Argonne’s Autonomous Discovery Initiatives: Merging AI and Robotics to Accelerate Science
Aug. 17, 2023 -- Contrary to what you might have seen in movies, science is not a clean, linear timeline of “eureka” moments where a researcher at a small lab bench hunches over a microscope and eventually saves the world.
Today’s global problems — from plastic pollution to climate change to a years-long pandemic — take hundreds or even thousands of people working for thousands of hours on every aspect of the solution. And there are missteps, mistakes and even outright failures along the way. But new technological advances will help researchers make pivotal discoveries faster than ever before.
Enter autonomous discovery.
The U.S. Department of Energy’s (DOE) Argonne National Laboratory is working to speed the rate of discovery by automating certain aspects of laboratory research. Autonomous discovery will incorporate robotics, artificial intelligence (AI) and machine learning in the search for knowledge. Robotic arms can be programmed to fill and transport delicate or dangerous sample materials. AI can quickly process millions of data points to help researchers focus on viable solutions. These automated assists from robots and AI on some of science’s more monotonous tasks will free up scientists’ hands and brains to work on things that only humans can accomplish.
Rick Stevens is one of the researchers leading this effort. Stevens is the associate laboratory director for Computing, Environment and Life Sciences at Argonne. According to Stevens, “The way we do science, the way we sit at the lab bench and do science, hasn’t really changed in over 100 years.”
He said that while researchers have access to better tools and faster computers, they still run up against the limitations of performing experiments. A biologist might have to use a pipette to hand-transfer liquid from one test tube to another for several hours a day; a researcher can only analyze so many samples before they have to eat or sleep.
Kawtar Hafidi is the associate laboratory director for Physical Sciences and Engineering and a co-leader on the autonomous discovery effort. She summed up the current research roadblocks. “Science is inefficient, time consuming, labor intensive and costly. We need to free science from the constraints of human limitations.”
But just how fast is fast enough? To answer that, we can look into the biology lab and a current crisis in human health: drug-resistant bacteria. In the past century, penicillin was heralded as a wonder drug. Along with other antibiotics, the medicine saved the lives of countless people who would have died from bacterial infections. However, many strains of bacteria have evolved natural protections against antibiotics. And they’re evolving quickly. It takes the average pharmaceutical company about 10 years to develop a new antibiotic. But it only takes about four years for bacteria to develop an immunity against it. This gap between the need for an answer and our time-to-solution is just the kind of challenge that autonomous discovery will tackle.
“I want to speed the rate of discovery on an order of magnitude,” Stevens said. “By 100x or even 1,000x.”
“He’s doing his Steve Jobs impression,” Hafidi joked. But the cross-disciplinary team she and Stevens lead is taking that audacious look toward the future and bringing together several areas of Argonne’s scientific research and technological expertise to do it. They’re looking at similarities in the way different scientific disciplines conduct experiments and creating solutions that will be easy to implement for every type of experimental process.
It takes teamwork and a lot of trial and error to revolutionize science. And it takes creativity to imagine how humans will conduct research in the next 10, 50 or 100 years. Hafidi and Stevens’ team is studying the feasibility of creating a large-scale facility that is infinitely reconfigurable and can run thousands of experiments 24/7 with little to no human intervention. It seems like a goal Steve Jobs himself would appreciate.
“We want to make the impossible possible and the merely hard, easy,” Stevens said.
Making the Impossible Possible
Back to drug-resistant bacteria. Researchers hypothesize that they can solve this problem by creating or identifying new antimicrobial peptides. These molecules are part of all living things’ immune response systems. Researchers have to determine which peptides will work best at fighting harmful bacteria and which specific structures on the peptides will produce the anti-microbial effects they need.
As Casey Stone, one of the leaders in Argonne’s Rapid Prototyping Lab, explained it, “When designing even a moderately small length peptide, there are many, many more possible sequences for that peptide than any one scientist or group of scientists could ever experimentally test.”
