Inside Argonne’s push toward autonomous scientific discovery

The future of scientific discovery will use all the technological tools at humanity’s disposal, and no scientific field is better equipped to do so than autonomous discovery, which combines robotics and artificial intelligence (AI).

The U.S. Department of Energy’s (DOE) Argonne National Laboratory is a leader in this emerging specialty through its Robotic Autonomous Platforms for Innovative Discovery, or RAPID, labs. Inside these labs, robots conduct physical experiments and AI-assisted computer programs analyze the results and help determine next steps. This combination of AI and robotics has the potential to accelerate scientific discovery from tens to hundreds of times greater than without these platforms.

“We can go home, and we can sleep, and we can eat. [The robot] doesn’t get tired. It’ll just work overnight, and it’ll do more experiments for us. Then we can come back and look at the data and figure out next steps.” — Ilke Arslan, deputy associate laboratory director for Physical Sciences and Engineering

Depending on your point of view, the advent of autonomous robots carrying out scientific experiments might sound exhilarating or terrifying. But Ian Foster, director of Argonne’s Data Science and Learning division and an Argonne Distinguished Fellow, says that while right now is an exciting time, we’re still in the early stages of a new scientific practice.

“I think of these systems as being like the very first computers that we had when I arrived at Argonne 30 years ago,” said Foster, who is also the Arthur Holly Compton distinguished service professor of computer science at the University of Chicago. As he explains, new breakthroughs in large language models (LLMs) and chip technology are only beginning to reveal their scientific capabilities. But we already know they can do more than a human scientist.

“A human can plan and execute one experiment at a time, but they can’t perform 1,000 experiments simultaneously,” he said.

Foster’s whole career has been about projecting and manifesting the remotely possible. For his doctorate, in the 1980s, he worked in an AI group on logic programming, through which he discovered Argonne’s ongoing studies in parallel computing. Today, he co-leads Argonne’s Autonomous Discovery initiative, focusing on the computer science side — designing software that controls the robots and working on the AI methods for the experiments’ protocols.

“It’s all closely related,” Foster said. ​“My past work was about connecting computing and storage systems within an institution, and often across institutions, to tackle problems that we couldn’t tackle otherwise. Now what we’re doing is connecting computers, AI, storage and robotics. It’s all about enabling them to work together to do things that we couldn’t otherwise do.”

The interplay of robots and LLMs

When Foster mentions connections, he’s thinking as broadly as possible. The bridges between the RAPID labs have the potential to form across all scientific disciplines. Within these facilities, robots and AI-powered computer systems are experimenting in various areas of biology, chemistry and materials science. The RAPID labs are taking the first steps toward a future where discoveries are accelerated across many different scientific fields.

“For scientific inquiry, it’s a meandering path. We need to be agile and flexible, to pivot, to be able to move on from the workflow we’ve been doing before and to do it in an automated fashion. We’re trying to sort of take the first steps towards that.” — Dion Antonopoulos, Biosciences division director

Right now, it’s small steps. There are no humanoid robots wheeling around a lab with test tubes carrying out science experiments — though Ilke Arslan, Argonne’s deputy associate laboratory director for Physical Sciences and Engineering, says those might be just 10 to 15 years away. Arslan is a physicist by training and co-leads Argonne’s Autonomous Discovery initiative alongside Foster. On a spring day, she walks through a white-walled room about the size of an elementary school classroom in Argonne’s RAPID-200 lab. This space is used to study materials and chemistry problems in energy storage, quantum and microelectronics.

When you walk in, you see a command station with six monitors stacked in two rows of three. This is what Arslan calls ​“command central,” where AI-powered computers relay instructions to the robot. The robot in this case is an arm, a serpentine chromatic appendage around the size of a large table lamp, with a black claw attached to its top end. The arm moves back and forth across a slim motorized rail, picking up solutions and putting them into analysis instruments in a glass-walled container about the size of a small aquarium. At the moment, this station is experimenting with creating membranes and characterizing them to determine their efficacy and whether they’re able to filter out unwanted elements.

Before these machines were set up, Foster and others experimented with them in trial runs in Argonne’s Rapid Prototyping Lab (RPL). This is where researchers can test machines with water or other nontoxic chemicals before they’re moved to the RAPID labs, where the actual chemical solutions and materials will be introduced. Before robots conduct actual science experiments, the RPL is where every scenario and command will be gamed out, tested and amended to ensure the experiments are done safely, without any surprises.

This methodology also applies to the computers, which are feeding instructions to the robots and then processing and analyzing the results of the experiments. The computers are working from LLMs designed by Foster and his team to operate solely with the information provided during training. But this can be a lot of information.

