edoardo 36c1db0df2 code alignment between the two versions. There are now two possibilities: 1)optimize the average objective function, this generally leads to different performance of the converted modes; 2) optimize using an epigraph formulation. This assures that whatever the final epigraph value, the performance of the converted modes will be greater (in terms of mode overlap) than the epigraph itself. The epigraph is implemented as a constrained optimization, which is stricter than optimizing the average

2026-07-28 22:41:42 +02:00

misc

added fiber_modes.ipynb useful utility to compute fiber modes

2026-06-04 14:29:31 +02:00

fiber_modes.ipynb

added fiber_modes.ipynb useful utility to compute fiber modes

2026-06-04 14:29:31 +02:00

hdf5_vis.py

another parameter test

2026-05-25 16:44:50 +02:00

metapeg_v2_unc.py

fixed plots between epochs

2026-05-31 10:51:01 +02:00

metapeg_v2.py

fixed plots between epochs

2026-05-31 10:51:01 +02:00

metapeg.ipynb

modified version of metapeg, using epigraph and linewidth/spacing constraints

2026-05-24 17:47:07 +02:00

metapeg.py

modified version of metapeg, using epigraph and linewidth/spacing constraints

2026-05-24 17:47:07 +02:00

metaprop_ad.py

modified filter parameters. Using a conic filter with radius 0.5

2026-06-05 08:32:06 +02:00

metaprop_asm.py

general testing of parameters. Regarding metaprop_fibermode_doublet, the spatial sampling has been fixed and the grid size is controlled by the number of pixels. I have seen that the mask-mask and mask-output plane distance is crucial and should not be chosen randomly, at the moment i have tested empirically some dimensions. The parameters in this commit seem to work ok, but we need to understand. Then, intensity conversation is still an issue, since the pattern at the output has 95/100, meaning 5/100 is lost somewhere. We really need to understand if this is an issue or not. With these settings, we might want to increase the window size, but then see the next point. I have introduced boundary boxes on phase masks, this is to avoid the algorithm using pixels that are clearly outside and only see noise. If we extend the simulation domain, we ought to extend these boundaries, otherwise the issue will present itself again. For what pixelation is concerned, honestly i expected worse, though there is still the need to apply some kind of filter to eradicate spurious single pixels. That should be doable using autograd, since the heavylifting is done by the adjoint, using autograd just for the filter part should not require too much computing effort. This needs to be tested.

2026-06-19 17:17:45 +02:00

metaprop_fibermode_doublet_epigraph.py

code alignment between the two versions. There are now two possibilities: 1)optimize the average objective function, this generally leads to different performance of the converted modes; 2) optimize using an epigraph formulation. This assures that whatever the final epigraph value, the performance of the converted modes will be greater (in terms of mode overlap) than the epigraph itself. The epigraph is implemented as a constrained optimization, which is stricter than optimizing the average

2026-07-28 22:41:42 +02:00

metaprop_fibermode_doublet_multi.py

code alignment between the two versions. There are now two possibilities: 1)optimize the average objective function, this generally leads to different performance of the converted modes; 2) optimize using an epigraph formulation. This assures that whatever the final epigraph value, the performance of the converted modes will be greater (in terms of mode overlap) than the epigraph itself. The epigraph is implemented as a constrained optimization, which is stricter than optimizing the average

2026-07-28 22:41:42 +02:00

metaprop_fibermode_doublet_shiftSource.py

added metaprop_fibermode_doublet_shiftSource that introduces some modifications needed to properly account for out-of-axis excitations. Initially, it was observed that previous code was only capable of optimizing an on-axis source to an on-axis target. If for example the source was shifted away from the center, then the optimization would fail. Even if the target was shifted out-of-axis to match the position of the source, the algorithm would not work. Several bugs produced this issue. The first was related to how the theoretical model was translated to code, in particular how we handled gradient calculations. In theory, we considered column vectors, for which taking the transpose simply turns them into row vectors but does not change the element wise product. In the code, we instead deal with 2D arrays, and at that point transposing is no longer irrelevant. Check notes for more details about this bug. The second bug was related to how we handle FFTs. In particular, the simulation space is centered in the center of the array. FFTs work considering that the signal starts at the corners. This center mismatch did not cause any issue for the on-axis case, due to the symmetry of the problem. We did not observe both of these two bugs because in the on-axis case everything is centered.

2026-07-14 17:13:02 +02:00

metaprop_fibermode_doublet.py

general testing of parameters. Regarding metaprop_fibermode_doublet, the spatial sampling has been fixed and the grid size is controlled by the number of pixels. I have seen that the mask-mask and mask-output plane distance is crucial and should not be chosen randomly, at the moment i have tested empirically some dimensions. The parameters in this commit seem to work ok, but we need to understand. Then, intensity conversation is still an issue, since the pattern at the output has 95/100, meaning 5/100 is lost somewhere. We really need to understand if this is an issue or not. With these settings, we might want to increase the window size, but then see the next point. I have introduced boundary boxes on phase masks, this is to avoid the algorithm using pixels that are clearly outside and only see noise. If we extend the simulation domain, we ought to extend these boundaries, otherwise the issue will present itself again. For what pixelation is concerned, honestly i expected worse, though there is still the need to apply some kind of filter to eradicate spurious single pixels. That should be doable using autograd, since the heavylifting is done by the adjoint, using autograd just for the filter part should not require too much computing effort. This needs to be tested.

2026-06-19 17:17:45 +02:00

metaprop.py

added metaprop, implementing adjoint inverse design through the angular spectrum method

2026-06-02 23:17:21 +02:00

mode_converter.ipynb

first debug on metapeg. At the moment there is only one frequency being optimized. Although we are not using the epigraph yet, it would still be convenient to still consider a small range around the central frequency and optimize the average objective between them. This would still be better than considering only one frequency, but for the moment I need to understand if this model can work or not. Before running a full batch, more debugging is required to better understand how everything works. In this release, the code has not been tested in the results part, but the main optimization function run a couple of rounds without throwing errors. We still need to assure the normalization we make on the far fields is reasonable or not. The next step is to run a dummy simulation until the end and observe the results generation works, then the true test would be to let it run a true optimization with the right resolution.

2026-05-13 22:43:54 +02:00

mode_converter.py

fixed observed far-field component

2026-05-13 10:43:51 +02:00

sim_opt_design.py

1) minor changes to metapeg_v2, but we no longer modify the chunk property and leave it to the default value (=True). Setting it to False causes segmentation errors in the adjoint loop, but works for direct simulations (for example in the normalization run) 2) added sim_opt_design, which is used to simulate an optimized design given the values of an optimized MaterialGrid after and outside the main optimization algorithm

2026-05-27 14:55:09 +02:00

S

Description

An inversly designed mode converter based on a custom ASM propagator. FDTD simulations (meep) are used to design the unit-cell and verify the full design.

45 MiB

Languages

Jupyter Notebook 90.8%

Python 9.2%