In his lab at Osnabrück University, 2021 Schmidt Science Fellow Sebastian Musslick is excited about the possibilities of AI to accelerate discovery and push beyond the limitations of human brain power.
Collaborating with 2021 Schmidt Science Fellow Suhas Mahesh, they published a paper laying out where we currently stand with the automation of science and its future challenges and possibilities.
It is the urge to understand how the human brain works that drives Sebastian’s curiosity.
“I think it’s one of the greatest puzzles of humanity,” he says. “How does this blob of cells give rise to emotions and consciousness and essentially come up with all of the scientific theory we have? It’s pretty remarkable to me.”
Starting out in psychology, it was this question that drew him to a PhD in computational neuroscience.
But towards the end of his doctorate, he became frustrated by the field’s slow progress, in part because of the limitations of scientists’ brain capacity.
As a Schmidt Science Fellow, he pivoted to automated scientific discovery, the ambition to automate the entire empirical research process using AI and machine learning.
Sebastian explored how scientists in other disciplines used automated discovery techniques in their work, and this led to the collaboration with materials scientist Suhas Mahesh.
His collaboration with Suhas was enabled by a Schmidt Science Fellows Catalyst Grant.
The program, now in its third year, provides Fellows with early-stage seed funding for emerging ideas that may find it hard to attract investment from other sources at this stage, ensuring innovative collaborations have the chance to realize their potential.
So what can automation do for science and scientists? Firstly, says Sebastian, it can dramatically accelerate the process.
To demonstrate this, Sebastian has developed an AI-powered, open source Automatic Research Assistant (AutoRA) and collaborated with Suhas to put the system to work on very different problems.
Sebastian is using AutoRA to independently generate new discoveries for behavioural research studies.
The tool can build the experiments, recruit people around the world to take part, collect the data and then use that to formulate new theories and models of how humans behave, in an iterative, ‘closed loop’ manner.
AutoRA can do all this in less than a day, when previously it would have taken two or three months.
“It’s incredibly exciting and unprecedented. It enables us to search for much broader sets of theories”
“It’s incredibly exciting and unprecedented. It enables us to search for much broader sets of theories of how humans might behave and much broader sets of experiments,” says Sebastian.
It is here that the greatest opportunities for automation of science lie, he says.
Human scientists can only think about, and do, a limited number of things at the same time, but AI can ‘think’ much more broadly and join the dots in a way that people can’t.
AI also has the capability to address some of the ‘bottlenecks’ that restrict scientific discovery, as Sebastian and Suhas explore in their paper published in PNAS.
Take, for example, the long and laborious process of creating, testing, and evaluating new mathematical formulae or computational models to explain scientific data. AI has now discovered computational models that have so far escaped human intuition.
In 2023, Google DeepMind announced that its AI tool GNoME had discovered an incredible 380,000 new stable materials that could power future technologies. That’s the equivalent of 800 years’ worth of knowledge, if you consider the rate of discovery of stable materials made by people until now.
In a similar vein, Suhas has used a version of the closed-loop AutoRA system that Sebastian developed to discover new photovoltaic materials.
Suhas is now a Program Scientist in the AI & Advanced Computing Institute at Schmidt Sciences, where he’s leading the creation of a new program on AI for Science to be launched in 2026.
He is in no doubt of AI’s potential to create a paradigm shift in research.
“A favourite scientist of mine says there are two ways science advances: one is you find a new rock to look under, or you take an existing rock and find a new way of looking underneath it,” he says.
And AI is a new way of looking under every rock.
“It is very rare to find a tool that works across the entirety of science. The last time was probably the discovery of calculus. This doesn’t happen very often, and we need to take advantage of that.”
Some might reasonably question the wisdom of handing over the entire scientific process to the machines. What are the risks and the challenges?
It is imperative, says Sebastian, that we are clear with what we want to achieve in science and anchor science in society’s goals so there is always a ‘human in the loop’.
Automation must be constrained by the evolving needs of society so that it doesn’t speed off down undesirable paths.
He says we should also guard against machines learning at the expense of our own knowledge and understanding.
“We have to gain scientific understanding ourselves, in terms of explanation, characterization, control, prediction, and so on. We have to ensure that humans stay in the loop, otherwise we miss this epistemic goal of science,” explains Sebastian.
Technology is also a major barrier. AI is currently limited in what it can do by itself because robotic engineering is lagging far behind.
Suhas agrees that interoperability, ensuring equipment across the lab can work together seamlessly, is a current challenge in automation.
But perhaps the greatest challenge, says Sebastian, is how to handle the increased demand for peer review of scientific articles: a system that is already stretched to near breaking point.
“As we apply AI and automated science in our daily work, the sheer amount of research outcomes and articles that we can produce will vastly increase. Who should review those papers? That’s a big problem we have to solve. Because we don’t just want to hand over the review process to the AI scientists, unless we can perfectly verify what they’re doing. And we’re really far from that point at the moment.”
In the next five years, Sebastian predicts that automated scientific discovery will rapidly spread across scientific disciplines. Natural and engineering sciences are currently ahead of the crowd, but behavioural and social sciences will likely be next.
All of this may change the jobs of scientists so that they take a more creative, oversight role on setting research goals while machines do the heavy lifting, says Sebastian. So science education must keep up with AI.
To this end, the Schmidt Science Fellows Science Leadership Program has recently introduced sessions focused on AI in Research.
Sebastian says: “I think we will have to re-educate scientists in such a way that they are aware of the limitations of these automated techniques, to be aware of the potential mistakes that they can make, or the risks that they may encounter, and the potential biases that these systems have with respect to finding explanations in the data, for instance.”
AI should not, at least for now, take away the job of being able to evaluate science, Sebastian believes; it’s important that scientists still understand the scientific process from top to bottom.
But he has no doubt that changes are on the near horizon that will transform the job of the research scientist: watch this space.
Photos: Claudine Gossett
