With resolution 1000-times greater than a light microscope, electron microscopes are exceptionally good at imaging materials and detailing their properties. But like all technologies, they have some limitations. To overcome these limitations, scientists have traditionally focused on upgrading hardware, which is costly. But researchers at the U.S. Department of Energy's (DOE) Argonne National Laboratory are showing that advanced software developments can push their performance further by using an artificial intelligence (AI) framework.
Electrons act like waves when they travel, and electron microscopes exploit this knowledge to create images. Images are formed when a material is exposed to a beam of electron waves passing through it. These waves interact with the material, the interaction is captured by a detector, and measured. These measurements are used to construct a magnified image.
Along with creating magnified images, electron microscopes also capture information about material properties, such as magnetization and electrostatic potential, which is the energy needed to move a charge against an electric field. This information is stored in a property of the electron wave known as phase. Phase describes the location or timing of a point within a wave cycle, such as the point where a wave reaches its peak.
When measurements are taken, information about the phase is seemingly lost. As a result, scientists cannot access information about magnetization or electrostatic potential from the images they acquire.
Retrieving phase information is a decades-old problem. It originated in X-ray imaging and is now shared by other fields, including electron microscopy. To resolve this problem, the Argonne researchers proposed leveraging tools built to train deep neural networks, a form of AI.
Neural networks are essentially a series of algorithms designed to mimic the human brain and nervous system. When given a series of inputs and outputs, these algorithms seek to map out the relationship between the two. But to do this accurately, neural networks have to be trained. That's where training algorithms come into play.
“Tech companies like Google and Facebook have developed packages of software that are designed to train neural networks. What we've essentially done is taken those and applied them to the scientific challenge of phase retrieval,” said group leader Mathew Cherukara.
Using these training algorithms, the research team demonstrated a way to recover phase information. But what makes their approach unique is that it also enables scientists to retrieve essential information about their electron microscope.
“Normally when you're trying to retrieve the phase, you presume you know your microscope parameters perfectly. However, that knowledge might not be accurate,” said Professor Tao Zhou. “With our method, you don't have to rely on this assumption. Instead, you actually get the conditions of your microscope — that's something other phase retrieval methods can't do.”
Their method also improves the resolution and sensitivity of existing equipment. This means that researchers will be able to recover tiny shifts in phase, and in turn, get information about small changes in magnetization and electrostatic potential, all without requiring costly hardware upgrades.