✨ There is plenty to see at the bottom: discovering phases and ferroic variants in (real) atomically resolved images
A decade ago, Steve Pennycook and I have posed that electron microscopy can become transformative tool for exploring physics of condensed matter, as summarized in our opinion in Nature https://lnkd.in/g9jZmkbv.
At that time, there was extensive discussion on whether pushing the resolution of STEM is justified - after all, there are no smaller details then atomic spacing. We argued that physics of solids - polarization, octahedra tilts, chemical expansion - is hidden in these structural details. By now, this viewpoint is well established and STEM did in fact approach limit of resolution imposed by thermal vibrations.
That said, just getting atomically resolved images is not enough - we need to extract useful information from them. We wrote "Large, multi-dimensional data sets will pose challenges for data collection, storage and analysis. New approaches will be needed for extracting relevant knowledge and linking it to theory"
Indeed, the classical approach to do so requires physical model, e.g. relationship between atomic column coordinates and physical order parameter fields. This approach has two limitations:
- we substitute the richness of image data by the point estimates of atomic positions
- we ignore the unknown effects and phenomena, i.e. physical model becomes a potentially biased regularizer of the data. It is necessary when data is noisy, but by now it is really oversampled
Can we do better?
In this tutorial notebook, Kamyar Barakati illustrates how we can use the family of Variational Autoencoders with rotational and translational invariances to analyze this data, and particularly the use of conditional (on the atom type) VAE for mapping phases, order parameter fields, and detection of mis-tilt effects.
In this particular case, we can iterate between the physics based analytics and VAE analysis. However, the VAE approaches work for any data - so you can use this approach to dive into atomic world of your own!
Now, what would be the next step?
Once you go through the notebook, you will notice that analysis relies on building proper descriptors (atom-centered image patches) and of course there are hyperparameters of VAE. These (in this case) can be tuned based on expected physics, but do affect what we learn - e.g. see https://lnkd.in/eqRD8NDG.
So the next step is building autonomous human in the loop workflows: https://lnkd.in/ef_iYt-Q.
Stand by for further news!
https://lnkd.in/eyHsztn4
