How one should choose the time discretization where the problem will be evaluated?
For time-domain EM problems, you will want to use a fine discretization where the fields are changing quickly (early in time – when the source is shut-off, either with a step-off function or if a waveform is used), and then you can slowly expand the time-stepping as the simulation runs because the fields change more slowly.
One thing to note in the computation, is that when solving the time-domain, each distinct time-step you use requires that the system matrix be re-factored. So if I used a time-mesh of
time_steps = [(1e-6, 10), (3e-6, 10), (1e-5, 10), (3e-5, 10), (1e-4, 10)]
then we will need 5 matrix factorizations. So it is a bit of a trade-off between the total number of time-steps needed and the number of factorizations.
A good test is to run a simulation over a half-space and check that the time-decay curves are smooth. For example, in this one https://github.com/simpeg/em-notebooks/blob/master/notebooks/TDEM_vmd_sounding_over_sphere.ipynb, there is a slight kink in the half-space time decay at 2e-4ms, so we might want to go in and check if that is where the time-discretization changes and instead maybe take a couple more steps at a finer discretization.