Data Science within Electron Microscopy

Richard Jinschek - A Case Study of Compressive Sensing applied in Electron Microscopy

Electron Microscopy is a field in materials characterization that uses electrons instead of photons in order to take images of materials and structures but also to analyse composition of materials at an atomic scale. This is important for many different applications in industry, such as metals, semiconductors, computer chips, catalysis, batteries, fuel cells, etc. and its impacts are visible everywhere in our day-to-day life.

The bottleneck with using electrons, however, is that they could affect the material and potentially could cause damage to the specimen that is being looked at. This can raise additional questions on top of the original research question such as, “Are we seeing the original structure of what we want to image or are the electrons changing this as we are looking at it?”

This is where my industrial partner, SenseAI Innovations, comes in. They specialize in AI imaging software primarily used in Electron Microscopy where they implement the principles of compressive sensing by using sparse scanning and reconstruction techniques. With this they can reduce the total electron dosage that a sample is exposed to during an experiment by only scanning a portion of the available positions and then reconstructing the sampled data into a full two-dimensional image that can be used for analysis.

The algorithm used by SenseAI is known for its speed and being agnostic to the data it sees, meaning it is widely applicable to different datasets for reconstructing and denoising. Due to factors in how data is acquired within a microscope, it becomes difficult to optimize all of the parameters when reconstructing the data we collect and as a result making it harder to see the structural features we are interested in.

My work focused on investigating the individual steps within the algorithm to explore how to extract the features needed for analysis of the materials. We do this at two different steps, first by preprocessing the data before it is seen by the algorithm and secondly by looking at how a reconstruction is performed and modifying this step where we alter the weighting of different elements to benefit how parts of the image are reconstructed.

These methods both help the algorithm used by SenseAI to extract more information out of the sparsely sampled data while also filtering some of the noise out from the higher frequency data needed for analysis. This is extremely useful for imaging highly beam-sensitive samples that are currently not robust enough when using such a scanning electron beam. Within the wider electron microscopy community, the exploration of data that was previously not attainable is becoming more and more viable.

Figure 1. Shows a reconstruction from data artificially sampled at 25% on the left, and then for comparison how the two described methods, method 1 in the middle and method 2 on the right, yield more information out of the reconstruction using the same parameters.

Keywords:

Electron Microscopy, Dictionary Learning, Compressive Sensing, Data Science