FAQ  •  Login

BDF1 not converging

<<

samuelacuna

Newbie
Newbie

Posts: 12

Joined: Wed Sep 14, 2016 11:41 pm

Unread post Wed Nov 02, 2016 4:11 pm

BDF1 not converging

For those on the simEngine3D track...

Has anyone been able to converge their BDF of order 1? This is the first integration at the first time step, which will be used to seed the order 2 BDF eventually.

My correction term (deltaZ) has a few values in it that just wont converge to zero. They just start to oscillate around ± 0.03.
My initial conditions seem correct (at least they match what I computed during inv dynamics analysis).

If you've figured this out, I'd like to compare system values (psi, g, phi, F, tauHat, etc.) to see where things are going wrong. Email me.

Samuel
saacuna@wisc.edu
<<

Dan Negrut

Global Moderator
Global Moderator

Posts: 833

Joined: Wed Sep 03, 2008 12:24 pm

Unread post Wed Nov 02, 2016 4:27 pm

Re: BDF1 not converging

Most likely the Jacobian is not healthy.
One quick thing you can do: use finite differences to compute your Jacobian numerically using your set of equations. Then compare to what you're getting in the Newton step to see whether they are close to each other.
Does this make sense?
Dan
<<

michaeltaylor

Newbie
Newbie

Posts: 27

Joined: Fri Sep 04, 2015 12:51 pm

Unread post Thu Nov 03, 2016 6:55 pm

Re: BDF1 not converging

Samuel,

Did you have any luck solving your problem? I've been dealing with the same issue since Tuesday and I haven't had any luck figuring it out ??? . My values for t = 0 match the Inverse dynamics solution and at time step 2 I get crazy qdd and lambda values, and then things return to normal at time step 3 if I use BDF order 1 and a couple time steps later for BDF order 2. If I force the Inverse dynamics solution for step2in the dynamics analysis, then this issue still occurs, but now at time step 3 :o . In all of my steps, even the crazy one, NR converges in 4-5 iterations. My Phi_q is identical between the inverse dynamics solution and the dynamics solution when it is going crazy :-\ .

If I disable the driving constraint, then problem 1 works fine the entire time with BDF order 1 and order 2.

I've pretty much run out of ideas of things to try or look at :'( . This issue is causing garbage results for problem 2 since it launches the pendulum during the second time step.

Please let me know if anyone is seeing the same thing as well and if you've had more luck solving it.

Mike

See the plot
Attachments
HW8_PendulumAccel_BDF1_OddStep2.png
SimEngine3D Odd Step #2 for BDF order 1
HW8_PendulumAccel_BDF1_OddStep2.png (78.11 KiB) Viewed 1226 times
<<

danielpiombino

Newbie
Newbie

Posts: 14

Joined: Wed Sep 14, 2016 11:45 pm

Unread post Fri Nov 04, 2016 1:28 pm

Re: BDF1 not converging

I'm not sure if this is the same thing or not, but my solution spikes drastically for the first couple points. The solver converges in 1 iteration for each of those points, but clearly not to the right point. My acceleration, velocity, and force plots all have similar spikes, but the height is about the same as the rest of the plot for those. I tried lowering my time step to 1e-3 but nothing changed.

-Dan
Attachments
Reaction Torque 1e-2.jpg
Reaction Torque 1e-2.jpg (23.6 KiB) Viewed 1208 times
<<

michaeltaylor

Newbie
Newbie

Posts: 27

Joined: Fri Sep 04, 2015 12:51 pm

Unread post Fri Nov 04, 2016 1:52 pm

Re: BDF1 not converging

It looks the same as mine currently does.

See attached BDF order2 with h = 0.05.

Mike
Attachments
HW8Prob1__ReactionTorque.png
BDF2 h =0.05 with initial spike
HW8Prob1__ReactionTorque.png (102.03 KiB) Viewed 1202 times
<<

f13-759-gsubramani

Newbie
Newbie

Posts: 27

Joined: Mon Sep 09, 2013 9:12 am

Unread post Sat Nov 05, 2016 9:37 am

Re: BDF1 not converging

While trying to debug my my iteration matrix I noticed that the euler parameter normalization constraints are different from the ones I have implemented. I used p'p - 1 = 0 as my euler parameterization constraint but in the slides(10/17 slide 17) it uses 1/2(p'p - 1) = 0. This would mean the iteration matrix is different from the one in the slides. I hope this helps someone still caught off guard.
<<

Dan Negrut

Global Moderator
Global Moderator

Posts: 833

Joined: Wed Sep 03, 2008 12:24 pm

Unread post Sat Nov 05, 2016 9:51 am

Re: BDF1 not converging

Samuel,
When you say that it's different, do you mean that there is a mistake in the slides?
If so, will you please point out which slide I need to modify.
Thank you, Samuel.
Dan
P.S. Does you simEngine3D converge now?
<<

f13-759-gsubramani

Newbie
Newbie

Posts: 27

Joined: Mon Sep 09, 2013 9:12 am

Unread post Sat Nov 05, 2016 8:01 pm

Re: BDF1 not converging

Hi Dan,

There is no mistake in the slides. The slides are consistent and correct. It is just that, on slide 9/24 slide 24 they are p'p -1 = 0 and in the recent one, that has the iteration matrix slides(10/17 slide 17) it is 1/2(p'p - 1) = 0. Since I implemented the older version, I was caught off guard when building my iteration matrix. The partial derivative is different as in the older version it is 2p and in the newer version it is just p. As far as the slides are concerned, they are correct in their context. I made the mistake by copying the matrix blindly. I should have looked at what I was doing a bit more closely.

I am getting converging results!

Sorry, I should have signed off....
Thanks!
Guru
<<

michaelvignos

Newbie
Newbie

Posts: 7

Joined: Fri Sep 04, 2015 12:51 pm

Unread post Sun Nov 06, 2016 8:46 pm

Re: BDF1 not converging

Guru,

Thanks for your input. Your suggestion of changing the definition of the Euler parameter normalization constraint from p'p - 1 to 0.5*(p'p - 1) helped with my convergence.

Thanks!

Mike
<<

samuelacuna

Newbie
Newbie

Posts: 12

Joined: Wed Sep 14, 2016 11:41 pm

Unread post Sun Nov 06, 2016 8:55 pm

Re: BDF1 not converging

Guru,
This fixed my code! Thanks.

Dan,
I spent a lot of time trying to figure out what was wrong... the euler parameter normalization constraint being scaled by 0.5 was a surprise.
Perhaps next time you teach this class, I would recommend emphasizing that this difference matters in simulation, because even though I saw this in the slides, I just assumed they were mathematically equivalent.
thanks.

- Samuel

Return to ME751 Fall 2016

Who is online

Users browsing this forum: No registered users and 1 guest

cron
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group.
Designed by ST Software.