If we want to reduce fossil-fuel use enough to keep the world from warming catastrophically, we need to wring a lot more power out of the sun. A number of expert groups, including the Intergovernmental Panel on Climate Change, have concluded that we’ll need to get about a third of the world’s electricity from solar sources by 2030.
Right now, with less than a decade left to go, we’re at “about a tenth of that,” says Tonio Buonassisi, a professor of mechanical engineering and head of the Institute’s Photovoltaics Lab. To meet the goal, we will need to dramatically accelerate the deployment of solar power.
With that in mind, the PV Lab has been continually reconfigured over the past few years as Buonassisi and his colleagues bring in anything they can think of that might hasten the search for new solar materials. They’ve tried everything from stopwatches and personality tests to machine-learning algorithms and pipetting robots.
The result is a lab pace that’s orders of magnitude faster. They’ve condensed processes that once took six months or a year into two weeks; analysis of x-ray diffraction spectra that once required two to three hours can now be done in 5.5 minutes. “Right now it’s all about speed,” says Buonassisi.
The PV Lab started working with machine-learning algorithms in 2012. “We didn’t think of them as this whole new way of doing science,” Buonassisi says. “We just thought of them as a productivity tool.” But by 2017, he realized that the main question the lab was circling around was too complex to answer through traditional means. “I needed to change the way I did research in order to address it,” he says.
That question was how to find viable alternatives to silicon for use in capturing solar energy. At the moment, 95% of the world’s solar cells rely on silicon semiconductors. This element is abundant—found in almost all dirt and sand—and solar cells made with it are relatively efficient and tough. Your average silicon solar panel can convert roughly 20% of the energy from the sunlight that hits it, and it can work day after day for decades without breaking down.
But turning silicon into the thin, pure wafers needed for these cells is expensive and relatively difficult and energy intensive. It also often requires rarer materials, like silver. Experts, including Buonassisi, are working on improving these processes. But if we want to make solar a major part of the grid, bringing in some easier-to-manufacture materials will expand the possibilities—and, Buonassisi says, it could accelerate competition, driving innovation up and prices down across the board.
One promising class of materials is perovskites, natural and lab-made compounds with a crystalline structure that makes them good semiconductors. Perovskites are simpler and faster to manufacture than silicon wafers. Because they are compounds rather than elements, there could be a huge number of them that have yet to be created, says PV Lab member Jim Serdy. Different perovskites can also be stacked within a single solar cell, to absorb different wavelengths of light and squeeze more energy out of each sunbeam.
If perovskites manage to shoulder their way into the solar energy market, that may be “the critical path” for satisfying the world’s energy needs sustainably, Serdy says. But whether it’s possible is “dependent a great deal on how quickly we can discover these new compounds and their properties.”
The sheer number of perovskites is exciting—there could be thousands out there that match up well with different applications. But that also makes it a daunting task to search for the perfect material—one that accomplishes a desired task with the right combination of stability, efficiency, and cost-effectiveness, and that can also be manufactured easily and at scale.
In the past, scientists attempting to discover or invent a new material would start with some educated guesses. They’d create a few materials in the lab, test them, and then use what they learned to try again. Finding just one good option might take a year or more. “Statistically speaking, it’s like the monkeys at a typewriter”—banging away until they happen to write something useful, says Buonassisi.
The PV Lab’s researchers still follow this basic procedure. (“The scientific method isn’t going out of style,” he says.) They’ve just turbocharged it. Algorithms trained on theoretical knowledge and previous results help them make smarter guesses. And high-throughput experimentation and automated analysis allow them to test those guesses faster and run many tests in parallel. With all these advances, “we’re able to accelerate the whole process,” says research scientist Shijing Sun, team leader for the lab’s Accelerated Materials Development Program.
High-efficiency research doesn’t just speed up the pace of discovery—it also makes people bolder, says Buonassisi. “When you’re stuck in a mode of low throughput, you tend to play it safe,” he says. If you know you have more chances, though, “you can be a lot more ambitious.”
In the fall of 2018, for instance, Sun’s team began looking for more stable perovskites. (Although some perovskite solar cells are now as efficient as silicon ones, they tend to be more prone to degradation.) They started by identifying what they call their “search space,” in this case a group of 5,000 different possible materials to evaluate—all combinations of cesium, methylammonium, formamidinium, and lead iodide, mixed in different proportions and synthesized in different ways.
