In biology, defects are generally bad. But in materials science, defects could be intentionally tuned to provide materials useful recent properties. Today, atomic-scale defects are rigorously introduced in the course of the manufacturing strategy of products like steel, semiconductors, and solar cells to assist improve strength, control electrical conductivity, optimize performance, and more.
But whilst defects have turn into a strong tool, accurately measuring various kinds of defects and their concentrations in finished products has been difficult, especially without cutting open or damaging the ultimate material. Without knowing what defects are of their materials, engineers risk making products that perform poorly or have unintended properties.
Now, MIT researchers have built an AI model able to classifying and quantifying certain defects using data from a noninvasive neutron-scattering technique. The model, which was trained on 2,000 different semiconductor materials, can detect as much as six sorts of point defects in a cloth concurrently, something that might be unattainable using conventional techniques alone.
“Existing techniques can’t accurately characterize defects in a universal and quantitative way without destroying the fabric,” says lead creator Mouyang Cheng, a PhD candidate within the Department of Materials Science and Engineering. “For conventional techniques without machine learning, detecting six different defects is unthinkable. It’s something you may’t do some other way.”
The researchers say the model is a step toward harnessing defects more precisely in products like semiconductors, microelectronics, solar cells, and battery materials.
“Immediately, detecting defects is just like the saying about seeing an elephant: Each technique can only see a part of it,” says senior creator and associate professor of nuclear science and engineering Mingda Li. “Some see the nose, others the trunk or ears. But it surely is amazingly hard to see the complete elephant. We want higher ways of getting the complete picture of defects, because we now have to know them to make materials more useful.”
Joining Cheng and Li on the paper are postdoc Chu-Liang Fu, undergraduate researcher Bowen Yu, master’s student Eunbi Rha, PhD student Abhijatmedhi Chotrattanapituk ’21, and Oak Ridge National Laboratory staff members Douglas L Abernathy PhD ’93 and Yongqiang Cheng. The paper appears today within the journal .
Detecting defects
Manufacturers have gotten good at tuning defects of their materials, but measuring precise quantities of defects in finished products remains to be largely a guessing game.
“Engineers have some ways to introduce defects, like through doping, but they still struggle with basic questions like what form of defect they’ve created and in what concentration,” Fu says. “Sometimes in addition they have unwanted defects, like oxidation. They don’t all the time know in the event that they introduced some unwanted defects or impurity during synthesis. It’s a longstanding challenge.”
The result’s that there are sometimes multiple defects in each material. Unfortunately, each method for understanding defects has its limits. Techniques like X-ray diffraction and positron annihilation characterize just some kinds of defects. Raman spectroscopy can discern the variety of defect but can’t directly infer the concentration. One other technique often known as transmission electron microscope requires people to chop thin slices of samples for scanning.
In a number of previous papers, Li and collaborators applied machine learning to experimental spectroscopy data to characterize crystalline materials. For the brand new paper, they desired to apply that technique to defects.
For his or her experiment, the researchers built a computational database of two,000 semiconductor materials. They made sample pairs of every material, with one doped for defects and one left without defects, then used a neutron-scattering technique that measures different vibrational frequencies of atoms in solid materials. They trained a machine-learning model on the outcomes.
“That built a foundational model that covers 56 elements within the periodic table,” Cheng says. “The model leverages the multihead attention mechanism, identical to what ChatGPT is using. It similarly extracts the difference in the information between materials with and without defects and outputs a prediction of what dopants were used and in what concentrations.”
The researchers fine-tuned their model, verified it on experimental data, and showed it could measure defect concentrations in an alloy commonly utilized in electronics and in a separate superconductor material.
The researchers also doped the materials multiple times to introduce multiple point defects and test the bounds of the model, ultimately finding it might make predictions about as much as six defects in materials concurrently, with defect concentrations as little as 0.2 percent.
“We were really surprised it worked that well,” Cheng says. “It’s very difficult to decode the mixed signals from two various kinds of defects — let alone six.”
A model approach
Typically, manufacturers of things like semiconductors run invasive tests on a small percentage of products as they arrive off the manufacturing line, a slow process that limits their ability to detect every defect.
“Immediately, people largely estimate the quantities of defects of their materials,” Yu says. “It’s a painstaking experience to envision the estimates by utilizing each individual technique, which only offers local information in a single grain anyway. It creates misunderstandings about what defects people think they’ve of their material.”
The outcomes were exciting for the researchers, but they note their technique measuring the vibrational frequencies with neutrons could be difficult for firms to quickly deploy in their very own quality-control processes.
“This method could be very powerful, but its availability is proscribed,” Rha says. “Vibrational spectra is an easy idea, but in certain setups it’s very complicated. There are some simpler experimental setups based on other approaches, like Raman spectroscopy, that might be more quickly adopted.”
Li says firms have already expressed interest within the approach and asked when it can work with Raman spectroscopy, a widely used technique that measures the scattering of sunshine. Li says the researchers’ next step is training an analogous model based on Raman spectroscopy data. In addition they plan to expand their approach to detect features which might be larger than point defects, like grains and dislocations.
For now, though, the researchers consider their study demonstrates the inherent advantage of AI techniques for interpreting defect data.
“To the human eye, these defect signals would look essentially the identical,” Li says. “However the pattern recognition of AI is sweet enough to discern different signals and get to the bottom truth. Defects are this double-edged sword. There are various good defects, but when there are too many, performance can degrade. This opens up a brand new paradigm in defect science.”
The work was supported, partly, by the Department of Energy and the National Science Foundation.