Stone explained that even if human researchers could analyze one sequence per minute, it would take about 31,000 years to test all possible combinations.
By using the tools of autonomous discovery, researchers could test all the combinations in a few months or years.
But autonomous discovery goes beyond just analyzing huge amounts of data. There are some big advantages to using robotics in the lab. Besides handling dangerous materials, like those drug-resistant microbes, robots can perform menial tasks like pipetting — moving small samples of a liquid from one part of an experiment to another — with speed and accuracy.
Another big advantage: robots don’t get bored or tired.
Rafael Vescovi is a data scientist with the Data Science and Learning division who works with Stone in the Rapid Prototyping Lab. When he worked as an intern, he conducted research at Argonne’s Advanced Photon Source, a DOE Office of Science user facility. During this time Vescovi felt the limitations of being, well, human.
“When I developed the experiment, the bottleneck was how much time I could stay awake and how many samples I could do in the time that they gave me,” Vescovi said. “Everything I did was manual, so we would go hours pressing buttons, pulling and doing things.”
When he continued into his masters and Ph.D. programs, Vescovi developed a robot to help with his experiment. “I realized like, ‘Wow! I can do it 50 times faster because the robot is better than me!’”
That wow factor is on full display as the Rapid Prototyping Lab takes a maker culture approach to automating labs. It has become a sandbox environment where researchers and a steady flow of interns break down experimental processes, hack robotic arms and 3D print grippers and accessories that will meet the needs of several different scientific workflows.
As they look for new ways to integrate robots into the experimentation process, Stevens explained that the Rapid Prototyping Lab should “try lots of things, fail quickly and fail often. That’s the way to succeed.”
The 4th Industrial Revolution
Industrial robots have played a large role in manufacturing for decades, and pharmaceutical companies have put those techniques to use with robots that perform the monotonous tasks necessary for high-throughput testing.
The real seismic shift from a robot-powered assembly line to autonomous discovery is the incorporation of AI. “We’re coming to the point of what people are calling the fourth industrial revolution. That revolution is being driven by AI,” Stevens said.
Argonne is uniquely situated to generate and process data. The lab also has the capacity to use those huge data sets to train AI. From there, AI will be able to not only supervise experiments, but also determine what data seem promising and decide which experiments to run next.
“An easy way to think about it is like having a robot scientist,” said Kyle Hippe. Hippe is a first-year research scientist at Argonne who recently completed an internship at the Rapid Prototyping Lab. “This scientist would do everything from generating your hypothesis to testing it to analyzing your results and then formulating a new hypothesis in that kind of loop.”
Dion Antonopoulos, director of Argonne’s Biosciences division, stressed that while robotics are taking the drudgery out of benchwork, “the incorporation of AI is what the whole program is about.”
Antonopoulos said that AI is excellent for detecting patterns in results that humans might miss. When the system detects these patterns across huge data sets, it can run the next logical experiment automatically. “If it’s straightforward and easy enough to do, the system should just do it already,” he said.
Hippe and other AI experts are working on closing that experimental loop so that human scientists can focus on other things. “Biologists are generating more data than we could ever consume in a reasonable lifetime. The role of a biologist is becoming less and less about working at a lab bench and more and more about understanding that data,” said Hippe. He and his mentors believe that if they can train AI to drive experimentation, “we can put the pedal to the metal in all of these experiments.”
Training AI
While some researchers are working on automating specific workflows in biology or chemistry, others are figuring out how to expand across many different types of sciences at once. This is where an autonomous system could really kick discovery into high gear.
“Imagine a system that could support hundreds of experiments, both big and small, at the same time,” said Ian Foster, director of the Data Science and Learning division. “We’re talking about a system that could use both usual and unusual tools, schedule different experiments to run efficiently and continuously with those tools and detect and correct errors.”
