At HRL Laboratories in Malibu, California, materials scientist Hunter Martin and his team load a grey powder as fine as confectioner’s sugar into a machine.
They’ve curated the powder recipe—mostly aluminum, blended with some other elements—down to the atom. The machine, a 3-D metal printer, lays the powder down a single dusting at time, while a laser overhead welds the layers together. Over several hours, the machine prints a small block the size of brownie.
HRL’s parent companies, Boeing and General Motors, want to 3-D print intricate metal parts in mass for their sleek new generation of cars and planes. Airbus has already installed the first-ever 3-D printed metal part on a commercial airplane, a bracket that attaches to its wings. But the tech is limited by the quality of today’s metal powders, says Martin. Most useful alloys aren’t printable because the atoms in the powder grains don’t stack correctly—leading to a weak, brittle weld.
So Martin’s group, which largely works at Boeing and GM’s forward-thinking HRL’s Sensors and Materials Laboratory, figured out how to alter the recipe of a strong alloy so it was compatible with a 3-D printer. Their secret weapon: a machine learning software made by Bay Area-based company, Citrine Informatics. It turns out, algorithms can learn enough chemistry to figure out what materials Boeing should use in their next airplane body.
Martin’s test block took more than 2 years of work. Scanning through the periodic table, his team came up with 10 million possible recipes for improving the powder. Then, they had to figure out which ones to try to make—using Citrine’s machine learning algorithms.
When companies upgrade their products—the next Prius, smartphone, or raincoat—they first consider how to upgrade the materials they’re made of. They could be improving quality, like making a harder glass for the iPhone, or figuring out how to make a cheaper battery. “Everything has to start with, what are we going to make it out of?” says materials scientist Liz Holm of Carnegie Mellon University, who has collaborated with Citrine in the past.
So in 2013, when Mulholland was still in business school, he and Citrine co-founders Bryce Meredig and Kyle Michel thought they could speed up that process. A crucial step is to pick the first recipe in the right ballpark, which usually takes the touch of an experienced researcher who has worked with similar materials for years. But instead of relying on one scientist’s limited experience, why not ask an algorithm fed with decades of experimental data?
To create these algorithms, they had to trawl for the data from those decades of experiments. They wrote software to scan and convert the data printed in heavy reference books from another era. They fed their algorithms the results of supercomputer simulations of exotic crystals. They built a friendly user interface, where a researcher can select from drop-down menus and toggle buttons to describe the type of material they want. Other than HRL, the Citrine team has partnered with clients such as Panasonic, Darpa, and various national labs in the last four years.
But even still, materials science projects suffer from a lack of data. “We have to do some creative things to really make the most of the data available,” says Mulholland. Unlike, say, the algorithms underpinning Google Translate, which are trained with millions of words, you might only have a thousand data points or fewer for a class of materials. Some companies want to work with materials only discovered a few years ago. To give the algorithms more to work with, Mulholland’s team teaches the algorithms general rules about physics and chemistry.
Sometimes they even have to resort to handwritten data. “There are times when we have to scan papers and notebooks from our customers, which is truly awful,” says Mulholland. “The norm is close to what my lab notebooks used to look like. It’s a series of hard-to-read notes, interspersed with chemicals dripped onto pages.”