Energy is not conserved when all damping is turned off (unless I missed something)

[kkkmail]

Take PlanetDemo and turn off all damping.
Turn off the speed (set it to 0) so that all the small spheres simply fall to the planet.
Run the demo and observe that it eventually get into a "deep freeze state": most of the spheres get glued to the surface of the planet.

Correct behavior with no damping: small spheres should form an atmosphere-like structure. In short - there should be higher sphere density closer to the surface of the planet and all of them should keep moving.

Perhaps I did not turn off all damping. I'd appreciate to know how to do that for that example.

[kkkmail]

The issue is that the integrator does not conserve energy. I think that I even saw something like that mentioned somewhere in the code or the documentation: if you specify large frequency for SpringSettings even with zero dampingRatio, then you get the effective energy loss.

This is not physics.

The only easy way to circumvent that that I found is to set dampingRatio to some negative value after, of course, disabling Validate in SpringSettings.

Then, with some luck and careful calibration of negative dampingRatio (around -0.2625f for the models that I ran) and some relatively small frequency (5.0f) it seems possible to come up close to conservation of energy in the systems without friction.

The problems with that are that negative dampingRatio is heavily model dependent and that if the system has different types of dynamic bodies, then it is completely unclear which ones would get the "extra" energy and which ones would lose.

[RossNordby]

Sorry, traveling a lot right now, but short version: yeah, there is going to be some inherent damping in constraints. Individual constraints solved in isolation will produce results like an implicit integrator does, so when the behavior is too high frequency relative to the timestep or substep rate, the unrepresentable motion will get damped out.

Using higher solver step frequencies (either via more timesteps or higher substep counts) will tend to help, but there will still be some energy loss. I would not recommend negative damping ratios.

The standard interventions here would be:

1. 0 damping ratio,
2. Use a maximum recovery velocity of float.MaxValue since the demos default of 2 will damp out a whole lotta stuff,
3. Make sure the solver's effective frequency is significantly higher than the constraint frequency.
4. 0 friction.

This is not physics.

alas, the tradeoffs of real time video game simulation

[kkkmail]

Thanks for the comments. The funny thing is that I bumped into your wonderful library when looking for a simple way to showcase some thermodynamics, e.g. put a gas into a box and "observe" how its "temperature" changes when you change the volume of the box. Obviously, you need energy conservation for this to work. Anyway, some combination of low frequency, some negative damping ratio, float.MaxValue maximum recovery velocity, along with a careful choices of velocity integration count and substep count seems to produce a stable enough system, which can be used for such a purpose. I'd be glad to post the parameters here once we are done with these experiments.