[tecmul]

Hi all,

I heard somewhere that in transient simulations, the time step should not change too quickly. Is this true? If yes, why?
For example, if a constant time step of 0.1 seconds produces results of acceptable accuracy, why would reducing the time step to 10^-6 seconds mid-run be a bad thing? Neglecting the huge increase in computation time of course.

[sbaffini]

Assuming we are talking about methods that can actually handle that in an accurate way, I would make a distinction between methods only requiring previous the time step solution and those requiring more time levels.

I don't think the former (e.g. runge kutta) have such issues.

The latter do, in the same way a sudden change in grid spacing is bad. But with time integration it is all over your grid.

[tecmul]

Originally Posted by sbaffini

The latter do, in the same way a sudden change in grid spacing is bad. But with time integration it is all over your grid.

It isn't obvious to me why multi time level schemes would suffer from a sudden change in the time step. An increase in the truncation error maybe? Could you explain a little more?

[arjun]

The fluxes have time step dependency and when you suddenly change the timestep the change in flux is huge, most often you do not have number of inner iterations to handle this.


If you wish to make sudden change then at least for few time steps you would have to have very large number of inner iterations.

[sbaffini]

Originally Posted by tecmul

It isn't obvious to me why multi time level schemes would suffer from a sudden change in the time step. An increase in the truncation error maybe? Could you explain a little more?

As I said, it is in the exact same way it happens for a grid, non-uniformity introduces further errors which were not present for constant steps. Problem is that the nature of the additional terms is typically different from that of the errors also present for the uniform steps and their magnitude an order smaller. Also, the error sign depends from the change of the step (increase vs decrease).

Not sure what to add, Ferziger explains the matter pretty clearly.

[FMDenaro]

Originally Posted by tecmul

Hi all,

I heard somewhere that in transient simulations, the time step should not change too quickly. Is this true? If yes, why?
For example, if a constant time step of 0.1 seconds produces results of acceptable accuracy, why would reducing the time step to 10^-6 seconds mid-run be a bad thing? Neglecting the huge increase in computation time of course.

Would you use a mesh having a cell of size 0.1 with an adjacent cell of size 10^-6?