It's important to know the magnetic structure of the material for many applications. Information like this is hard to find. There are only a limited number of machines that can support the analyses of magnetic structures.
The magnetic structures of only a small number of materials have been calculated. Even on large, state-of-the-art supercomputers, lengthy calculations are required to predict magnetic structures. Power demands grow as the size of the crystal structures under consideration goes up.
The MIT assistant professor of nuclear science and engineering and the MIT assistant professor of electrical engineering and computer science have found a way to use machine learning to streamline this process. Smidt says that the approach might be cheaper.
The results were published in a journal. The first authors of the paper are three MIT undergrads and a PhD student.
They were given a big challenge to design a neural network that could predict the magnetic structure of the materials. The neural networks that were co-invented by Smidt are what they used. The advantage of this kind of network is that we won't get a different prediction for the magnetic order if a crystal is translated or rotates. It's helpful for looking at 3D materials.
There are elements of structure.
A database of nearly 150,000 substances compiled by the Materials Project at the Lawrence Berkeley National Laboratory was used by the MIT group. Magnetic order and magnetic propagation were assessed by the team using this input.
There are three categories for ferromagnetic, antiferromagnetic, and nonmagnetic. The ferromagnetic material has small magnets with their own poles. The atom has a magnetic moment from the south to the north. The direction of the combined magnetic field produced by all of them is the same in a ferromagnetic material. In an antiferromagnetic material, the magnetic moments of the atoms point in a different direction than their neighbors, canceling each other out in an orderly pattern. All the atoms in a nonmagnetic material are nonmagnetic. The material's magnetic moments would point in random directions so that the net result is zero magnetism.
Magnetic propagation relates to the periodicity of a material. If you think of a crystal as a 3D arrangement of bricks, a unit cell is the smallest possible building block. The propagation value of the material is zero if the magnetic moments of the unit cells are aligned. The material is given a non-zero propagation value when moving from one cell to another.
There is a network solution.
The goals weren't worth much. How can machines learn to do things? To train the neural network to find correlations between a material's magnetic structure and its crystalline structure, the students took a portion of the Materials Project database. The students learned that they achieved the best results when they included not just information about the atoms' lattice positions, but also the atomic weight, atomic radius, and electronegativity. A lot of weights are fine-tuned during the training process.
Heiberger says that a weight is similar to the coefficients m and mx. The actual equation is a lot messier, with not just one coefficient but perhaps a hundred; x, in this case, is the input data, and you choose m so that y is predicted most accurate." Sometimes the equation needs to be changed to get a better fit.
The testing phase will come next. Heiberger says that the weights are kept as-is and that you compare the predictions you get to previously established values.
The model was able to predict magnetic order and propagation with an accuracy of 78 percent and 74 percent, respectively. Even if the material had magnetic atoms, the prediction of the order of nonmagnetic materials was still accurate.
The road is being plotted.
The MIT investigators believe that this approach could be applied to large molecule that have difficult to discern atomic structures. Smidt says that the strategy there is to take as big a unit cell as possible and try to approximate it as a somewhat disorganized crystal.
The authors wrote that the current work represents the first step insolving the grand challenge of magnetic structure determination. Smidt says that the "full structure" in this case means determining the specific magnetic moments of every atom.
There are some tricky details to be worked out but we have the math in place. It's a project for the future, but it looks like it can be done.
The undergraduates have already finished their work and won't be involved in that effort. They all enjoyed the experience. "Pursuing a project outside the classroom gave us the chance to create something exciting that didn't exist before."
The research began when the students were first-years. The work shows how we can expand the first year learning experience to include a research product. Every teacher wants to be able to support this kind of learning. It's great to see their hard work and dedication result in a contribution.
"This was a life-changing experience that I will remember for the rest of my life." It would be fun to combine computer science with material things. That decision turned out to be a good one.
There is more information about machine learning magnetism. The article is titled "Isci.2022.105192."
Journal information: iScience
MIT News ( web.mit.edu/newsoffice/) is a popular site that covers news about MIT.
All Rights Reserved © 2024
