AI Drives Autonomous Scanning Microscopy

Modern scanning microscopes achieve sub-atomic spatial resolution and sub-picosecond time resolution, enabling researchers to study materials at unprecedented scales. These capabilities generate massive datasets, imaging one cubic millimeter at 10-nanometer resolution produces a quadrillion data points. Acquiring and storing such datasets takes an extraordinary amount of time. Yet, the scientifically interesting information often concentrates in sparse regions containing interfaces, defects, or specific structural features. Traditional approaches require scanning every point, but prior knowledge about where these features exist is rarely available.

FAST solves this problem by teaching a computer to make smart decisions about where to look. The system uses artificial intelligence to study the data as it collects it, then figures out which spots to measure next. Like a photographer focusing on the most interesting parts of a scene, FAST focuses the microscope on areas where materials show differences or changes, skipping over less interesting flat regions. In a test at the Advanced Photon Source, scientists used FAST to scan a thin layer of material called WSe2. The AI found every tiny bubble trapped under the material by scanning just 20% of the sample, finishing in 40 minutes instead of 200 minutes.

FAST works without knowing anything about the sample beforehand. Scientists trained the AI using an ordinary photograph that has nothing to do with microscopy, yet it successfully guides many different types of experiments. The system runs on a small computer at the beamline and makes decisions almost instantly, adding less than 2% to the total experiment time. This makes FAST practical for everyday use at X-ray facilities and other research centers where scientists use scanning microscopes.