When Alán Aspuru-Guzik, a Mexico City–born, Toronto-based chemist, looks at climate-change models, his eyes gravitate to the error bars, which show the range of uncertainty surrounding any given prediction. “As scientists,” he says, “we have a duty to contemplate worst-case scenarios.” If climate change proceeds as expected, humanity might have a couple of decades or so to come up with materials that don’t yet exist: molecules that enable us to quickly and cheaply capture carbon, and batteries—made of something other than lithium, a metal that is costly and difficult to mine—to store the global supply of renewable energy.
And what if the situation gets worse than we expected it to? The need for new materials will go from urgent to extremely urgent to dire. Could we quickly come up with the things we need?
Aspuru-Guzik (one of MIT Technology Review’s 35 Innovators Under 35 in 2010) has devoted much of his life to versions of this question. Materials discovery—the science of creating and developing useful new substances—often moves at a frustratingly slow pace. The typical trial-and-error approach, whereby scientists produce new molecules and then test each one sequentially for the desired properties, takes an average of two decades, making it too expensive and risky for most companies to pursue.
Aspuru-Guzik’s objective—which he shares with a growing number of computer-savvy chemists—is to shrink that interval to a matter of months or years, enabling humanity to quickly amass an arsenal of resources for fighting climate change, like batteries and carbon-capture filters. The goal is to revive the moribund materials industry by incorporating digital simulations, robotics, data science, artificial intelligence, and even quantum computing into the discovery process.
Imagine computer programs that use precise knowledge of molecules’ electronic structure to create new designs; imagine robots that make and test these molecules. And imagine the software and robots working together—testing molecules, tweaking designs, and testing again—until they produce a material with the properties we’re looking for.
That’s the idea, at least. Actually executing it is another matter. The structures of molecules are mind-bogglingly complex, and chemical synthesis is often more art than science, defying efforts to automate the process. But advances in AI, robotics, and computing are bringing new life to the vision.
Aspuru-Guzik cochaired a 2017 workshop in Mexico City where 133 participants—including Nobel Prize–winning scientists and representatives from 17 national governments—came together to focus the global research community on this goal. The conference was a pivotal moment, helping take the field of accelerated materials discovery from a niche area of inquiry to a worldwide priority for many of those attendees. After the event, Canada, India, and the EU, among others, began investing in initiatives to speed up material research.
The work itself is ambitious and technically difficult because it spans so many disciplines. But as a chemist, software engineer, AI pioneer, quantum computer programmer, robotics enthusiast, and serial entrepreneur, Aspuru-Guzik just may have the right mix of computational expertise and imagination to connect the multiple tools essential to making it happen. He has emerged as one of the more convincing evangelists for the new way of doing chemistry.
“Alán can see beyond what people think is possible,” says Joshua Schrier, a Fordham University chemist and frequent collaborator. He is the kind of innovator, says Schrier, who changes the way everybody around him practices science.
For Ryan Babbush, head of the quantum algorithms team at Google, Aspuru-Guzik’s most prominent character trait is his creative restlessness. “Alán spends his time and energy on the newest thing, the most uncharted territory,” he says. “He doesn’t stick around and focus on incremental developments.”
That can be a problem given the time and hard work it takes to bring a new material to market—an undertaking that requires dogged, narrowly focused research and endless business patience. But ultimately, Babbush says, Aspuru-Guzik is interested in reimagining the process of materials discovery, equipping scientists in the community with the computational and automation tools they need to speed up their job.
Today, Aspuru-Guzik is building a lab in Toronto where AI algorithms design novel molecules, and robots quickly make and test them. The lab is a kind of prototype, meant to demonstrate how materials discovery might work in the future. “I want to enable a whole new era, the age of materials on demand, where every lab can easily create new compounds,” he says. In the future, he hopes, we’ll be better positioned to address the next global crisis. “The problems of the world require molecules,” he adds. “And right now, we suck at making them.”
Aspuru-Guzik speaks exuberantly, digressively, and very quickly. When I first visited his office at the University of Toronto, he showed me a collection of lucha libre (Mexican wrestling) masks—bright blue, green, and pink balaclavas adorned with Aztec patterns. “The masks are a humanization tool,” he says. “You bring a Nobel Prize winner or an executive from Hitachi into your office, and after talking for a while, it’s good to stop and say, ‘Pick a mask. Take a selfie.’” It’s hard not to view the masks as a metaphor for his multifaceted life.