For their first round of experimental investigation, the group asked an algorithm to select 28 materials that provided a broad sample of the possibilities, says Sun. After synthesizing these materials, the team subjected them to high-throughput tools and techniques developed by Serdy and one of the lab’s technical associates, Janak Thapa. These tools allow them to test the materials’ stability quickly by exposing them to high temperatures, high humidity, and illumination—juiced-up versions of the conditions they might experience on a sunny rooftop.
“We are basically putting them into a sauna,” says Armi Tiihonen, a postdoctoral researcher in the lab. “We were aiming for extreme acceleration, to get the materials to degrade fast, because we didn’t want to waste months of time.”
To measure their stability, the team trained cameras on the materials, set to take photos every five minutes. Perovskites change color as they break down, often fading from near-black to a pale yellow. After the samples had spent about five days in the sauna, the team analyzed the photos to determine each material’s rate of degradation. (They also analyzed some of the samples more deeply using x-ray diffraction, to confirm visual observations and see how the materials’ structure changed as they degraded.)
Then they fed these results back to the first algorithm and asked it to pick 28 more materials—some similar to those that had been most successful in the experimental stage, and some from parts of the space that remained unexplored.
Even the most experienced materials scientists would have trouble making that kind of call, says Sun. “I can make a decision if we have 10 materials,” she says. “If we have 5,000 materials, I can’t really think about what to do next.”
The team went through this cycle a few times—choosing materials via the algorithm, creating and testing samples in the real world, and providing feedback to the algorithm. By the end of the fourth round, they had found a cluster of materials that were 17 times more stable than the most commonly used perovskite—as well as three times more stable than the lab’s previous record holder, which they had found through more traditional means. (Their findings and methods were published in February by the journal Matter.)
Other PV Lab projects have had similar success. In 2019, Sun’s team set out to find lead-free perovskites. They identified two materials that were completely new, along with four that had never before been made in the thin-film form necessary for use in solar cells. “In the past, it probably would have taken us over a year,” says Buonassisi. With the new methods, they were done in two months.
In another experiment, a perovskite solar cell built with one of these new materials proved more stable under harsh environmental conditions than the best one they had ever made with their previous methods, showing that the improvements in these individual materials carry through to the solar devices that are made with them.
The team is “fusing together the simulation and the experiment” to quickly identify and test promising materials, Buonassisi says. “We’re getting closer and closer to the point of being able to imagine something and then being able to realize it in real life.”
Life in the fast lane
Researchers in the PV Lab take the ever increasing pace in stride. In 2020, after 10 years of studying and working in bench science, Thapa started delving into machine learning; by the end of the year, he had coauthored his first paper on the subject.
“The personality of the lab is adaptability,” he says; members learn to do whatever the group needs. That’s true even for the lab’s undergrads. The goal is for any student who comes through to learn “how to be in a project and be in a team, and be a well-rounded, contributing member of the STEM community,” says Sara Bonner, the lab’s program administrator.
These goals can lead to unusual practices. Years ago, in order to figure out where they could save time, the lab borrowed a tool from factory floors of the early 20th century: “We literally had people with stopwatches, watching every step of the laboratory process and timing it,” says Buonassisi. On the basis of this analysis, they optimized their methods and invested in new equipment. They improved their sample preparation efficiency by 350%, going from 28 minutes per sample in 2015 to about eight in 2018.
More recently, he asked everyone in the lab to take personality tests, so they could learn to build on each other’s strengths and work together better. He sees these exercises as investments. “If we spend the time to develop this tool set that allows us to work more productively, then we can solve 10 times as many problems,” he says.
When the stakes are high, a breakneck pace can actually be a relief. At the start of her career, Tiihonen says, the work was so slow-moving that her goals always seemed out of reach. But now she and her colleagues can “actually achieve what we want.”
Sun likes the way the new techniques allow her to expand her area of expertise—where in the past the team might have focused on one parameter, or one class of perovskites, they now have “the opportunity to be able to step into more projects, and really get closer to that dream solar cell material,” she says.
The team keeps searching for bottlenecks in the process and widening them however they can. For the past few years, Buonassisi has been spending a lot of time in Singapore as part of the Singapore-MIT Alliance for Research and Technology. There and at MIT, he is beginning to incorporate robots that can perform some of the steps in the lab’s research pipeline. In Singapore, for example, a formulation robot mixes different chemicals into the compositions required for sample-making faster and more precisely than a researcher could painstakingly pipette them. It can do the nitty-gritty physical steps about four to 10 times as fast as a person can, and its precision helps improve reproducibility. What’s more, this robot can be remotely controlled, so lab members or collaborators anywhere can “queue up jobs and run them,” says Buonassisi. Meanwhile, he and a few collaborators are also working on a super-high-throughput tool that will help researchers in his MIT lab search through even more possibilities at once.
