max_tick_distance affecting weighting potential calculation

[JunwenDiao]

Hi!
By having different values of max_tick_distance, the weighting potential calculation for undepleted region behaves unexpectively:
to use the ICPC example file as an example:

sim = Simulation(SSD_examples[:InvertedCoaxInCryostat])
sim.detector = SolidStateDetector(sim.detector, contact_id = 2, contact_potential = 500u"V")

by using

calculate_electric_potential!(sim, device_array_type = CuArray, refinement_limits = [0.2, 0.1, 0.05, 0.01], verbose = false, depletion_handling = true,grid = Grid(sim, max_tick_distance = 0.5u"mm"))

and

for contact in sim.detector.contacts
    calculate_weighting_potential!(sim, contact.id, device_array_type = CuArray, refinement_limits = [0.2, 0.1, 0.05, 0.01], verbose = false, depletion_handling = true, grid = Grid(sim,max_tick_distance = 0.5u"mm"))
end

the weighting potential of undepleted region has value of 1 as expected:



but if I change the value of max_tick_distance to 0.1mm, the weighting potential for contact 1 is no longer having the whole undepleted region with value of 1:

while the weighting potential for contact 2 looks normal:

It seems like there is a threshold for what that is going to happen, and the weighting potential changes a lot with different max_tick_distance. A comparison with max_tick_distance being 2mm and 2.1mm:

Did I use the arguments incorrectly or can this be a potential bug?
Thank you!

[fhagemann]

Are you changing max_tick_distance just for the weighting potential calculation, just for the electric potential calculation or both?

And I assume everything is running with depletion_handling being true?

[JunwenDiao]

setting the max_tick_distance for both and depletion_handling = true for both weighting potential and electric potential

[fhagemann]

So, I can reproduce this using Float32 and Float64 precision.
I would need to look deeper into this to understand where this is coming from and how it can be avoided.
I opened an issue from this --> #530

[fhagemann]

A quick check revealed the following:
in the initial state of the weighting potential, all grid points except for the contact grid points are set to 0.
Then, we calculate the weighting potential using the SOR algorithm "into the bulk".
For the undepleted region, this apparently needs WAY more iteration steps if the max_tick_distance is set very small and therefore there are many grid points in the undepleted region.

When using calculate_electric_potential! or calculate_weighting_potential!, the maximum number of iterations is 50_000, and every 500 steps it checks whether it converges to 1e-7 (there was a typo in the docstring, that should be fixed by #531).

For max_tick_distance = 0.1u"mm":

apply_initial_state!(sim, WeightingPotential, 1, Grid(sim; max_tick_distance)))
plot(sim.weighting_potentials[1])

SolidStateDetectors.update_till_convergence!( sim, WeightingPotential, 1, 1e-7, depletion_handling = true,  max_distance = 0.1u"mm")
plot(sim.weighting_potentials[1])

Now, when increasing this (max_n_iterations = -1, meaning no limit and n_iterations_between_checks = 500_000) will force the simulation algorithm to run at least 500_000 SOR iteration steps before checking for convergence and possibly finishing the potential calculation. When I do this, the simulation takes significantly longer, but it looks like I'm getting the correct result still for max_tick_distance = 0.1u"mm":

calculate_weighting_potential!(sim, 1, 
    max_n_iterations = -1, n_iterations_between_checks = 500_000, 
    refinement_limits = [0.2, 0.01], depletion_handling = true, max_distance = 0.1u"mm")

To me, this means that the implementation of "convergence" would need to be revisited, as it seems like the algorithm is finding convergence too early, even if it did not yet relax to its final solution.

[JunwenDiao]

Thank you!