Aspuru-Guzik grew up in a half Catholic, half Jewish family of writers, musicians, and architects.As a 19-year-old chemistry student at the National Autonomous University of Mexico, he was returning from an overnight rave in the city of Cuernavaca when the car he was riding in veered off the road and crashed. Surgeons had to open his belly to repair his intestines and cauterize the ruptured blood vessels, leaving him with a scar that runs, like a median line, down the center of his abdomen.
After this early brush with mortality, he became committed to a life of intellectual adventurousness. If a field of inquiry intrigued him, he’d pursue it, even if it was esoteric or beyond his expertise.
At the time, there was great excitement over the possibility of using computer-based modeling to design molecules with desired properties, forgoing slow and tedious experiments. Scientists talked about a new era of virtual chemistry, only it didn’t work very well. Computers were too slow and molecules too complex.
While browsing journals in the university library, Aspuru-Guzik came across a paper about the challenges of doing molecular chemistry inside a computer. In 1926, the physicist Erwin Schrödinger had published an equation to predict the behavior of subatomic particles, like electrons and protons. If you can mathematically model a molecule at the subatomic level, you can begin to make inferences about the resulting material: how it combines with other materials, how hard or soft it is, or how quickly it decomposes. At least that’s the idea. But for most materials the Schrödinger equation becomes too complicated for even today’s largest supercomputer.
To make the math doable, Aspuru-Guzik set about creating versions of the equation that require fewer approximations, making them more accurate—a project that became the focus of his doctoral studies at the University of California, Berkeley. The goal was to streamline the calculations to the point where a computer could handle them but not so far that the model became scientifically useless. Using Aspuru-Guzik’s algorithms, a researcher could model—that is, simulate—a random molecule and immediately make predictions about the properties of the resulting substance.
Other scientists had designed similar algorithms, but the ones Aspuru-Guzik came up with as a grad student were impressive enough to get him a job at Harvard when he finished as a postdoc at Berkeley. As an assistant professor at Harvard—and as director of the Aspuru-Guzik research group, a 40-person team of computer scientists, biologists, engineers, physicists, and chemists—he threw himself into an initiative called the Harvard Clean Energy Project. Most solar panels use silicon to transform sunlight into electricity. But were there cheap, easy-to-make organic substances that could do the job?
Aspuru-Guzik’s passions (from top left) range from stickers for street art to lab robotics to Mexican lucha libre masks to automated fluid handling.
Over six years, Aspuru-Guzik and his team ran simulations of 2.3 million different organic molecules to see which might have photovoltaic properties. He was hardly the first researcher to practice virtual chemistry, but he was doing it at an unprecedented scale. The increased computing capacity of the era meant that a single molecule could be simulated in a matter of minutes; in the 1990s, such simulations had taken days. Most important, he had access to seemingly limitless server space, much of it borrowed from other people’s devices. In a system akin to the old SETI@Home program, people who wanted to support the project could download a screen saver that would temporarily lend their hard drive to Aspuru-Guzik and his team. “We had one of the biggest supercomputers in the world,” he says, “but it was distributed all over the planet.”
In the end, Aspuru-Guzik discovered many organic materials that could, theoretically, be used for photovoltaic cells. The problem was that these winning molecules were too complicated to be manufactured cheaply. “My mistake,” he says, “was not consulting with organic chemists at the beginning to find out which molecules were easily makeable.”
With the Clean Energy Project, Aspuru-Guzik had basically been doing combinatorial chemistry—the old trial-and-error approach—inside computers instead of inside a lab. Then, beginning in 2012, researchers in Toronto and elsewhere made a series of breakthroughs in deep learning and other methods of machine learning. Like many chemists looking for new materials, Aspuru-Guzik transitioned to AI, which enabled him to discover molecules in a faster, more deliberate way. “The computer simulations are like a machine gun shooting randomly in the air in the hopes of getting a hit,” he says. “AI is a sniper. It chooses a target and takes aim.”
