Updates for casadi 3.6

[valentinsulzer]

Release notes: https://github.com/casadi/casadi/releases/tag/3.6.0

Will add as found

• time grid points should be passed when creating interpolator instead of when solving
• rootfinder errors

[valentinsulzer]

I don't know how to fix the failing sensitivity test so skipping for now, @martinjrobins @jsbrittain could you take a look at this?

[martinjrobins]

sure, this is also reported in #2789 (comment).

[valentinsulzer]

I think it might be a different error here

[martinjrobins]

Did PR #2883 fix the problem? I just tried test_sensitivities on spm, spme and dfn with idaklu solver and it seemed to work ok.

[valentinsulzer]

If you can remove this and the tests still pass then I'm happy

PyBaMM/tests/integration/test_models/test_full_battery_models/test_lithium_ion/test_dfn.py

Lines 16 to 18 in 639b6b4

	def test_sensitivities(self):
	# skip sensitivities test for now as it is failing with casadi 3.6
	pass

[martinjrobins]

ahh ok, I wasn't aware you'd added that

[martinjrobins]

near as I can figure this is the problem. Say you have a model like this

$$\frac{du}{dt} = -u$$

$$ v = a * u $$

$$ u(0) = 1$$

$$ v(0) = 1$$

For $a=1$, sundials calculates the initial condition for $s = \frac{dv}{da}$ as $s(0) = [0, 0]$, presumably because v(0) does not depend on $a$, but if you correctly write the initial condition for $v$ as:

$$\frac{du}{dt} = - u$$

$$ v = a * u $$

$$ u(0) = 1$$

$$ v(0) = a$$

For $a=1$ this is the same equations, but then sundials finds the initial condition should be $s(0) = [0, 1]$. Not exactly sure what the expected behaviour is really, or why sundials isn't calculating the correct ic for $s$ even when you don't give the correct symbolic form for $v(0)$

Note the problem is only for the initial condition, as soon as you start integrating the solution becomes correct

[valentinsulzer]

$s(0) = [0, 1]$ is the correct initial condition in the second case, right?

[martinjrobins]

s(0)=[0,1] is the correct initial condition in the second case, right?

yup, and I believe it should be for the 1st as well, so that is why I'm a bit confused that the calcic function in sundials (ie. calculate consistent initial conditions) isn't finding it, it just returns [0, 0] for the 1st and [0, 1] for the 2nd.

[martinjrobins]

ahh ok. You have to call IDAGetSens after IDACalcIC in order for the solution to actually show up in your sensitivity vector..... fixed now