Neural Network Reads Defects in Metal ✎ express
Inside a piece of aluminum, atoms are arranged in neat rows, like bricks in a wall. But sometimes a row ends before reaching the edge, and a "step" appears in the crystal, which physicists call a dislocation. These defects determine how strong, ductile, and how a metal behaves under load. Seeing them directly is hard: you need an X-ray microscope that captures weak reflections from distorted atomic planes, and a trained eye to decipher the pattern of dark lines and figure out exactly where the layer has shifted.
Now a neural network is trying to do this job. It was trained on thousands of computer-generated images where single dislocations in aluminum were X-rayed at different angles. The network was preloaded with knowledge of crystal lattice symmetry: that is, the rules by which atoms can shift at all. Thanks to this, the model doesn't just memorize images; it "understands" which shift variants are physically possible. On test data it had never seen, the accuracy of dislocation type recognition reached 93%.
On a real sample with so-called cross-slip (when a dislocation changes its slip plane), the network correctly identified the shift type in 72.7% of cases. This is lower than in simulations, but already comparable to manual work by a human who spends hours comparing an image with calculated templates. The authors believe that training on simulations with physical constraints is the right path to fully automated defect analysis in materials.
For now, the network can only handle single dislocations in aluminum, but the same approach can be transferred to other metals and alloys. This will speed up inspection of parts for aviation, reactors, and microelectronics, where every defect counts.
Dislocations were theoretically predicted long before they were seen: in 1934, three scientists independently realized that metal deforms precisely through the motion of such "steps."