Can create and simulate ODE without providing initial conditions for all variables

[TorkelE]

Pretty serious, leads to a silent and potentially nasty error.

# Fetch packages
using ModelingToolkitBase
using OrdinaryDiffEqDefault
using ModelingToolkitBase: t_nounits as t, D_nounits as D

# Create model.
@variables X1(t) X2(t)
@parameters k1 k2
eqs = [
    D(X1) ~ -k1*X1 + k2*X2
    D(X2) ~ k1*X1 - k2*X2
]
@mtkbuild sys = System(eqs, t)

sim_conds = [X1 => 3.4, k1 => 1.2, k2 => 0.13]
oprob = ODEProblem(sys, sim_conds, (0.0, 1.0))
sol = solve(oprob)

oprob[X1] # 3.4
oprob[X2] # 1.0
sol[X1][1] # 3.4
sol[X2][1] # 1.0

[TorkelE]

(should note that if a parameter value is forgotten, an error is thrown)

[AayushSabharwal]

Yeah this is a difficult one. If only MTKBase is loaded, all variables have a guess of 1.0. I did this since without proper mtkcompile, the initialization system has a lot of unknowns (almost every single one in the ODE) and requiring users to provide guesses for them is highly unreasonable. As a side effect, this means that cases like this are possible, since you'll just end up with an underdetermined initialization.

CC @ChrisRackauckas

[TorkelE]

At least for a pure ODE it should be fairly easy to detect this? Alterantively, give a really specific initial guess like 1.0214124151225503, and if that is unchanged, there might be something fishy? Having this as behaviour will inveitably lead to lots of silent and nasty errors when people forget to specify an initial condition for one of their 15 ODE variables.

[ChrisRackauckas]

Wait why doesn't this warn? i would think even in the MTK base one, you know if you make a NonlinearLeastSquaresProblem so it should throw the warning that its underdetermined?

[AayushSabharwal]

This absolutely should warn, I'd be very surprised if it doesn't. MTK goes through the same code path.

[AayushSabharwal]

At least for a pure ODE it should be fairly easy to detect this?

Yes. I can add that validation. Maybe even a general heuristic: if length(observed(initsys)) + length(equations(initsys)) == length(observed(ode)) + length(algebraic_equations(ode)) then something is very wrong.

[AayushSabharwal]

Yep, just verified that this throws a warning.

[ChrisRackauckas]

okay so then @TorkelE can you show yours is actually silent?

[ChrisRackauckas]

BTW I don't think we should do the default guess, just have an easy way to set such a guess, there's already an issue on that.

[AayushSabharwal]

BTW I don't think we should do the default guess, just have an easy way to set such a guess, there's already an issue on that.

I built that infrastructure for this. Removing the default guess is a one-line change. For users, adding the default guess is a single kwarg.

[TorkelE]

I can an initialisation-related warning, can confirm that, but otherwise everything proceeds like fine. But the problem is undefined, and there still should be an error rather trying to guess where the ODE should start.