Alternative approach for stop_when_dft_decayed

[mochen4]

Fixes #1831.

With the alternative approach, I couldn't find a default tol/decay_by parameter that would pass the single precision tests. The simplest solution I found is to use the minimum_run_time, as I currently have in this PR.

Alternatively, I could set a specific decay_by for each unit test in test_adjoint_solver.py

[codecov-commenter]

Codecov Report

Merging #1847 (4b8ac79) into master (2baeaef) will decrease coverage by 0.02%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1847      +/-   ##
==========================================
- Coverage   73.40%   73.38%   -0.03%     
==========================================
  Files          17       17              
  Lines        4896     4892       -4     
==========================================
- Hits         3594     3590       -4     
  Misses       1302     1302              

Impacted Files	Coverage Δ
python/adjoint/optimization_problem.py	56.77% <ø> (ø)
python/adjoint/wrapper.py	98.50% <ø> (ø)
python/simulation.py	76.81% <100.00%> (-0.04%)	⬇️

[stevengj]

Would be good to visualize the field in the single-precision case when the decay stagnates — i.e. what modes are causing it to have such a high Q? The fields at the Nyquist frequency, or maybe at zero frequency?

(If it is the zero frequency, then using an integrated source might fix it?)

[mochen4]

I noticed the decay stagnated for the eigenmode test in the single-precision case at the traditional forward simulation:

meep/python/tests/test_adjoint_solver.py

Line 401 in 5453a98

S12_unperturbed = forward_simulation(p, MonitorObject.EIGENMODE, frequencies)

but shouldn't it be basically identical to the forward simulation in OptimizationProblem (which decays fine)? @smartalecH

meep/python/tests/test_adjoint_solver.py

Line 398 in 5453a98

adjsol_obj, adjsol_grad = adjoint_solver(p, MonitorObject.EIGENMODE, frequencies)

[smartalecH]

As discussed, I think the issue is that forward_simulation isn't placing DFT monitors in the design region (where you are seeing some high-frequency speckling). The simulation itself should be identical between the two, but the adjoint_solver routine is explicitly waiting for the "speckle" to converge.

[mochen4]

Here are a few plots of the field when it stagnates

[mochen4]

The field's high-frequency pattern is probably due to the round-off at single-precision.