At the U.S. Department of Energy's Argonne National Laboratory, researchers have developed a project called DONUT — a neural network capable of analyzing data from scanning nanodiffraction X-ray microscopy. The solution helps determine crystal lattice orientation, deformations, and defects in experimental batteries and catalysts. Previously, scientists spent weeks manually comparing acquired images with computer simulations.
DONUT's architecture is based on the law of X-ray scattering, meaning the neural network does not require large volumes of expertly labeled data. It trains on the first frames of experiments and processes diffraction patterns 100 times faster than classical methods.
The technology could become the backbone of autonomous laboratories, where artificial intelligence would adjust experimental conditions in response to changes in the sample.