How materials for computer chips, solar panels, and batteries are developed looks to be in the early stages of a radical change. The same goes for research related to areas like superconductors and thermoelectrics.

The reason? The new possibilities created by machine learning in materials science.

“This is something that is set to explode in people’s faces, as it were. Within the last five years, there has been a huge growth in materials science research teams using AI/machine learning techniques. The amount of scientific papers on the subject has been growing almost exponentially,” says Dr. James Warren, director of the Materials Genome Program in the Material Measurement Laboratory of NIST.

“We already see real-world advances based on the research, but I think we are only at the beginning. Machine learning could benefit every step of the scientific process for developing and improving new materials.”

Early Days but Real-World Solutions

I’m not an engineer, nor a scientist. I get to ask really smart people really stupid questions for a living. That is pretty much how I define being a journalist.

The way I think about materials science is that it’s about stuff. That’s also how I think about parts of engineering and manufacturing. It’s about putting stuff together. The quality of your finished product relies on the quality and abilities of the stuff used to make it.

This is why materials science is critically important to technological progress. Want a better computer chip? You need the right materials. More efficient batteries for self-driving cars or solar panels? Same answer.

A concrete example of how machine learning can aid the development of new materials comes from Stanford University where a team led by Evan Reed, assistant professor of Materials Science and Engineering, has been using it to develop better electrolytes for lithium-ion batteries.

Electrolytes are often composed of a range of materials. Finding the optimal combination and composition of said materials can be difficult.

“We have developed a machine learning model that has been outperforming experts’ intuition when predicting which materials to use,” Reed says.

New Examples  Abound

Valentin Stanev, a research scientist at University of Maryland, has been using machine learning in superconductor research.

“We have a list of all superconductors that we know of, but we still don’t have a good way of figuring out if something is going to be a superconductor. I applied machine learning to the process to help find ways to develop such a framework,” he explains.

Stanev sees big potential for machine learning in other areas too, such as the development of thermoelectric materials, which absorb heat and turn it into electricity.

“A huge percentage of our energy production is wasted as heat. Being able to catch just a small percentage of that will have an enormous impact,” he says.

Beyond superconductors and thermoelectric materials, scientists think machine learning could lead to advances in hydrogen storage units for fuel cells. In healthcare, it could help make new materials that better control how drugs dissolve a stint. It could also lead to new metallic glasses, a subset of metals without a crystalline structure, which have many possible applications, including nanotube construction.

Machine learning might even have applications in scientific processes themselves.

“Many processes in materials science rely on some sort of classification or fitting. Traditionally, this has been done by hand or some simple linear model after significant data processing,” explains Shyam Dwaraknath, computational chemist postdoctoral fellow at Berkeley Lab. “Machine learning makes these tasks much easier while improving the quality, speed, and amount of data that can be extracted. This has yielded automated methods for constructing phase diagrams, predicting structures for new compositions, and even analyzing micrographs in place of humans.”

Data Is the Magical Ingredient

There is still some way to go, though. The machine learning and materials science revolution is very much nascent. One area of development is sorting out where machine learning does and doesn’t make sense.

“The materials science community is actively seeking to identify the areas where ideas from machine learning could have an impact, with ongoing work ranging from materials selection problems to faster and more efficient data collection and analysis,” Evan Reed says.

Shyam Dwaraknath adds, “We’re just now entering the age of big data in materials science with large databases of well-curated and directly comparable data, but the true complexity of materials is far larger than that. For comparison, all the data on the internet, about a sextillion bytes, is just now reaching the number of atoms in a grain of sand.”

Another unsolved challenge? How to turn new, theoretical materials science insights into actual materials and solutions—especially on an industrial scale.

“It is like the difference in knowing the ingredients and knowing the exact recipe for, say, a soufflé. You need to know the exact process. That is the difference between ending up with a nice, light soufflé or a brick,” James Warren says.

The Up-Swinging Curve

While there are challenges, all scientists interviewed have high expectations when it comes to machine learning’s potential in materials science.

Valentin Stanev says new applications of machine learning in the scientific process could reduce the time needed to run experiments by up to 80%.

“You can have a machine learning toolbox built into your experimental setup. It looks at the results coming out of the experiment and can algorithmically decide what experiment to do next and from these deduce the general outcome of a series of experiments. In a way, you may only need to run 10 or 20% of the experiment to get the full picture,” he explains.

Other possibilities include handing partial control of experiments over to an AI system that autonomously makes decisions on what next steps to take.

And according to Evan Reed, machine learning might even be used for a kind of reverse engineering.

“Imagine that you need a battery that has a certain set of properties. You feed those into the machine learning model that then automatically runs through all available, known materials and suggests a range of batteries consisting of different materials that meet your specifications.”

James Warren sees potential uses coming sooner rather than later.

“Many of these advances are not nearly as far off as people think—in many cases we are talking a few years, tops. A lot of people in the community have a sense of, ‘What the hell just happened?’ Hopefully, others will too,” he says with a semi-laugh.

Warren believes machine learning is a key to future advances in the space, helping scientists push back the theoretical limits of materials and perhaps leading to development of many exciting new kinds of materials.

Image Credit: Jackie Niam / Shutterstock.com

Marc is British, Danish, Geekish, Bookish, Sportish, and loves anything in the world that goes 'booiingg'. He is a freelance journalist and researcher living in Tokyo and writes about all things science and tech. Follow Marc on Twitter (@wokattack1).

Follow Marc: