Transistor channels in modern chips are 15 to 18 atoms wide. At this scale, a single missing atom changes how electrons flow. But until a 2026 study in Nature Communications (Karapetyan et al., Cornell/TSMC/ASM), nobody had directly imaged the atomic-scale defects inside working transistor structures. The manufacturing process was too complex, the features too small, and the standard imaging methods too blunt.
The technique is electron ptychography — a computational method where an electron beam scans across the sample while a pixel array detector (EMPAD, co-developed at Cornell) records the full scattering pattern at each position. By comparing how patterns change between positions, the algorithm reconstructs an image with sub-angstrom resolution. This detector has produced the highest-resolution images in the world, recognized by Guinness World Records.
What they found were “mouse bites” — small notches where atoms are missing from the edges of the channel. The name is apt: the defects look like something took a bite out of the channel wall. They form during fabrication — the etching, deposition, and annealing steps that build transistor structures atom by atom don't always leave every atom where it should be. At the dimensions of ten years ago, these imperfections were negligible. At 15 atoms wide, they're a significant fraction of the channel.
The defects aren't new. They've been causing variability in chip performance — different transistors on the same die behaving slightly differently — for years. What's new is seeing them. The imaging doesn't reveal a problem that didn't exist before. It reveals the specific geometry of a problem that was previously diagnosed only by its electrical symptoms. Knowing that a transistor underperforms is different from knowing that three atoms are missing from its northwest wall at position 12.
The practical value is in the feedback loop: now that fabrication defects can be mapped at atomic resolution, the manufacturing steps that produce them can be individually diagnosed and corrected. The bottleneck was never fixing the problem. It was finding it.