The advances in microscopy techniques have been key to great discoveries in fields like biological samples or material sciences, where researchers have been able to see up to atomical level. Now, combining the latest advances in microscopy with machine learning methods, an interdisciplinary team of mathematicians and physicists may go one step further in image resolution.
How can AI models help to overcome current resolution limitations in electron microscopy? Read in this week’s Helmholtz AI project showcase how researchers at the Forschungszentrum Jülich (FZJ) and Helmholtz Munich are working jointly on this challenge.
Could you introduce yourself, giving your affiliation, area of work, and of course, the project title?
I’m Benedikt Diederichs, former worker at the Ernst-Ruska-Center, Research Center Jülich, currently working at the Helmholtz Center Munich, the second partner of our project. My area of work is at the junction of machine learning and inverse problems, and that is exactly where the project EDARTI (Electron Diffraction Inversion by Artificial Intelligence Approaches) is situated.
In simple words, what specifically is your project about? And, how and why do you think it is a high risk, high gain endeavour?
My project is about fusing scanning transmission electron microscopy (STEM) with machine learning.
Roughly, STEM works as follows: you shoot electrons through a thin specimen or sample at a lot of different places. As electrons go through the sample, they get scattered and form peculiar patterns in the microscope detector depending on where in the sample they pass through. In a typical scan, thousands of such patterns are recorded - one for each place where you shoot the electron beam. And by studying these patterns you can find out many interesting properties of the specimen, for example the position and type of every atom.
Currently, disentangling the useful information from noise and uncertainty of the measurement process is a very challenging problem. However, simulating such diffraction patterns is well understood: we know how electrons behave when they go through different points in the samples. That makes it feasible to train a machine learning algorithm to improve the reconstruction process.
Using machine learning in STEM is a risky endeavor, as it is a completely new approach. We must first prove that we are really able to simulate the experimental conditions accurately, so that transference to experimental data is possible. As STEM is an important technique in materials science and increasingly in structural biology as well, pushing its limits would result in advances in many other areas.
How important has the Helmholtz AI funding and platform been to carry out this project?
Without Helmholtz AI funding my project would not have happened. As tackling challenging ML problems requires often intricate knowledge of the process we want to understand, bringing together practitioners with more theoretical working scientists has become even more important. Hopefully, Helmholtz AI is here to stay to make more projects like this possible.
Any other comments you wish to add?
Thousands of such patterns form the input for a single reconstruction