Adjoint Sensitivities don't work properly for domains that aren't ConstantTPDomain about reactionmechanismsimulator.jl HOT 4 CLOSED

Do you have an MWE on this? I'd like to take a look. Tape compilation requires that there aren't branches. When you say "rate coefficients change", that makes me thing you have a callback changing parameters, in which case you might need to use ReverseDiffAdjoint().

from reactionmechanismsimulator.jl.

Yes, so I've been digging through this the last few days. When I originally implemented I mostly focused on InterpolatingAdjoint, I was aware of the issues with ReverseDiffVJP(true), but that was the only thing that worked properly straight out of the box and while there is branching related to temperature and pressure I thought it would work in more cases that it seems to. ReverseDiffVJP(false) gives me NaNs, which I haven't investigated yet. TrackerVJP() gave me errors, which I've been working on recently. With TrackerVJP after fixing some array concatenation issues I've run into StackOverflowErrors, some of the rate coefficients aren't calculated vectorized, and I suspect through the operations they're causing the vectorized ones to convert to Array{TrackedReal}.

I've attached a jupyter notebook MWE, feel free to take a look. Honestly though I feel like at this stage I'm just trying to figure out how the Reverse mode tools are interacting with the code and trying to make sure it can handle the different types properly.

ConstantVH2Combustion_adjoint_MWE.ipynb.zip

from reactionmechanismsimulator.jl.

Interesting, I'll have to catch up on this in a bit. I wonder if those NaNs are true adjoint NaNs (as in bugs), or from numerical accumulation. It'll take some digging.

from reactionmechanismsimulator.jl.

Closing with merging of #95

from reactionmechanismsimulator.jl.