Ptychography gridding pattern

[erh3cq]

Problem: Ptychography reconstruction on Dectris (and other hybrid detectors with segments) is creating a 'gridding pattern' upon reconstruction. The grid can be seen in real and reciprocal space, but is most obvious in reciprocal space.

Trouble shooting:

1. The periodicity matches the probe step size. I have run acquisitions with various combinations of max outer angle, binning, and probe step size.
2. @smribet rean Bayesian Optimization of position, aberrations, and mixed state. This helped to some minor degree but the issue is still prominent.
3. I ran the "ptycho01_gold_on_carbon.ipynb" file then ran it again with simulated boarder pixels by averaging (e.g. each boarder pixel is a 2x1 pixel that is then assigned to two 1x1 pixels with each having the same intensity). The artifact did not appear. This would suggest it is not the averaging of the boarder pixels that creates the issue.

Expected behavior: artifact free reconstruction

Things to try @cophus suggested a series of filter that @smribet or @gvarnavi could implement

py4DSTEM version 0.14.9 (from github phase) Python version 3.12.1 Operating system Windows

[tcpekin]

I saw extremely similar grid artifacts using a Dectris when either the pixel or step size was incorrect. The mitigation by updating probes and probe position kind of also hints at that... Idk could just be a me problem, but I am haunted by grid patterns that look like that 😂

[gvarnavi]

Thanks for the report @erh3cq! I've been sitting on a tutorial notebook which investigates gridding-artifacts for the past couple of weeks, and haven't found time to finalize/post it. Attaching it here in the meantime, you can find the necessary data files here.

iterative-ptychography_02.zip

Unfortunately, gridding is a pathological artifact in "blind" ptychography (i.e. where one tries to solve for the probe and the object simultaneously). This is well-known in the X-Ray ptychography literature, and in-fact why they have switched away from raster-scans and instead opt for spiral scans.

That being said, the notebook investigates regimes where this is more/less pronounced - and possible remedies. I summarize them here briefly with this figure, but do have a look at the notebook:

• Lower batch-sizes hide a lot of sins, and allow the probe (which is the first to absorb gridding, and then transfer it to the object) to update more frequently
  • "Dummy" mixed-state hides even more. You can use mixed-state with 2 probes and a large batch_size to essentially dump a lot of the probe artifacts into a sacrificial, low-intensity probe mode. This includes gridding, but it also absorbs the mean-diffraction pattern for free
• Gridding can be overcome with sufficient dataset redundancy, i.e. finer probe sampling
  • Note you don't want to achieve this using larger defocus, since you need a very comfortable cropping box to avoid self-interaction (the notebook suggests even 0.1% self-interaction results in aggressive gridding!)
• Vacuum probe measurements are great for initializing the probe, but if it's even slightly off from the measured intensities' BF-disk you probably don't want to be forcing it, as even slight model-mismatch results in gridding (note you might not have a choice for low-dose datasets)

Finally, you mention implementing a filter for this. Indeed - the gridding artifacts are predictable, and we can easily add a Fourier filter which zeros-out those pixels. You get what you ask for with Fourier-filtering, so indeed this can reduce/remove gridding but our tests have shown the reconstructions to be fairly sub-optimal in other ways.

[gvarnavi]

To clarify, and connect with @tcpekin comment above:

I saw extremely similar grid artifacts using a Dectris when either the pixel or step size was incorrect

Gridding is a pathological artifact due to the translation periodicity ambiguity introduced by raster scan patterns, and can be exacerbated by ANY kind of model-mismatch. The notebook above investigates the probe aperture and cropping box mismatch, but indeed calibration mismatch (including pixel-sizes and rotation) also result in gridding..

[gvarnavi]

Closing this issue, now that the tutorials have been updated to include gridding-artifacts.