Foster said a system like that would need to work smarter. His team is working on a range of AI methods that are trained to understand and handle complex and wide-ranging tasks. “We don’t want to have several different AIs trained to do several different types of science. We want one that understands all aspects of science. We want common methods and tools at every level.”
One method the team is exploring is large language models (LLMs). Even with the buzz surrounding recent breakthrough large language models like ChatGPT, scientists are just beginning to look for ways to integrate them into autonomous systems.
Building a broad and deep understanding of science is a central promise of generative AI that will be trained on enormous datasets and be able to predict patterns that it can use to develop better experiments. Researchers looking to develop large language models for science are training them with a variety of data, from scientific literature to chemical formulas to genetic information. This gives them the ability to make intelligent predictions for an extensive range of scientific questions, and — scientists anticipate — insight into how to ask better questions and develop better hypotheses.
Computational biologist Arvind Ramanathan described it like this: “Think of AI as an undergraduate student working in your lab or at your job. You have to train them and prompt them to look for the things you need.”
Ramanathan said that while the promise of the technology is great, researchers also have to understand the ethical and safety implications of using AI for science. “When it comes to keeping safety in focus,” Ramanathan explained, “humans are the guardrails.”
The better we train LLMs and other types of AI, the better they are at helping us solve big problems. AI can find hidden routes or new pathways to discovery.
To Serve Humans
If researchers truly “close the loop,” meaning they take humans out of the experimentation process completely, where does that leave humans? Are we science-ing ourselves out of science jobs?
Ilke Arslan said that just the opposite is true. Arslan is the director of the Center for Nanoscale Materials, a DOE Office of Science user facility, and the Nanoscience and Technology division at Argonne. She and other Argonne researchers are using autonomous discovery to study polymer deconstruction and reconstruction in plastics recycling. Their goal is to bring about a more sustainable plastics future.
“The need for the human to look at the results and assess whether it makes sense will never go away,” she said. “Computers can be trained to look for certain things and make decisions about next experiments in that way, but no computer can think and analyze like a human.”
In addition to making experimentation faster and safer, this new way of doing science may actually reduce people’s barriers to performing the experiments. “Autonomous discovery is moving us further away from being ‘classically trained scientists’ who specialize in a narrow topic with a specific domain that takes at least five years of advanced training to master.” Arslan explained that while earning a Ph.D. might still take five years, students might train in broad areas like biology or physics, tied with a good grasp of coding or robotics, rather than spending five years perfecting a technique in the lab.
“Broad experience and knowledge across disciplines and domains will be of even greater value,” said Arslan.
Hippe agreed. As a researcher who is just starting his career in science, he’s not worried about losing out to a robot. “The jobs that we lose in the sense that we’re automating them, we gain in the sense that we have to maintain this infrastructure. In the same way that the assembly line got rid of some manufacturing jobs, we’re shifting jobs to other places.”
This means additional jobs for everyone from engineers who design and build the machines to the technicians who keep them running to facilities crews who maintain the autonomous labs.
In every case, Arslan stressed that there are very human drivers for success. “We need creativity and curiosity!” she said. “The curiosity to solve any scientific problem will put our next generation of scientists on the right track.”
Flexible Thinkers Wanted
So autonomous discovery will actually create the jobs of the future. They’ll just look a little different than they did 100, 50 or even 10 years ago when a lone scientist sat hunched over a lab bench waiting for a eureka moment.
New and different ways of thinking are in high demand. Hafidi emphasized that workforce development will be key to the success of autonomous discovery. She and the team are working to enhance STEM training around AI, coding, robotics and automation.
“Students and early career researchers don’t have the biases that established researchers might have,” she said. “We need people like that. People who can be flexible thinkers.”
Hafidi stressed that there are two things students and early career scientists can bring to the scientific revolution: “We’re ready to welcome people with fresh eyes and crazy ideas.”
Source: Gillian King-Cargile, Argonne National Laboratory