First, he had to train a neural network by feeding it a data set describing the molecular composition and chemical properties of 100,000 organic substances. The AI program could start recognizing patterns—that is, correlations between a given molecule and the substance it forms. It could then use this knowledge to invent candidate molecules to be synthesized and tested in the lab. With the help of AI, Aspuru-Guzik discovered new organic light-emitting diodes, or OLEDs, that were brighter than typical LEDs. He also identified new chemicals to be used in future organic flow batteries, massive industrial batteries that won’t require metals like lithium.
Meanwhile, he threw himself into the nascent field of quantum computing. The Schrödinger equation is hard to run on classical computers precisely because electrons and protons don’t obey the laws of classical physics. They operate, instead, according to quantum mechanics: they can be entangled (behaving in concert with one another, even if they aren’t connected), and they can exist in so-called superposition (occupying multiple opposing states at the same time). The math required to model these complex phenomena is dizzyingly complex, too. But because the qubits in quantum computers also obey the laws of quantum mechanics, the devices are better suited, at least in theory, to simulating molecules.
In practice, though, somebody had to figure out how to make the simulations work. In 2014, Aspuru-Guzik and a team of researchers released the Variational Quantum Eigensolver (VQE), a program to model molecules, albeit on small, error-prone quantum devices that, unlike all-purpose quantum computers, actually exist today. While the Schrödinger equation is a kind of abstraction—a mathematical formula meant to describe subatomic particles—the VQE uses quantum bits to mimic the behavior of the particles in a molecule, much as players in a reenactment might perform the Battle of Gettysburg.
In time, as companies develop more powerful quantum computers, the VQE could enable chemists to run strikingly accurate simulations. These models might be so precise that scientists won’t need to synthesize and test the materials at all. “If we ever reach this point,” Aspuru-Guzik says, “my work in materials science will be done.”
When Donald Trump was elected president of the United States in 2016, Aspuru-Guzik’s career was flourishing, but suddenly the prospect of remaining in the country no longer appealed to him. One week after the election, he began emailing colleagues in Australia and Canada, looking for a new job.
The University of Toronto offered him a prestigious government-funded position meant to lure top-tier researchers to the country and a cross-appointment at the Vector Institute for Artificial Intelligence, a nonprofit corporation cofounded by machine-learning pioneer Geoffrey Hinton that is quickly making Toronto a global hub for AI. The biggest inducement, however, was a promise to build a radical new materials lab called the Matter Lab, a project Aspuru-Guzik had dreamed of for years.
“In the Matter Lab, we only attack a problem after asking three questions,” says Aspuru-Guzik. “Does it matter for the world? If not, then fuck it. Has somebody else already done it? If the answer is yes, there’s no point. And is it remotely possible?” Here, the word “remotely” is key. Aspuru-Guzik wants to tackle challenges that are within the range of feasibility, but barely so. “If a material is too easy,” he says, “let other people find it.”
Located in a postwar brick building in downtown Toronto, the lab is unlike any other at the university. The ceiling is adorned with maroon and burgundy acoustic panels, an homage to the beloved Mexican architect Luis Barragán. Tucked away in an inconspicuous corner is a typical lab bench—a table with flasks, scales, and beakers beneath a fume hood—where graduate students can practice chemistry in much the same way their grandparents’ generation did. One gets the sense that this workstation isn’t often used.
In the center is a $1.5 million robot—a nitrogen-filled glass-and-metal enclosure housing a mechanical arm that moves back and forth along a track. The arm can select powders and liquids from an array of canisters near the sides of the enclosure and deposit the contents, with exacting accuracy, in one of a number of reactors. “The robot is like a tireless lab assistant who mixes chemicals 24/7,” says Aspuru-Guzik. It can make 40 compounds in a mere 12 hours.
Two additional features make the Matter Lab’s experimental setup unique. The first is software that Aspuru-Guzik and his collaborators designed, called ChemOS. It includes an AI system that generates candidate molecules and a program that interfaces with the robot, directing it to synthesize candidates on demand.