Although machines may be faster, humans are generally more adaptable. Using robots and similar tools when they’re helpful, rather than automating everything, allows the lab to speed up while preserving the “human-produced flexibility” that Buonassisi says is especially important for early-stage R&D.
But the ultimate speed-up, he says, comes when others adopt these methods and improve upon them. The PV Lab open-sources everything it does—from the stability-seeking algorithms to the blueprints for the machines—in order to “push these technologies out, and get more people excited about them and working on them,” he says. “We don’t have all the time in the world to wait.”
Finding the perfect recipe
An AI-driven research tool
looks to the past to find better ways
of making materials.
Often in order to forge ahead, we have to glance behind us. Elsa Olivetti, the Esther and Harold E. Edgerton Career Development Professor in MIT’s Department of Materials Science and Engineering, and her lab have been working on a set of algorithms—or as she likes to call it, a data science pipeline—that lets researchers search the scientific literature of the recent past to find clues about how to build the things we need for the future.
Olivetti’s group—which is focused on finding sustainable and affordable ways to design and develop materials—is always on the hunt for new tools, she says. A few years ago, she was talking to Gerbrand Ceder, then an MIT faculty member and creator of the Materials Project—a database of information about known and predicted materials, which researchers can use to find compounds that have the precise properties they’re looking for, even if they’ve never been made before.
Olivetti saw an opportunity to go further. While knowing what to make is a vital first step, “how one makes the material is what one needs to know in terms of the environmental and economic impact,” she says. In many cases, she thought, people have already done the work of making the material, and have meticulously recorded and published what steps were involved and how things went. Why not harness this resource?
Say you’ve been tasked with making a chocolate cake that bakes quickly and uses affordable ingredients. You could start from scratch: mixing components, tweaking ratios, and baking cake after cake until you hit on something that works. You could also trawl old cookbooks, watch online tutorials, and talk to trusted friends. But what if you had a machine that could look through millions of cookbooks, videos, and comment sections of baking websites, and compile the information it found into a new recipe that fit your purposes?
This is essentially what Olivetti’s tool does. Its users might want to make a solid-state electrolyte for a lithium-ion battery, or a low-carbon-
emission cement replacement. Rather than trying to read and synthesize past work in the area on their own or with a few colleagues, they can ask the tool to look through as much of the literature as it has access to—currently millions of papers and patents.
Olivetti’s tool combines natural-language-processing algorithms—which scan through papers to pull out relevant information—with neural networks, which recommend new recipes based on what has worked in the past. It seeks out information about the material in question, but also about different materials that may have related properties.
It has been challenging to create a set of algorithms that can distill so many papers, each with their domain-specific vocabulary and stylistic quirks, into useful recipes, Olivetti says. But the effort is already returning unexpected insights.
In 2019, she and some colleagues were working with zeolites, porous materials vital for applications from industrial catalysis to air purification. The size and arrangement of their pores affects what zeolites can be used for, but exactly how to control this attribute during synthesis wasn’t known. By using their algorithm to crunch the literature, Olivetti and her colleagues were able to deduce the crucial steps in making zeolites more or less porous—using the combined findings of past researchers to save future ones from endless rounds of trial and error.
In addition to serving up existing recipes for materials, an algorithm like this could help create new ones, says Olivetti. She can imagine incorporating a text-mining step into a workflow like Buonassisi’s, to bring a historical dimension into AI-driven attempts at materials synthesis.
It may also be possible, says Olivetti, to use similar technology to “pull broad themes from the field, or understand emerging trends—cool, broad opportunities that we’re just beginning to scratch the surface of.”
This new data poisoning tool lets artists fight back against generative AI
The tool, called Nightshade, messes up training data in ways that could cause serious damage to image-generating AI models.
Rogue superintelligence and merging with machines: Inside the mind of OpenAI’s chief scientist
An exclusive conversation with Ilya Sutskever on his fears for the future of AI and why they’ve made him change the focus of his life’s work.
The Biggest Questions: What is death?
New neuroscience is challenging our understanding of the dying process—bringing opportunities for the living.
Driving companywide efficiencies with AI
Advanced AI and ML capabilities revolutionize how administrative and operations tasks are done.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.