura14]

Thank you for getting with the update!