The second distinct feature is the “closed loop” nature of the production process. To explain how this works, Aspuru-Guzik points to a pair of narrow hoses at the back of the robot. “That’s where the pee-pee comes out,” he says. Once a reaction is finished, the resulting liquid runs through the plastic hoses to an analytical machine the size and shape of a mini-fridge, which separates out unwanted by-products. The refined material will flow into another robot that will test it to learn about its properties. Then the robot will feed the results of the experiment back into the ChemOS program, enabling the AI to update its data and instantly generate a new, better slate of candidate molecules, until—after rounds of predictions, syntheses, and tests—a winner emerges.
The idea of an automated, closed-loop discovery system has, partly because of Aspuru-Guzik’s tireless advocacy, become increasingly popular among the new practitioners of chemistry. Peers in Vancouver, New York, Champaign-Urbana, and Glasgow are building similar facilities. These labs are intended as all-purpose, automated spaces of molecular creation. That’s why Aspuru-Guzik doesn’t speculate too much about what, specifically, the Matter Lab will produce next. Such decisions will be dictated by curiosity, perhaps, or by the imperatives of a global crisis.
Making a mark
In 2020, Aspuru-Guzik experienced a period of early-pandemic weight gain, which caused his surgical wound to reopen. At the same time, he felt trapped and bored by the 2D world of Zoom calls and frustrated at not being able to roam freely about his lab. His harried work life had left little space for the kind of aimless—or seemingly aimless—pursuits that, in the past, had fostered creative breakthroughs. He needed a change.
A few months later, he began doodling on his computer, drawing a lucha libre mask resembling Screamin’ Jay Hawkins, the rock ’n’ roll pioneer known for his operatic vocals and macabre stage antics. He named the character Bruho (a variation of “brujo,” Spanish for sorcerer) and decided to impose his artwork on the urban landscape. He bought a sticker printer and began plastering the Bruho avatar on mailboxes and streetlights. Soon he was part of the city’s bustling street-art scene.
Today, Aspuru-Guzik has two goals for the near future. The first is to design a modular, affordable version of his closed-loop system that can serve as a model to scientists around the world. He wants to build an all-in-one lab box, containing the ChemOS package along with synthesis and characterization robots. With this device, a user will punch in a description of a given material, and the system will immediately simulate and test candidate molecules. If we are to usher in a new era of materials on demand, Aspuru-Guzik reasons, the technology has to proliferate—and it has to be easy to use.
His second medium-term goal is to make his mark, artistically, on the city of Toronto.
A few days after my visit to the lab, I joined him and his crew for a night of stickering and postering. Like his materials work, this too was collaborative. Our eight-person group included Soap Ghost, an aloof young woman with full-sleeve tattoos; Urban Ninja, a wiry middle-aged man who arrived pulling a cart with a bucket of wheat paste, a homemade liquid adhesive; and Life, a flinty insomniac, his hair split down the middle, one half dyed blond like Cruella de Vil’s. “I’ll keep going until sunrise,” he told me, gamely. Everyone had bundles of stickers or rolls of posters they’d designed themselves.
In Toronto, this kind of street art—which doesn’t require spray paint—is punishable by fines (even though the police often look the other way), so we moved quickly and furtively. Ninja took us down an alleyway to a bare plywood wall of a boarded-up building, and we descended on it with our brushes, covering the surface with the paste and papering it with images—a bearded Buddha, a ukulele-playing rat, a Bruho figure, robed like a Jedi. The assemblage didn’t make a whole lot of visual sense, but it had a kind of anarchic beauty to it. Within an impossibly short time frame, emptiness had given way to multiplicity, and Aspuru-Guzik was thrilled. “This wall was blank a minute ago,” he exclaimed. “Look at it now.
Simon Lewsen is a Toronto-based magazine writer.
Inside the machine that saved Moore’s Law
The Dutch firm ASML spent $9 billion and 17 years developing a way to keep making denser computer chips.
The 50-year-old problem that eludes theoretical computer science
A solution to P vs NP could unlock countless computational problems—or keep them forever out of reach.
The US is worried that hackers are stealing data today so quantum computers can crack it in a decade
The US government is starting a generation-long battle against the threat next-generation computers pose to encryption.
This new startup has built a record-breaking 256-qubit quantum computer
QuEra Computing, launched by physicists at Harvard and MIT, is trying a different quantum approach to tackle impossibly hard computational tasks.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.