MIT Develops AI Model to Identify Atomic Defects in Materials

MIT scientists have developed an artificial intelligence model designed to identify and quantify atomic-level defects within materials, a longstanding challenge in materials science. The breakthrough, published in the journal Matter, leverages machine learning to interpret data from a noninvasive neutron-scattering technique, allowing unprecedented insight into structural imperfections that can influence product performance.

Defects at the atomic scale, while often detrimental in biological contexts, are frequently engineered into materials like steel, semiconductors, and solar cells to enhance electrical, mechanical, or chemical properties. However, accurately measuring both the type and concentration of these defects without damaging the material has been a complex problem, forcing manufacturers to rely on indirect estimations or destructive testing.

The MIT team, led by lead author Mouyang Cheng and senior author Mingda Li, trained its AI model on a database of 2,000 semiconductor materials. For each material, they created paired samples: one with intentionally introduced defects through chemical 'doping' and one without. By analyzing neutron scattering data—specifically, the vibrational frequencies of atoms within these materials—the model learned to distinguish up to six types of point defects simultaneously, a task not achievable with current standalone methods.

Traditional techniques such as X-ray diffraction, Raman spectroscopy, and positron annihilation each reveal only part of the defect landscape. More comprehensive analysis used to require cutting thin samples for transmission electron microscopy, a process that can damage the material and is not scalable. The research team reports that their neural network model, employing a 'multihead attention' mechanism similar to those used in large language models like ChatGPT, can discern nuanced data differences to infer both the type and concentration of defects present.

Verification experiments demonstrated the model’s ability to estimate defect concentrations as low as 0.2 percent in alloys and superconductors. The system also performed accurately even when multiple types of defects were present at once, significantly reducing uncertainty for engineers aiming to optimize material performance.

Despite these advances, the researchers note that the practical deployment of their approach may currently be limited by the complexity and availability of neutron-scattering equipment. Raman spectroscopy, a more common and accessible technique in industry, is slated for adaptation in future versions of the AI model. The team is also working to expand the system’s capabilities to detect larger structural features such as grains and dislocations.

The research was supported by the U.S. Department of Energy and the National Science Foundation. If successfully adapted to more widespread laboratory techniques, the model could help manufacturers in sectors such as electronics, microelectronics, and solar energy achieve more accurate quality control without resorting to destructive testing.

“As products become more dependent on precisely engineered defects, our ability to measure these features with AI will play a crucial role in advancing materials science,” said Mingda Li. “AI’s pattern recognition can reveal subtle differences invisible to conventional analysis, opening a new era in quantifying and exploiting material defects.”

For further information, see the full paper published in Matter.

Source: news.mit.edu{:target="_blank"}