Nigel Browning1 2 3 B. Layla Mehdi1 2 Houari Amari1 Heath Bagshaw1 Matthew Bilton1 Andrew Stevens3 Christopher Buurma3

1, Mechanical, Materials and Aerospace Engineering, University of Liverpool, Liverpool, , United Kingdom
2, Fundamental and Computational Science, Pacific Northwest National Laboratory, Richland, Washington, United States
3, Sivananthan Laboratories, Bolingbrook, Illinois, United States

Transmission electron microscopy (TEM) is a widely used method to observe and quantify the atomic scale structure, composition, chemistry, bonding, electron/phonon distribution and optical properties of nanostructures, interfaces and defects in many materials systems. In addition to observing static structures, in-situ gas and liquid stages now permit dynamic experiments to be performed inside the TEM to observe complex structural and chemical transformations. In all of these cases, the goal in designing and implementing the TEM experiments is to acquire the most information about the sample while the least amount of damaging electron dose is delivered to it –a modern TEM experiment must therefore first determine the “optimal sampling” conditions for the material being studied. In addition, as many TEMs now have direct electron detectors capable of recording 100-1000 images per second or more (resulting in terabytes of data for each experiment), data compression and the use of automated analytics play a key role in interpreting the results of these experiments. The use of compressed sensing and inpainting methods is now being taught as part of the regular senior undergraduate/graduate student course structure in advanced TEM methods. In addition, the use of machine learning to improve the analysis of the sub-sampled datasets and data analytics to extract key parameters from a series of images are also key parts of the course. The sub-sampling/inpainting methodology for optimal sensing is hardwired into the acquisition mechanism of the microscopes used for the practical aspects of these microscopy courses, allowing students to directly modify the means by which images are acquired and test its effect on the speed, resolution and precision of the images obtained/reconstructed/analyzed during their training. In this presentation we will discuss the use of TEM images, and in particular obtaining the best TEM image for the lowest dose, to teach the concepts of compressed sensing, inpainting and machine learning as part of a core materials science method. Overall, we have found that the direct atomic scale images of the structure permit students to quickly get a grasp of the main mathematical and data concepts and how to best implement them in their experimental design. Here we will also discuss the application of these same methods onto other materials characterization techniques (and imaging tools in general), the way that users are trained on those methods, and the precision of the results that are obtained.