Faster Chemist II: Fast Harder

More tales from the accelerating edge of science

Jason Hein in his lab. UBC photo

Last week I told you about the Acceleration Consortium at the University of Toronto, which has received $200 million from the federal government to combine AI and automation to speed up the discovery of new materials. Response to that piece has been fantastic. I mentioned that I’d done way more reporting than I could fit into a single post, even a long one. Several of you told me you hoped for a sequel. Here it is.

The first post concentrated on Alán Aspuru-Guzik, the Mexican chemist who runs the Acceleration Consortium. But any good team has many players. And people who arrive at surprising destinations usually get there by surprising paths. Let me tell you about the head of the Acceleration Consortium’s Vancouver branch.

---

By 2011 Jason Hein had come a long way for a kid from Manitoba. After earning his BSc and his PhD from the University of Manitoba, he’d spent five years as a researcher at the Scripps Research Institute, a private, non-profit medical research colony in San Diego, California. If lab work is your thing, Scripps is hard to beat. One of the labs Hein worked in was led by Barry Sharpless, only the fifth person in history to win two Nobel Prizes.

Then Hein’s career took an odd turn. For his first faculty position, in 2011 he went to the University of California at Merced, the newest campus in the statewide University of California system. Merced is in the San Joaquin Valley southeast of San Francisco. Higher-education attainment rates there were historically low. Unemployment and poverty rates were well above California averages. Almost two-thirds of Merced students receive Pell grants, the most basic form of U.S. federal student aid.

To Hein, in many ways these were features, not bugs. The new campus would make a real difference in the community and in students’ lives. Instead of having to fit in, a young prof with an interest in biology, chemistry, physics and engineering could design his own research program. Hein was busy and well-liked.

But this still wasn’t Scripps. “The only grad students that ever went there, it was almost like a punishment,” Hein told me. “It’s like, ‘Well, you didn’t get in [to some prestige program], so I guess you’ve got to go to UC Merced.’”

Again, this situation had a substantial bright side. A chemistry program in the San Joaquin Valley was bringing new knowledge to a population that hadn’t had much access to it. But Hein’s forte was synthetic chemistry, the study of complex chains of reactions to make new materials. That took serious technical skill in the lab, and a lot of these students were arriving without much lab background.

“Nevertheless, having come from Manitoba, where we had to make do with limited resources and a challenging environment, I felt at home,” Hein told me. “Manitoba was already a crucible of sorts, where you had to exhibit skill and determination to navigate an unsupportive setting. I started seeing similar challenges emerging at Merced, where only the exceptionally resilient would thrive. This meant that many qualified and driven students would never complete their degrees unless something changed.”

There’s a familiar toolkit of techniques for bringing students up to speed on unfamiliar material. Basically it involves various sorts of remedial work. Hein could lay on a bunch of short courses or tutorials, or invite the students to up their game with extra lab time. None of that was ideal at Merced. First-generation university students aren’t typically busting out with spare time. Many of Hein’s were needed at home, especially those who came from migrant worker families, where a few hours’ manual labour could help pay bills.

So Hein decided he would automate as much of his lab work as possible.

“I turned to my tech roots and said, ‘We’ll build systems that can do that.’ Like, I don’t need a student that can come in as an expert kineticist. I can have a robot that knows how to do that part of it. It allows the student to focus on why we’re doing the reactions. The robot does the dexterity part. And it worked.”

Between 2005 and 2015, a bunch of cheap and highly adaptable platforms for electronics hobbyists came on the market, including Arduino and Raspberry Pi. These allowed Hein and his students to customize gadgets for handling lab tasks. A lot of those gadgets were available on eBay, a little raggedy but costing a fraction of their original retail price.

All this gadgetry didn’t free Hein’s students of the need to understand chemical reactions. In fact they had to be extremely precise about the steps they wanted the automated lab equipment to go through as it performed its experiments. That meant front-loading a lot of work. But once that work was done, the machines could take over and the students could go home to help their families.

“As soon as this started happening, it felt like I had tapped into a live wire,” Hein said. “I began recapturing my original dream of pursuing excellent science without sacrificing work-life balance.”

Second-hand hardware and off-the-shelf controllers meant Hein and his students had the luxury of failing fast: try something, find what’s not working, and adjust. He was able to make rapid progress. But the experimental design and the analysis of products was still all being done by human beings. That’s when Hein met Alán Aspuru-Guzik, who had started using AI algorithms to do some of the work people had been doing.

“I met Alán and his whole thing was, ‘Well, look, you’re building these robots. What if we put an AI in front of the planning part of it?’”

It was an intoxicating idea. By 2015, Hein had moved to UBC, and by 2019 he and his colleagues had built a superb self-driving lab called ADA, which today is hot on the track of new materials for solar cells, improved batteries, and other useful products.

There’s something kind of wonderful about the path Hein took to get to his position in the Acceleration Consortium. Chemists know how hard a lot of their work is, but they don’t love it. It’s exclusionary and it slows progress. By looking for solutions to the first of those two problems — the way the field’s technical complexity barred entry to too many newcomers — he began to glimpse solutions to the second.

Self-driving lab groups are comparing notes constantly now, Hein said. “It's like a hackathon now.” Many leaders in the field are getting ready to gather in Toronto at the end of the month for the consortium’s annual conference. Two of the big priorities now: standardization and democratization.

Right now, everyone in the world who has a self-driving lab built it themselves, using a large pile of grant money and considerable ingenuity. But the labs can’t talk to one another because they were designed as one-offs. There’s no industry standard, so there’s a vast amount of duplicated effort. Too much time designing self-driving labs and not enough time using them. It’s like the days before 1963, when ASCII code made it possible for a computer made by one company to send text to a computer made by any other company. “Everybody’s got to get going in the same direction,” Hein said. “Otherwise you get people running perpendicularly to each other.”

But the potential of these labs that can think and work for themselves is so vast that simply spreading the word is another big priority. So far, everyone who starts thinking about self-driving labs has come up with different ideas for how to use them. So Aspuru-Guzik and Hein and their colleagues want more people to start thinking about them.

That’s why Hein was especially excited that the Acceleration Consortium has landed a young researcher from Utah who’s only five years past finishing his bachelor’s degree. Sterling Baird is Acceleration Consortium’s first Director of Training and Programs. His job is to get researchers who haven’t been using automated labs to start thinking about it. He’s even made a Youtube video that shows how to make a rudimentary self-driving lab for under $100, as a way of getting more people familiar with the notion.

The more people start to imagine ways to incorporate AI into their research, the faster the whole field will evolve, Hein hopes. So far most labs are simply swapping out a single researcher for a single AI algorithm in ways that don’t really change the design of an experiment, he said. “It's akin to having a forklift that can only handle lightweight cardboard boxes, just as a single worker would.” But what happens when labs start to be rebuilt in ways that never would have been possible with humans alone? That’s the difference between a narrow improvement and a whole new ballgame. And it’s questions like this that are driving Aspuru-Guzik, Hein and their colleagues.