Opened 6 years ago
Last modified 3 years ago
#5328 new defect
ODE solver behaves erratically in two MultiBody examples with dynamic state selection
| Reported by: | Francesco Casella | Owned by: | Karim Adbdelhak |
|---|---|---|---|
| Priority: | blocker | Milestone: | 1.19.0 |
| Component: | Run-time | Version: | v1.14.0-dev-nightly |
| Keywords: | Cc: | Andreas Heuermann, Lennart Ochel |
Description
Steps to reproduce:
- set
-d=newInst,stateselection - simulate
Modelica.Mechanics.MultiBody.Examples.Elementary.SpringDamperSystem
The simulation terminates successfully, but the solver statistics and the fairly long simulation time point to some issue with the solver
solver: dassl 53347 steps taken 139199 calls of functionODE 76637 evaluations of jacobian 394 error test failures 25259 convergence test failures 4.2913s time of jacobian evaluation
Most importantly, if you cross-check body1.w_a[3] and its derivative, they are definitely not consistent with each other until about time = 2.4; afterwards, they seem more in synch with each other, at first sight.
This is strange, because the state selection includes body1.w_a[3] as a state, so the relationship among the two should be derivative/integral, with a relative tolerance of less than 1e-6. If you reduce the tolerance to 1e-7, the shape of the state curve changes, but it remains inconsistent with the derivative during the first two and a half seconds.
Maybe the indication of the state selection is incorrect, and what I am seeing is a wrongly computed dummy derivative. The reported state selection is the expected one, maybe the solver is doing something else in the background.
This problem does not show up if the old frontend is used to flatten the model, and I am currently investigating what is the difference among the two. Nevertheless, this kind of things should never happen anyway.
Change History (20)
comment:1 by , 6 years ago
comment:3 by , 6 years ago
| Cc: | added |
|---|---|
| Priority: | critical → blocker |
Replying to sjoelund.se:
#5332 was resolved: how are the results now?
Unfortunatly, they are exactly the same, so the problem is not the same as the one in #5332.
Apparently, it affects Modelica.Mechanics.MultiBody.Examples.Elementary.HeatLosses and Modelica.Mechanics.MultiBody.Examples.Elementary.SpringDamperSystem in a very similar way.
See also #5390, there is an issue on the first of the two models, maybe they are related.
Lennart, Karim, Andreas, would you mind having a look? This is also necessary to get MSL 3.2.3 100% working.
comment:4 by , 6 years ago
| Summary: | ODE solver behaves erratically in one MuliBody example → ODE solver behaves erratically in two MuliBody examples (issue with state selection?) |
|---|
comment:5 by , 6 years ago
| Summary: | ODE solver behaves erratically in two MuliBody examples (issue with state selection?) → ODE solver behaves erratically in two MultiBody examples (issue with state selection?) |
|---|
comment:6 by , 5 years ago
The dynamic behaviour is identical for the old and new frontend. The problem might be related to dynamic state selection. This is currently under further investigation.
comment:8 by , 5 years ago
The situation has gotten slightly worse after the last commits. With the latest nightly, the statistics now read:
solver: dassl 76030 steps taken 195615 calls of functionODE 110064 evaluations of jacobian 621 error test failures 36280 convergence test failures The simulation finished successfully.
comment:10 by , 5 years ago
I also checked the state selection against Dymola. The state selection is exactly the same: same statically selected states, same dynamic state selection set, same selection of actual dynamic states during simulation: body1.Q[1], body1.Q[2], body1.Q[3].
comment:11 by , 5 years ago
Maybe something's wrong with the Jacobian (numerical or symbolical, I don't know), since there is a convergence failure every two steps, and there are more evaluations of jacobians than steps, which is very unusual.
@Karim, can you please check that?
comment:12 by , 5 years ago
| Milestone: | 1.14.0 → 1.15.0 |
|---|
Releasing 1.14.0 which is stable and has many improvements w.r.t. 1.13.2.
This issue, previously marked as blocker for 1.14.0, is rescheduled to 1.15.0
comment:13 by , 5 years ago
| Summary: | ODE solver behaves erratically in two MultiBody examples (issue with state selection?) → ODE solver behaves erratically in two MultiBody examples (issue with symbolic Jacobian?) |
|---|
comment:14 by , 4 years ago
| Milestone: | 1.15.0 → 1.16.0 |
|---|
Release 1.15.0 was scrapped, because replaceable support eventually turned out to be more easily implemented in 1.16.0. Hence, all 1.15.0 tickets are rescheduled to 1.16.0
comment:15 by , 4 years ago
| Summary: | ODE solver behaves erratically in two MultiBody examples (issue with symbolic Jacobian?) → ODE solver behaves erratically in two MultiBody examples with dynamic state selection |
|---|
comment:16 by , 4 years ago
@Karim, maybe you want to have a second look at this.
In both cases, we have a rigid body suspended by a spring, which requires to use the internal state variables of the rigid body, in particular the quaternions for the orientation, with a dynamic state selection.
The initial conditions are such that the body mostly oscillates vertically, but also has a bit of initial angular momentum that makes it oscillate slightly around its horizontal axis and eventually to slightly swivel like a pendulum. This motion is caught in a completely wrong way by the dummy derivatives, resulting in an exponential amplification of the motion which is completely unphysical.
I'm not sure what you should check exacly to spot where the error is. Maybe comparing the solution with Dymola's will give some hints.
When #5332 is resolved, this issue will probably be resolved automatically. Yet, it would be worth understanding how it is possible that a non-dummy state derivative is so blatantly different from the slope of the corresponding state trajectory, because this issue may pop up in other models.