Investigate using scipy_dae instead of scikits.odes

[moorepants]

https://github.com/JonasBreuling/scipy_dae

new as of 2024, looks like it might be pure python

[Peter230655]

I will play around with it a bit and report anything interesting

[moorepants]

Thanks, I will be interested in your findings.

[moorepants]

I'm also working on a conda package: conda-forge/staged-recipes#32184

[Peter230655]

I'm also working on a conda package: conda-forge/staged-recipes#32184

dumb question: What does this mean? Does it mean, it will be possible to install scipy-dae from anaconda?

[moorepants]

It will be possible to install scipy_dae with conda from conda forge.

[Peter230655]

VERY initial findings:
I used my n-link closed pendulum simulation to compare solve_ivp to solve_dae

• solve_dae takes twice as long to integrate
• accuracy is worse by a factor of 10 (I compare how well the energy is kept constant and how well the speed constraints are kept)

Nice:

• the signature is virtually the same as with solve_ivp
• the result object seems to be the same, e.g. result.nfev gives the number of times the r.h.s. was called.

[Peter230655]

I want to test the dae capability of scipy_dae.
My closed pendulum has two configuation constraints, which I convert to velocity constraints to use sympy mechanics Kane's method.
I do not know of a simple way to get Kane's method to give DAEs if there are configuration constraints.
Am I missing something obvious?

[moorepants]

I'm not concerned with speed for this book. I want an easy-to-install and clean API DAE solver for teaching students how to prepare the equations of motion for the solver and get a result. It should allow us to demonstrate lowering tolerances to ensure there is not constraint drift.

So far it is easy to install because it is pure python. We then need to see if it can replace scikit.odes in the holonomic constraint simulation chapter. Any other findings and learnings can be discussed in the PyDy repository or other relevant places.

[moorepants]

If you want to help for this specific issue about the book (and course), I suggest simply trying to replace scikits.odes with this package in the chapter on holonomic simulation.

[Peter230655]

I will certainly try.

[Peter230655]

Update:
I think, I managed a way to use the dae feature. I placed a question with that repository to find out if correct, no reply, yet.
I tried my way on a four link pendulum. solve_ivp is still an oder of magnitude more accurate than solve_dae.
So, unless I did it wrong (I will see when I get the reply) this may not work for your purpose.

[Peter230655]

Update:
using rtol = $10^{-3}$ and atol = $10^{-6}$ and using the standard integration method, solve_dae is more accurate by a factor of 10 to 100.

[Peter230655]

I will certainly try.

The way I did it with my n link pendulum was like this:

• I set up the eoms with the relevant velocity constraints, using sympy mechanics kane.
• I added the two configuration constraints to the bottom of the eoms
• I added two me.dynamicsymbols to the state vector

This seems to work, but I do not have confirmation that this is the right way.

Should I try to modify your lecture like ths?
(Of course I do not know whether i manage with all these .rst files.....)

[moorepants]

If you want to demonstrate that scipy_dae can be a replacement in my chapter, then I will have a look at that and consider changing the chapter.

You can also update the chapter with a PR here, but I will review it just like any suggested change to the materials. I also may not adopt it until next year's course. We have already started this year's course.

[Peter230655]

scikits.odes.dae(...) is complicated in my opinion vs. solve.dae(..) which follows solve_ivp very closely, and hence very easy to use,
So, yes it would definitely be nicer in your lecture.

However, I do not think switching to solve_dae is crucial for your lecture.
So, I will try to update the chapter with a PR.