Gandiva is a tool for use in Molecular Beam Epitaxy systems (MBE) outfitted with RHEED (reflective high energy electron diffraction), a method of in-situ crystal monitoring. It can parse video footage- either pre-recorded or in real-time- and generate waveforms based on visual input from RHEED. What this means is, with only a camera, you can digitize your RHEED data, and:
- Get waveforms generated with respect to time, for real-time analysis of crystal layer growth.
- Observe how many layers are made, the thickness of the film, and the growth rate per hour, simply by inputting the lattice constant of your material, and allowing it to run unopposed.
- Export and import data to and from the app, in .json/.csv, and as images.
It's currently in use in DRDO's SSPL, an MBE lab with III-V and II-VI fabrication facilities, as their RHEED software of choice.
It has a sleek interface with QT's Python bindings, uses OpenCV for >100FPS analysis speeds on most devices, and has an integration with Matplotlib to provide graphs that don't look out of place from a publication. It also uses the GNU All-Permissive License so you can genuinely modify it in any manner possible, beyond what's allowed in most open-source licenses. That license is just 'I'm putting a license here so you know I forfeit my right to a license' and it basically means you can modify it, use it commercially or noncommercially and there's no obligation that any derivatives are open source as well.
If you upload the following video to Gandiva:
The user interface displays this:
Sinusoidal behavior is very clearly visible from this and we can see it accurately tracks the three peaks. It can also provide real-time analysis with the 'Start Live' tool. Make sure to select the correct camera device for it, since there may be multiple.
Download Gandiva.exe from the releases page, run it, and don't delete the Gandiva shortcut on your Desktop.
Most RHEED-parsing algorithms are either slow, operate predominantly on images and not video, are too slow for real-time rendering, or use complex computer vision algorithms which require careful tuning. The algorithm implemented here is a bespoke solution that's robust to varying initial conditions, uses no neural networks (entirely heuristic-based), and is extremely performant (>1000% faster than needed for real-time).
I call the algorithm INterDecile Rank Averaging, because it takes the average intensity of all the pixels between the 10th and 90th percentile of brightness, averages their intensity, calculates the average intensity of the top 100 pixels, and calculates the intensity differential between them as the metric for how bright the RHEED diffraction spot is when compared to the baseline intensity of the null spots. We use noise-filtered zero crossing detection for waveform detection. We used to use peak counting but we found it had to be tuned frequently, whereas this model is robust to noise even of the same order of magnitude as the main wave's amplitude.
Thus our solution is a lot more easily modifiable, significantly faster, comparably performant, and qualitatively different from finding a region of interest, or other conventional methods for RHEED analysis.