How Keith Brown’s lab is hastening materials research using machine learning and autonomous research robots
His team has built a “self-driving lab” to handle the grunt work of experiments, freeing them up to think more creatively about solving today’s biggest challenges
By Kat J. McAlpine
Keith Brown started out his scientific career by studying physics before pivoting to spend time doing materials research. The experience jumping between those two areas helped him identify a need to speed up the time it takes for research to reach the real world.
“I was left with a cognitive dissonance between where materials are today, and where they could be,” he says. “There is so much amazing promise in new materials, but the current process of research, development, and translation can take 10 to 20 years to go from a discovery to a product. You’d think we’d be faster than that by now.”
Today, his lab group at Boston University, where he is a faculty member at the BU Photonics Center, is motivated by an urgent desire to discover new materials that solve big challenges. “Climate, energy, food, water… we can’t wait a decade or two for more sustainable, green, and biodegradable materials to become available,” says Brown, who was recently awarded tenure, is an associate professor of mechanical engineering and materials science engineering at BU College of Engineering and an associate professor of physics at BU College of Arts & Sciences.
Brown’s team at BU is dedicated to accelerating the materials science pipeline. “The amount of time experiments require to design, perform, and analyze their results has been a key bottleneck for materials science research,” Brown says.
That’s why Brown’s group has turned to robots for help. “We’re developing autonomous experimentation methods, using lab robots to do faster experiments and do them more reliably,” he says. “And we’re also using machine learning and other tools to help us select the next best experiments to run.”
One of their focuses is looking at commercially available materials to investigate whether their structure can be optimized to improve energy absorption, performance, comfort, and other key measures of efficiency. They’ve created automated 3D-printing robots to assist their research.
“With additive manufacturing, we can make really complicated structures that minimize the weight of components while allowing us to tune other mechanical properties to our liking,” Brown says.
His lab’s autonomous 3D-printing research robots are running experiments over 90 percent of the time, sometimes not even shutting down when the lab team takes a break over the winter holidays. “The system is designed to be very robust, running all the time,” Brown says. “This type of system is known as a self-driving lab, because after we give it a goal and instructions, it begins making real-time choices to design and complete experiments.”
Even more interesting, he says, is the partnership that emerges between the human researcher and the robotic system. “When a human chooses an experiment, they often choose between refining promising experimental results or finding new areas to explore,” Brown says. “Using machine learning, we can program the system to choose the experiments that will give us the most knowledge by balancing these considerations. Thus, the goal is to try to craft a large enough ‘playground’ of parameters that enable the robots to access any conceivable outcome.”
With support from the U.S. Department of Defense, the group is working to develop better padding for use inside rigid helmets. “
Padding plays a really important role—energy needs to be dissipated to protect the head of someone who’s falling out of a car or landing on the ground after parachuting out of a plane,” he says. “The goal is to find new materials or structures that absorb energy better than traditional pads while also increasing the comfort of that padding for people wearing helmets?”
To give their lab’s robots the right parameters to perform experiments in, Brown says the team first had to consider how to teach the idea of comfort to a machine learning algorithm. “If it takes a small amount of force to compress a little but a substantially larger amount of force to compress a lot, that is a material that might feel comfortable while also providing protection against impact,” he says.
He is also interested in metamaterials, which rely on shape and architecture to change the mechanical properties of materials. Using their autonomous robots, Brown’s team has discovered a new way to structure a polymer—resembling a screw or a piece of fusilli pasta— that is extraordinarily efficient at absorbing energy.
“We’ve found that even when this structure fails, it does so in a way that preserves its ability to absorb energy. We imagine this could be a new type of material to explore using in a car’s crumple zones, or built into other structures designed to withstand damage,” Brown says. Despite its strength, the plastic-based material is also biodegradable. “In a landfill, it would disappear over about a century,” he adds.
Brown says the plastic in their metamaterial could be swapped out for aluminum to achieve similar performance, and aluminum products are easily recycled. A peer-reviewed publication describing the team’s findings is forthcoming.
Collaborating with Jörg Werner, a BU College of Engineering assistant professor of mechanical engineering and materials science engineering, Brown’s lab is also researching new materials that could solid-state batteries used in cars, cellphones, etc. “He’s the materials synthesis expert, and we’re the autonomous experimentation folks,” Brown says. Their joint research, supported by the National Science Foundation, is challenging the team to develop more sophisticated methods to allow for countless rounds of experiments to be carried out on an extremely small scale.
“If you want to run a million or even tens of thousands of experiments, the amount of material going into those experiments is not trivial. It could quickly add up to a lot of necessary supplies. One of the benefits of automating experiments is that you can develop systems that only require very small amounts of material,” he says. “Normally you might use a few microliters of supplies in experiments, but we’re striving to do experiments using a billion times less material.”
Today, there is a small but growing community of materials scientists focused on developing self-driving labs. “I’ve been fortunate to be a part of this community. It seems that there were very few of us exploring this type of thing a few years ago, but the community is really taking off now,” Brown says.
As he’s cultivated and grown his lab team at BU—affectionally called “the kablab” by his team in honor of Brown’s initials, K.A.B.—his top criterion for selecting new members has been creativity. “Especially when you’re running research robots that can handle the grunt work, the hardest question to answer is what do you want those robots to do? We think about this deeply and creatively, read scientific literature to formulate strong experimental questions, and do our best to ensure the research goals we’re giving these robots are good ones.”