For hundreds of years, new materials were discovered through trial and error, or luck and serendipity. Now, scientists are using artificial intelligence to speed up the process.
Recently, researchers at Northwestern University used AI to figure out how to make new metal-glass hybrids 200 times faster than they would have doing experiments in the lab. Other scientists are building databases of thousands of compounds so that algorithms can predict which ones combine to form interesting new materials. Others yet are using AI to mine published papers for “recipes” to make these materials.
In the past, scientists and builders mixed materials together to see what formed. This is how cement, for instance, was discovered. Over time, they learned the physical properties of various compounds, but much of the knowledge was still based on intuition. “If you asked why Japanese watered steel was better at making knives, I don’t think anybody could have told you,” says James Warren, director of the Materials Genome Initiative at the National Institute of Standards and Technology. “They just had an artisan’s understanding of the relationship between that internal structure and awesomeness.”
Now, instead of using artisan’s knowledge, we can use databases and computations to quickly map out exactly what makes a material so much stronger or lighter — and that has the potential to revolutionize industry after industry, according to Warren. The time between discovering a material and integrating it into a product like a battery can be more than 20 years, he adds, and speeding up the process is bound to lead us to better batteries and glass for cell phones, better alloys for rockets, and better sensors for health devices. “Anything made out of matter,” says Warren, “we can improve.”
To understand how new materials are made, it’s helpful to think of a materials scientist like a cook, according to Warren. Say you have eggs, and you’re in the mood for something chewy and firm. Those are the properties of the dish you want, but how do you get there? To create a structure where both the white and the yolk are solid, you need a recipe that includes the step-by-step instructions for processing the egg — hardboiling it — just the way you want it. Materials science uses these same concepts: If a scientist wants certain material properties (say, light and hard to fracture), she will look for the physical and chemical structures that would create these properties, and the processes — like melting or beating metal — that would create these structures.
Databases and computations can help find answers. “We do quantum mechanical-level calculations of materials, calculations sophisticated enough that we can actually predict the properties of a possible new material on a computer before it’s ever made in a laboratory,” says Chris Wolverton, a materials scientist at Northwestern University who runs the Open Quantum Materials Database. (Other major databases include the Materials Project and the Materials Cloud.) The databases aren’t complete, but they’re growing, and already giving us exciting discoveries.
Nicola Marzari, a researcher at Switzerland’s École Polytechnique Fédérale de Lausanne, used databases to find 3D materials that can be peeled apart to create 2D materials of just one layer. One example of this is the much-hyped graphene, which consists of a single sheet of graphite, the material in a pencil. Like graphene, these 2D materials could have extraordinary properties, like strength, that they don’t have in their 3D form.
Marzari’s team had an algorithm sift through information from several databases. Starting from more than 100,000 materials, the algorithm eventually found about 2,000 materials that could be peeled into one layer, according to the paper Marzari published last month in Nature Nanotechnology. Marzari, who runs Materials Cloud, says these materials are a “treasure trove” because many have properties that could improve electronics. Some conduct electricity very well, some can convert heat into water, some absorb energy from the Sun: They could be useful for semiconductors in computers or batteries, so the next step is to investigate these possible properties more closely.
Marzari’s work is one example of how scientists are using databases to predict which compounds might create new and exciting materials. Those predictions, however, still need to be confirmed in a lab. And Marzari still had to tell his algorithm to follow certain rules, like looking for weak chemical bonds. Artificial intelligence can create a shortcut: Instead of programming specific rules, scientists can tell AI what they want to create — like a superstrong material — and the AI will tell the scientists the best experiment to run to make the new material.
That’s how Wolverton and his team at Northwestern used AI for a paper published this month in Science Advances. The researchers were interested in making new metallic glasses, which are stronger and less stiff than either metal or glass and could one day improve phones and spacecraft.
The AI method they used is similar to the ways people learn a new language, says study co-author Apurva Mehta, a scientist at Stanford University’s SLAC National Accelerator Laboratory. One way to learn a language is to sit down and memorize all the rules of grammar. “But another way of learning is just by experience and listening to someone else talk,” says Mehta. Their approach was a combination. First, the researchers looked through published papers to find as much data as possible on how different types of metallic glasses have been made. Next, they fed these “rules of grammar” into a machine-learning algorithm. The algorithm then learned to make its own predictions of which combination of elements would create a new form of metallic glass — similar to how someone can improve their French by going to France instead of endlessly memorizing conjugation charts. Mehta’s team then tested the system’s suggestions in lab experiments.
Scientists can synthesize and test thousands of materials at a time. But even at that speed, it would be a waste of time to blindly try out every possible combination. “They can’t just throw the whole periodic table at their equipment,” says Wolverton, so the role of the AI is to “suggest a few places for them to get started.” The process wasn’t perfect, and some suggestions — like the exact ratio of elements needed — were off, but the scientists were able to form new metallic glasses. Plus, doing the experiments means they now had even more data to feed back to the algorithm so it grows smarter and smarter each time.
Another way of using AI is to create a “cookbook,” or a collection of recipes for materials. In two papers published late last year, MIT scientists developed a machine-learning system that scans academic papers to figure out which ones include instructions for making certain materials. It could detect with 99 percent accuracy which paragraphs of a paper included the “recipe,” and with 86 percent accuracy the exact words in that paragraph.
The MIT team is now training the AI to be even more accurate. They’d like to create a database of these recipes for the science community at large, but they need to work with the publisher of these academic papers to make sure their collection doesn’t violate any agreements. Eventually, the team also wants to teach the system to read papers and then come up with new recipes on its own.
“One goal is to discover more efficient and cost-effective ways of making materials that we already make,” says study co-author and MIT materials scientist Elsa Olivetti. “Another is, here’s the compound that the computational materials science predicted, can we then suggest a better set of ways to make it?”
The future of AI and materials science seems promising, but challenges remain. First, computers simply cannot predict everything. “The predictions themselves have errors and often work on a simplified model of materials that doesn’t take into account the real world,” says Marzari from EPFL. There are all sorts of environmental factors, like temperature and humidity, that affect how the compounds behave. And most models can’t take those into account.
Another problem is that we still don’t have enough data about every compound, according to Wolverton, and a lack of data means algorithms aren’t very smart. That said, he and Mehta are now interested in using their method on other types of materials beside metallic glass. And they hope that one day, you won’t need a human to do experiments at all, it’ll just be AI and robots. “We can create really a completely autonomous system,” Wolverton says, “without any human being involved.”