Generating solution output for PDE takes longer than generating solution about methodoflines.jl HOT 2 OPEN

I guess I should link to the system that's causing the actual problem: https://data.earthsci.dev/dev/geosfp/#Using-data-from-GEOS-FP

This is also related to this PR: #240

from methodoflines.jl.

I understand that a lot of this code is in MethodOfLines.jl instead of this package, but the observed function that's highlighted in the screenshot above is in this package so I'm posting the issue here. In the larger system that this issue is trying to represent, observed takes up a much larger portion of the overall time (like, pretty much all of it), and it seemed like what was happening was that a new observed function had to be generated and compiled for each variable at each grid point, and then type interference had to be run each time it was called, and that seemed to be taking up a lot of time.

So I guess my question is whether this seems like a possible explanation for the problem, and if so if there's a way to statically type the observed function and/or possibly use one function that gets called for each variable rather than generating a new function for each variable.

It's possible, and somewhat related to the issue of how lowering currently occurs via scalars and gets O(n^2) amount of code on 2D PDEs. Function folding is required there, and that's the reason for large compile times, and is the biggest piece of work right now that is going on in MTK. One that is cleaned up, then I think preserving structure in observation functions and Jacobians is next. But it's all one thread of preserving structure in the code generator.

from methodoflines.jl.