pschlik, on 25 October 2023 - 04:36 PM, said:
Had another moment to test today. Numbers of substeps are more tame now, and in combination with more gradual change frame time pacing is definitely more consistent now. Also remained fairly stable even at 300% sim speed so the algorithm seems to be running fast enough at the moment. Still seeing some unusual indications from the simulated wheel adhesion though, it will still get stuck at unreasonably high values (for example, the MSTS Dash 9 reading 33% ahdesion while at a complete standstill). That can't be doing anything good for the stability of the model and there has to be a simpler way to avoid that than just throwing more substeps at the system. Surely it would be obvious without substepping that a force of 0 should require and ahdesion of 0?
Thanks for the feedback.
I am taking from this that you think that the Polach adhesion substep rate is "about right" now, is this correct?
Can you expand in more detail on how to duplicate the " it will still get stuck at unreasonably high values (for example, the MSTS Dash 9 reading 33% ahdesion while at a complete standstill)" scenario? The old model always had the maximum adhesion (given by locomotive adhesion) occurring at zero speed.
pschlik, on 25 October 2023 - 04:36 PM, said:
And this might be related but whenever I apply power ammeters are reading NaN! (Dynamic braking is fine though?)
Ok I can duplicate this, and I now roughly when this started to ok, but I can't see any obvious issues. I will investigate further (I am off on holidays for a week on Monday my time) so I am not sure whether I can fix it before then.
pschlik, on 25 October 2023 - 04:36 PM, said:
At this point, I'm not convinced that this model is a fool's errand but even after working through the weird stuff, it will clearly need more computing power than the previous model. Maybe some simplifying assumptions can be made, or the old model could be user-selected (though it could use some modifications to give more reasonable wheel creep), but I don't know the details and it's not the right time to worry about that anyway.
I would prefer not to have multiple user selected models, we should strive for the most efficient and realistic performance (this may still need to be determined, and it could be either the new model working at an acceptable level, or reverting to the old model)..