“We’re looking at using LLMs to extract protocol definitions from scientific papers,” he said. ​“We mine large amounts of scientific literature and prepare that data to train models that we hope will be more knowledgeable than other LLMs in particular areas of science.”

Essentially, the computers form the most well-trained and knowledgeable team of scientists imaginable for the experiments they’re conducting. And as Foster points out, they can relay and process information humans simply cannot.

“A genome is basically a sequence of a four-letter alphabet,” he said. ​“You’ve got billions of these genomes, and if you train LLMs on large numbers of them, then they can start to do interesting things like produce other credible genetic sequences.”

Advances in biosciences

If you walk about 10 minutes south of RAPID-200, you’ll arrive at a low-slung, nondescript building. Behind one of the doors in a long, dark gray hallway, Biosciences Division Director Dion Antonopoulos and his team are applying robotics and AI toward biosciences in RAPID-350, Argonne’s autonomous lab for bioactive factors.

The windowless room is similar to, but slightly larger in scale than, the materials and chemistry lab. There’s a large, glass-walled enclosure in the center, but the equipment in here looks like stacks of VHS players attached to a long, thin conveyor belt. This is where Antonopoulos and his team are coming up with new antimicrobial peptides — molecules to outmaneuver antibiotic-resistant microorganisms.

The robotic arms in this lab are smaller in size, and fittingly so, as they’re meant to drop microliters of liquids into trays, then move them into large incubators, and finally place them under a spectral spectrometer, which takes photographs of the plate and determines if the sample is a success.

“What’s different with this whole enterprise isn’t just the automation,” Antonopoulos said. ​“It’s the idea of, I have a problem, I have a computational model of that problem. I’m going to go test and validate that model in the lab, and I’m going to do that in an automated fashion. And when I get those results, I’m going to feed it back into the model. It’s a more expansive sort of laboratory work that we’re trying to do.”

Arslan calls this a ​“closed-loop scenario,” where the computer, LLMs and robots are experimenting and circulating information until Argonne scientists achieve their desired results.

“Data is analyzed by the computer, and AI would make the decision about which experiment to do next,” Arslan said. ​“So, you would tell it, ​‘I want to optimize this parameter,’ and if it did not optimize that parameter, it will go back and start over again with a different set of solutions to try to create the optimal parameter you want.”

Is this whole system working faster than a human scientist? Not yet. But as Arslan points out, the hard limits of what humans can do in a single day is part of why AI and robots are already doing more than human scientists can do.

“We can go home, and we can sleep, and we can eat,” she said. ​“This doesn’t get tired. It’ll just work overnight, and it’ll do more experiments for us. Then we can come back and look at the data and figure out next steps.”

And of course, AI technology is always improving, meaning the time between these stages will only become shorter.

The lab of the future

While they speak enthusiastically about present initiatives, Argonne researchers are even more excited about forthcoming projects. Arslan briefly mentions the prospect of formulating new cathode materials in batteries. During her tour of RAPID-200, she walks to a separate room with a display on the wall of a robot arm surrounded by a honeycomb formation; this is where Argonne scientists plan to carry out experiments that must be created and analyzed in a vacuum (like outer space) because the air in our environment would damage the novel functionality of the materials scientists are trying to create.

In biosciences, Antonopoulos relates an initiative to develop novel proteins that are tuned to specific wavelengths of light. While discussing it, he recounts the history of the building that houses the RAPID-350 lab as a facility where scientists have tested new materials in controlled settings for safety purposes. This is where new frontiers in scientific experimentation take place.

The hope now is to scale up. Antonopoulos vividly lays out a big-picture vision of a warehouse filled with self-driving computers and robotics that can mix and match a variety of experiments. It sounds almost like the assembly lines of a factory, but he’s wary of the notion of industrializing experimentation.

“For scientific inquiry, it’s a meandering path,” Antonopoulos said. ​“We need to be agile and flexible, to pivot, to be able to move on from the workflow we’ve been doing before and to do it in an automated fashion. We’re trying to sort of take the first steps toward that.”

It’s the beginning of a potential future in which all scientific experimentation is being performed by robots and AI. There’s a natural tendency to fear that human scientists will disappear, but Foster doesn’t see it that way. He views this as an accelerator for scientists.

“The expertise of the people employed will probably change over time, as it always does,” he said. ​“But I don’t foresee fewer people being employed or scientists taken out of the loop. I see more science getting done.”

And not just ​“more,” but exponentially more, rocketing toward a future where the rate of scientific breakthroughs transforms life on Earth.