Data mining used to find new materials

MIT researchers have integrated data mining tools and quantum mechanics to design a software which can help predict the crystal structures of materials. By using the same methods as online sales sites suggesting new books to customers, they claim they can determine in days the properties of atomic structures that might have taken months before.

In fact, this is a little bit more complex than that. But MIT researchers have successfully integrated data mining tools and modern methods of quantum mechanics to design a software which can help predict the crystal structures of materials. And to simplify, they say they've used methods used by online sales sites to suggest books to customers. And it seems to work: they claim they can determine in days the properties of atomic structures that might have taken months before. But read more...

Let's look at the introduction of this MIT document -- without forgetting it's a news release.

Using a technique called data mining, the MIT team preloaded the entire body of historical knowledge of crystal structures into a computer algorithm, or program, which they had designed to make correlations among the data based on the underlying rules of physics.

Harnessing this knowledge, the program then delivers a list of possible crystal structures for any mixture of elements whose structure is unknown. The team can then run that list of possibilities through a second algorithm that uses quantum mechanics to calculate precisely which structure is the most stable energetically -- a standard technique in the computer modeling of materials.

Two structure types strongly correlatedLet's illustrate the method with a couple of images on the right. "In Fe3C (a) and MgCu2 (b) the coordination of large (red: Fe, Mg) and small (blue: C, Cu) atoms is quite different, leading to a high energy on interchange of their positions." (Credit: MIT/Nature Materials)

And here is a more detailed explanation of why these particular materials have been chosen.

We observe a strong correlation between the Fe3C-type structure at a composition of AB3 and the MgCu2-type structure at A2B. These two structures occur together in 52 of the 87 alloys in which Fe3C is present, and using the fact that MgCu2 occurs in 7.04% of the alloy systems, the correlation ratio f is 8.49.
In other words, given that Fe3C is present at AB3, it is 8.49 times more likely that MgCu2 will form at A2B than if the structures were uncorrelated. In this case, the correlation can be easily understood as both structures (see on the right) form in systems where the two constituent elements A and B are of very different size.

After these highly technical details, let's go back to the MIT news release to learn more about the analogy between this software and the one used by Internet companies such as Amazon.

"We had at our disposal all of what is known about nature," said Professor Gerbrand Ceder of the Department of Materials Science and Engineering, leader of the research team. Ceder compared the database of crystal structures to the user database of an online bookseller, which can make correlations among millions of customers with similar interests. "If you tell me you've read these 10 books in the last year and you rate them, can I make some prediction about the next book you're going to like?"

And does this work? Apparently yes.

Ceder's team of computational modelers can already determine, in the space of just a few days, atomic structures that might take months or even years to elucidate in the lab. In testing on known structures of just two elements, Ceder's group found the new algorithm could select five structures from 3,000-4,000 possibilities with a 90 percent chance of having the true structure among the five.

This project started several years ago and you can read for example this earlier MIT news release, "MIT team mines for new materials with a computer" (November 17, 2003).

But the latest research work has been published by Nature Materials under the title "Predicting crystal structure by merging data mining with quantum mechanics" (Volume 5, Number 8, Pages 641-646, August 2006). Here are two links to the abstract and the full text (PDF format, 6 pages, 515 KB) of this scientific paper.

Sources: Eve Downing, MIT News Office, July 19, 2006; and various web sites

You'll find related stories by following the links below.

Newsletters

You have been successfully signed up. To sign up for more newsletters or to manage your account, visit the Newsletter Subscription Center.
See All
See All