The defect in the silicon was invisible to the tools manufacturers actually use. It showed up in the neutron.
Researchers at MIT and Oak Ridge National Laboratory built an AI model that reads vibrational spectra from neutron-scattering experiments and extracts something conventional analysis cannot: which defects are present, in what concentration, without destroying the sample. The model covers 56 elements in the periodic table, detects up to six types of point defects simultaneously, and resolves defect concentrations as low as 0.2 percent. It is described in a paper published this week in Matter.
The "wait, AI can do that?" is earned. Conventional defect characterization is a collection of workarounds. Transmission electron microscopy requires cutting a sample into ultrathin slices — invasive, slow, and useless for high-throughput quality control on finished parts. X-ray diffraction and positron annihilation each characterize a single defect type. Raman spectroscopy identifies defect categories but not their concentrations. Manufacturers introducing defects through doping — intentionally, because defects are often performance features — have no reliable way to confirm what they actually created and at what quantity.
"Existing techniques can't accurately characterize defects in a universal and quantitative way without destroying the material," said Mouyang Cheng, a PhD candidate in MIT's Department of Materials Science and Engineering and the paper's lead author. Without machine learning, resolving six distinct defect types in a single analysis is unthinkable.
The model uses multihead attention — the same architecture underlying large language models — to learn the difference in vibrational spectra between pristine and doped samples, then predicts which dopants are present and at what concentration. Neutron scattering measures how atoms vibrate in a solid. Those vibrations are sensitive to what is in the crystal lattice and where. A dopant atom at a lattice site produces a different vibrational signature than the same atom at an interstitial site or a vacancy beside it. To the human eye, those signals look nearly identical. The AI sees the difference and gets to ground truth.
"Defects are this double-edged sword," said Mingda Li, associate professor of nuclear science and engineering at MIT and the senior MIT author. "There are many good defects, but if there are too many, performance can degrade. This opens up a new paradigm in defect science."
The catch is the measurement hardware. Neutron scattering requires access to a nuclear reactor or spallation source — large, expensive facilities that are not standing inside a semiconductor fab. The method is powerful but not deployable. The team's next step is building an equivalent model using Raman spectroscopy data, which uses scattered light instead of neutrons and is already widely deployed in semiconductor manufacturing. If that translation works, every fab with a Raman tool gets quantitative, multi-defect-type, non-invasive inspection — something does not exist today.
"Companies already use Raman-based tools extensively for semiconductor defect detection, and several industrial partners have asked when a similar AI-driven model could work with Raman data instead of neutrons," said master?s student Eunbi Rha, a co-author.
The practical stakes are real. Defects govern the performance of semiconductors, battery materials, solar cells, and structural alloys. Intentional doping — introducing specific atoms to tune electrical properties — is central to chip manufacturing. But if you cannot measure what you introduced and in what quantity, you are flying half blind. The MIT-Oak Ridge model, if it reaches industrial deployment via Raman, would close that gap.
The work received funding from the Department of Energy and the National Science Foundation.
