[slack] <chrisrackauckas> oh yeah, MassActionJump has something tricky here.
@ChrisRackauckas have you seen this paper? https://papers.nips.cc/paper/9308-demystifying-black-box-models-with-symbolic-metamodels.pdf
I would imagine embedding this in some kind of specialized neural ode architecture would be an interesting approach for discovering reaction kinetics from data and other dynamical systems
I wonder if this can be integrated with modelingtoolkit and diffeqflux
Feels like sindy with the dictionary being learnable as well
[slack] <isaacsas> We should really add a function that gets at this more easily.
[slack] <chrisrackauckas> it's known but not intentional
[slack] <chrisrackauckas> it needs a fix.
[slack] <torkel.loman> It was just much easier to make a method which solved it until the time point of the callback and took the values at the last time point as initial condition to another simulation, but with modified parameters.
[slack] <torkel.loman> It's not super pretty, but works well enough for my simple purposes.
(biggest problem is to run it in monte carlo, where I lose performance, but it is not of that much importance in this case)
(biggest problem is to run it in monte carlo, where I lose performance, but it is not of that much importance in this case)
[slack] <torkel.loman> But yes, some way of enabling one to modify these kinds of things with callbacks as easily for continuous problems should probably be a long term goal.
[slack] <torkel.loman> ok, cheers. I'll skip out on writing a report on it then
[slack] <chrisrackauckas> did you profile it?
[slack] <chrisrackauckas> same integration method?
[slack] <ashton.bradley> So I wast able to use identical method, but I figure TSit5 should represent pretty well (?)
[slack] <chrisrackauckas> Tsit5 is only for non-stiff equations
[slack] <chrisrackauckas> Use Trapezoid if you want to compare to Trapezoid?
[slack] <chrisrackauckas> what's he parallelizing? The Jacobian calculation?
[slack] <chrisrackauckas> Did you use auto-sparsity and all of that?
[slack] <ashton.bradley> I didn’t use auto-sparsity… maybe I should try that next
[slack] <chrisrackauckas> use finite differencing
[slack] <chrisrackauckas> autodiff is the issue, not Trapezoid
[slack] <ashton.bradley> do you mean DiffEqOperators ?
[slack] <chrisrackauckas> no
[slack] <chrisrackauckas>
Trapezoid(autodiff=false)
[slack] <ashton.bradley> (it is already a finite difference method)
[slack] <ashton.bradley> ah, ok.
[slack] <chrisrackauckas> and are you also comparing the integration error?
[slack] <ashton.bradley> no, but it is imaginary time, so just trying to find a ground state.