Add periodic repetition in EME simulations

[caseyflex]

Major changes:

• Adds a field num_reps to every EMEGridSpec class, which can be used to periodically repeat that subgrid while reusing mode solves, overlaps, and interface s-matrices
• Adds a class EMEPeriodicitySweep which allows sweeping the num_reps of every EMEGridSpec in the EMESimulation.eme_grid_spec. This is currently done by assigning names to EMEGridSpec objects, and then passing num_reps as a list of dictionaries, one for each sweep_index, and each of which is a mapping from the name of the subgrid which will be repeated to the desired number of repetitions of that subgrid.

Minor changes:

• When multiple freqs are provided, uses the largest for automeshing instead of raising an error.
• Adds methods EMEExplicitGrid.from_structures and EMECompositeGrid.from_structure_groups which place grid boundaries at structure bounds
• Increases MAX_NUM_FREQS to 100

Todo:

• Check / extend test coverage

[tomflexcompute]

That's great @caseyflex . At the conference this week so hard to find a good chuck of time to carefully review it. Maybe @momchil-flex can take a look?

[yaugenst-flex]

Thanks @caseyflex, the code looks good to me, no comments there, only some high level ones regarding documentation.

This might just be my unfamiliarity with EME but from reading through the PR it wasn't really obvious to me when and how you would use this functionality, and for example what is the difference between a EMEModeSpec(num_reps=N) and an EMEPeriodicitySweep? A more detailed docstring at least for the latter would help greatly I think. Also perhaps a quick motivation in the docstring ("Periodic repetition [...] for X purpose") would go a long way. Is there a related example notebook?

[caseyflex]

Thanks @caseyflex, the code looks good to me, no comments there, only some high level ones regarding documentation.

This might just be my unfamiliarity with EME but from reading through the PR it wasn't really obvious to me when and how you would use this functionality, and for example what is the difference between a EMEModeSpec(num_reps=N) and an EMEPeriodicitySweep? A more detailed docstring at least for the latter would help greatly I think. Also perhaps a quick motivation in the docstring ("Periodic repetition [...] for X purpose") would go a long way. Is there a related example notebook?

Ah thanks @yaugenst-flex , I forgot the docstring. I'll add some details there. I think @tomflexcompute will make an example notebook once he has a bit of time.

Currently the api is like this:

unit_cell_grid_spec = td.EMEExplicitGrid(
    mode_specs=[mode_spec]*2,
    # boundaries=[0, p*ff, p],
    boundaries=[p*ff],
    name="unit cell")

port_cell_grid_spec = td.EMEUniformGrid(
    mode_spec=mode_spec,
    num_cells=1
)

eme_grid_spec = td.EMECompositeGrid(
    subgrids=[port_cell_grid_spec, unit_cell_grid_spec, port_cell_grid_spec],
    subgrid_boundaries=[0, p]
)

this defines a composite grid with two port subgrids (each having one cell) and a unit cell subgrid consisting of two cells which is meant to be periodically repeated, see the image below. The unit cell grid is named

Then there are two main ways you could use this. You could just set unit_cell_grid_spec.num_reps directly -- this would effectively repeat the unit cell that many times, and simulate a single structure. Alternatively, you can use EMEPeriodicitySweep like this:

n_list = [1, 50, 100]
sweep_spec = td.EMEPeriodicitySweep(num_reps=[{"unit cell": n} for n in n_list])

then this would simulate three different structures at once, with varying number of repetitions of the unit cell.

Let me know your thoughts on the API!

[yaugenst-flex]

Thanks for that explanation @caseyflex, that's really helpful, and the interface makes sense to me.

[tomflexcompute]

Thanks @caseyflex for the great addition of the EME capability!

[momchil-flex]

Thanks @caseyflex , just some small comments/questions

[momchil-flex]

Since this is user-facing it should have a proper doc string with parameters and return, and maybe an example.

[momchil-flex]

Should this and updated_copy_num_reps be private? Or is there a reason we might want them user-exposed?

[caseyflex]

Yeah, fair

[momchil-flex]

Same comment about full docstring.

[momchil-flex]

Do we want to lift this restriction eventually?

[caseyflex]

No, because the thing is the field may vary cell to cell on the repeated grid, so there is no way of capturing it on the original EME grid (before repetition)

[momchil-flex]

I'm not sure how I feel about using the name attribute for functional purposes. But this is not the first place in the code this happens. Just curious - is it used anywhere else in EME, and did you consider any alternatives?

[caseyflex]

It’s only here. I considered forcing the sweep num_reps to be a list of tuples, each of which contained a num_reps for the current grid spec, as well as a list of num_reps for the sub grids if it is composite (as a recursive data structure in the same format as the grid spec). But this was way too cumbersome

[tomflexcompute]

Btw the current limit on the number of frequencies is still a little limited I find. Just wondering if the limit is due to cost consideration? The credit cost scales linearly with number of frequencies right?

[caseyflex]

Btw the current limit on the number of frequencies is still a little limited I find. Just wondering if the limit is due to cost consideration? The credit cost scales linearly with number of frequencies right?

@momchil-flex it was actually just based on simulations taking too long, I remember you weighed in on setting it to 20, shall we relax it to maybe 500?