Enhancing Cryogenic Scanning Transmission Electron Microscopy Efficiency with Machine Learning

Cryogenic scanning transmission electron microscopy (STEM) is an important characterization technique to study quantum phenomena and electron beam sensitive materials. However, the cryogenic temperature regime is susceptible to spatial distortions that reduce the quality of atomic resolution data. These distortions are a hardware limitation caused by thermal instability and cryogen bubbling and result in non-linear errors that cannot be easily removed. A common solution is to increase the acquisition speed to reduce the visible distortions, though this strategy results in much noisier data unless many frames are acquired and properly aligned. Non-rigid registration provides an effective solution to this problem, albeit at considerable computational expense using conventional algorithms. Further complicating this is that many materials have a low electron beam tolerance even under cryogenic conditions, making sparse sensing an inherent problem in addition to any acquisition errors.<br/>To overcome these challenges without expensive hardware investments, it is possible to use algorithmically-driven data acquisition and reconstruction to perform error correction, reduce data redundancy, and increase signal strength. We have developed a series of algorithms to enhance multimodal cryogenic STEM data quality through computational techniques. These include developments in the field of 4D-STEM compressive sensing and non-rigid registration.