MIT researchers use AI to detect atomic defects in materials | MIT News

In biology, disability is usually bad. But in materials science, defects can be deliberately corrected to give useful materials new properties. Today, atomic-scale defects are carefully introduced during the manufacturing process of products such as metals, semiconductors, and solar cells to help improve energy, control electrical conductivity, improve efficiency, and more.

But since defects have become a powerful tool, accurately measuring the different types of defects and their concentrations in finished products has become a challenge, especially without opening or damaging the end product. Without knowing what defects are in their products, engineers run the risk of making products that don’t work properly or have unintended properties.

Now, MIT researchers have created an AI model that can isolate and measure specific defects using data from non-conventional neutron scattering techniques. The model, which was trained on 2,000 different semiconductor materials, can detect up to six types of defects in the material simultaneously, something that would not be possible using conventional techniques alone.

“Existing methods cannot accurately identify defects on a regular and quantitative basis without destroying the material,” said lead author Mouyang Cheng, a PhD candidate in the Department of Materials Science and Engineering. “In normal systems without machine learning, finding six different errors is unthinkable. It’s something you can’t do otherwise.”

The researchers say the model is a step toward using defects directly in products such as semiconductors, microelectronics, solar cells, and battery materials.

“Right now, fault finding is like the proverbial elephant sighting: Each process can only see part of itself,” said lead author and associate professor of nuclear science and engineering Mingda Li. “Some see the nose, others the trunk or the ears. But it’s very difficult to see the whole elephant. We need better ways to get a full picture of the disability, because we have to understand it to make the tools more useful.”

Joining Cheng and Li on the paper are Chu-Liang Fu, undergraduate researcher Bowen Yu, master’s student Eunbi Rha, PhD student Abhijatmedihi Chotrattanapituk ’21, and Oak Ridge National Laboratory staff members Douglas L Abernathy PhD ’93 and Yongqiang Cheng. The paper appears today in the journal Important.

Finding mistakes

Manufacturers have gotten good at fixing defects in their products, but estimating the exact number of defects in finished products is still a guessing game.

“Engineers have many ways to introduce defects, such as using doping, but they still struggle with basic questions such as what kind of defects they have created and what state they are in,” said Fu. “Sometimes they also have unwanted properties, like oxidation.

The result is that there are often many errors in each area. Unfortunately, each method of understanding errors has its limitations. Techniques such as X-ray diffraction and positron annihilation show only certain types of defects. Raman spectroscopy can detect the type of element but cannot directly identify the concentration. Another technique known as a transmission electron microscope requires people to cut small pieces of samples to be scanned.

In several previous papers, Li and collaborators applied machine learning to experimental spectroscopy data to characterize crystalline materials. In the new paper, they wanted to apply that approach to errors.

In their study, the researchers created a computer database of 2,000 semiconductor materials. They sampled pairs of each material, one doped with defects and one left intact, and used a neutron scattering technique that measured the different vibrational frequencies of atoms in solids. They trained a machine learning model on the results.

“That creates a basic model that includes 56 elements in the periodic table,” Cheng said. “The model uses a multihead approach, like the one used by ChatGPT. It also extracts the difference in the data between the materials with and without defects and predicts what dopants are used and at what concentration.”

The researchers fine-tuned their model, validated it in experimental data, and showed that it can measure the concentration of the element in an alloy commonly used in electronics and in a different superconductor material.

The researchers also ran the raw materials multiple times to introduce more defects and test the model’s limits, eventually finding it could make predictions about up to six defects in materials at once, with a defect concentration as low as 0.2 percent.

“We were really surprised that it worked so well,” Cheng said. “It is very challenging to determine the mixed symptoms in two different types of disabilities – let alone six.”

Model method

Typically, manufacturers of materials such as semiconductors conduct invasive tests on a small portion of products as they roll off the production line, a slow process that limits their ability to detect all problems.

“Right now, people overestimate the number of defects in their things,” said Yu. “It’s a painstaking experience to check the measurements using each method, which gives the location information in one letter anyway. It creates confusion about what flaws people think they have in their story.”

The results pleased the researchers, but they realize that their method of measuring vibrational frequencies with neutrons would be difficult for companies to quickly implement in their quality control processes.

“This method is very powerful, but its availability is limited,” said Rha. “Vibrational spectra is a simple idea, but in some applications it is very complicated. There are simple setups based on other methods, such as Raman spectroscopy, which can be quickly adopted.”

Li says companies have shown interest in the method 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 expand their method to detect features larger than point defects, such as grains and displacements.

For now, however, the researchers believe that their research demonstrates the potential benefit of AI methods for interpreting flawed data.

“In the eyes of people, these signs of disability would look the same,” said Li. “But AI pattern recognition is good enough to distinguish between different signals and get to the ground truth. Errors are a double-edged sword. There are many good errors, but if there are too many, performance can degrade. This opens a new paradigm in flawed science.”

The work was supported, in part, by the Department of Energy and the National Science Foundation.