In biology, defects are generally bad. But in materials science, defects can be deliberately adjusted to give the material useful new properties. Today, atomic-level defects are carefully introduced during the manufacturing process of products like steel, semiconductors, and solar cells to help improve strength, control electrical conductivity, optimize performance, and more.
But even though defects have become a powerful tool, accurately measuring the different types of defects and their concentrations in finished products has been challenging, especially without cutting or damaging the final material. Without knowing what defects their materials contain, engineers risk creating products that perform poorly or have unexpected properties.
Now, MIT researchers have created an AI model that is able to classify and quantify certain defects using data from a non-invasive neutron-scattering technique. The model, which was trained on 2,000 different semiconductor materials, can simultaneously detect six types of point defects in a material, something that would be impossible using traditional techniques alone.
“Existing techniques cannot accurately characterize defects in a universal and quantitative manner without destroying the material,” says lead author Mouyang Cheng, a PhD candidate in the Department of Materials Science and Engineering. “For traditional techniques without machine learning, detecting six different defects is unimaginable. This is something you can’t do any other way.”
Researchers say the model is a step toward more accurately exploiting defects in products such as semiconductor, microelectronics, solar cells and battery materials.
“Right now, detecting defects is like the proverbial thing about seeing an elephant: Each technology can only see a part of it,” says Mingda Li, senior author and associate professor of nuclear science and engineering. “Some people see the nose, some people see the trunk or the ears. But it’s extremely difficult to see the whole elephant. We need better ways to get a complete picture of the defects, because we have to understand them to make materials more useful.”
Cheng and Li are joined on the paper by postdoc Chu-Liang Fu, graduate researcher Bowen Yu, master’s student Eunbi Ra, PhD student Abhijatamedhi Chhotrattanapituk ’21, and Oak Ridge National Laboratory staff members Douglas L. Abernathy PhD ’93 and Yongqiang Cheng. This paper is published in the journal today Case.
fault detection
Manufacturers have become good at correcting defects in their materials, but measuring the exact amount of defects in finished products is still largely a guessing game.
“Engineers have many ways to introduce defects, such as through doping, but they still struggle with basic questions such as what kind of defect they have introduced and in what concentration,” Fu says. “Sometimes they also contain unwanted defects like oxidation. They don’t always know whether they have introduced some unwanted defect or impurity during synthesis. This is a long-standing challenge.”
The result is that almost every material has many defects. Unfortunately, each method of understanding defects has its own limitations. Techniques such as X-ray diffraction and positron annihilation characterize only certain types of defects. Raman spectroscopy can discern the type of defect but cannot directly estimate the concentration. Another technique known as a transmission electron microscope requires people to cut thin slices of samples for scanning.
In some previous papers, Li and colleagues applied machine learning to experimental spectroscopy data to characterize crystalline materials. For the new paper, they wanted to apply that technique to defects.
For their experiment, the researchers created a computational database of 2,000 semiconductor materials. They created sample pairs of each material, one doped for defects and one left without defects, then used a neutron-scattering technique that measures the different vibrational frequencies of atoms in solids. They trained a machine-learning model on the results.
“He created a fundamental model that included the 56 elements in the periodic table,” says Cheng. “The model takes advantage of the multidimensional attention mechanism, just like ChatGPT is using. It similarly extracts differences in the data between materials with and without defects and makes predictions about which dopants were used and at what concentrations.”
The researchers fine-tuned their model, verified it on experimental data, and showed that it could measure defect concentrations in an alloy commonly used in electronics and a different superconductor material.
The researchers doped the materials multiple times to introduce multiple point defects and test the limits of the model, ultimately finding that it could simultaneously predict about six defects in materials with defect concentrations as low as 0.2 percent.
“We were really surprised that it worked so well,” says Cheng. “It is very challenging to decode mixed signals from two different types of faults – let alone six.”
an ideal perspective
Typically, manufacturers of things like semiconductors run aggressive testing on a small percentage of products as they come off the manufacturing line, a slow process that limits their ability to detect every defect.
“Right now, people largely overestimate the amount of defects in their materials,” says Yu. “Checking estimates using each individual technique is a laborious experience, which only provides local information anyway. This leads to misunderstandings of what people think are the flaws in their material.”
The results were exciting to the researchers, but they noted that their technique measuring vibrational frequencies with neutrons would be difficult for companies to quickly implement into their quality-control processes.
“This method is very powerful, but its availability is limited,” says RHA. “Vibrational spectra is a simple idea, but in some setups it is very complex. There are some simpler experimental setups based on other approaches, such as Raman spectroscopy, that can be adopted more quickly.”
Li says companies have already expressed interest in this approach and asked when it would work with Raman spectroscopy, a widely used technique that measures the scattering of light. Li says the researchers’ next step is training a similar model based on Raman spectroscopy data. They also plan to extend their approach to detect features that are larger than point defects, such as grains and dislocations.
However, for now, the researchers believe their study demonstrates the inherent benefit of AI techniques for interpreting defect data.
“To the human eye, these fault signals will look essentially the same,” says Lee. “But AI’s pattern recognition is good enough to understand various signals and get to the ground truth. The flaw is this double-edged sword. There are many good flaws, but if there are too many, performance may deteriorate. This opens a new paradigm in flaw science.”
This work was supported, in part, by the Department of Energy and the National Science Foundation.
