more polyslab gradient bug fixes #2020
Conversation
|
I tested this a bit more (as mentioned) and I think it's good to go. |
bb7e202 to
1b30214
Compare
54959aa to
6b254a7
Compare
|
OK this is ready, freshly tested in notebooks and all seems good. The changes:
|
| @@ -1037,7 +1037,7 @@ def to_adjoint_point_sources(self, fwidth: float) -> List[PointDipole]: | |||
|
|
|||
| for freq0 in field_component.coords["f"]: | |||
| omega0 = 2 * np.pi * freq0 | |||
| scaling_factor = 1 / (MU_0 * omega0) | |||
| scaling_factor = 33 / (MU_0 * omega0) | |||
There was a problem hiding this comment.
Well that seems pretty arbitrary :D why 33?
There was a problem hiding this comment.
the honest answer is because I tested the PointDipole sources against every structure combination and the gradients were consistently 33x smaller than numerical ...
There was a problem hiding this comment.
A little worrying then. Have you tested different grid sizes at the monitor location? And the monitor being in different materials?
There was a problem hiding this comment.
It's hard to test everything, I already had like a 10 x 4 grid of different combinations of things to test, but no I did not test grid size and medium. The grid size should be handled by using interp in the source, the medium could have an effect but I think it would be some scaling by the refractive index, the overall constant might still be the same. I can look into it some more if it would be useful
There was a problem hiding this comment.
Yeah I'm just trying to think about things that this "constant" might actually depend on, like the mesh step around the dipole for some reason, or the material epsilon.
There was a problem hiding this comment.
it's very possible the gradients depend on these things in some way. This PR resolve the gradient direction and corrects the overall normalization so that the un-normalized gradients match pretty well (typically under 10% unnormalized and around 1% normalized), but under my test conditions. So There's another dimension to explore which is how these norms change under different background permittivity and grid resolution.. which gets pretty complicated. I think for sake of practical optimization what we have here is a good improvement, but I'll do a few more tests today to see if I can find some low hanging fruit to improve it
There was a problem hiding this comment.
I tested a few different Simulation.medium and also a few different grid sizes. The grad norms were constant across each, maybe a tiny bit of variation with dl, but didnt seem linear with dl and could be due to many things. Ultimately feel confident that these constants are good to give adj ~ numerical grads without normalization for a range of situations, and if not, we can revisit later. So I'll go ahead and merge this
Some minor but important polyslab gradient fixes. Surprised I missed this but glad I ended up catching it after all. Goes to show how tricky these things can be.
Found a sign error in the gradient contribution. This was not really affecting anything except in certain polarizations, it turned out to improve a customers' taper gradient accuracy, and fixed one of my numerical tests under certain conditions.
Also realized some condition for incorporating the slab_bounds into the area calculation was not being included as it was out of date with another PR. So after this fix + another tweak to the gradient norm, I am seeing POlySlab gradients matching numerical not only better than before but with the right magnitude.