Q&A with ORNL’s Rob Moore, Part 2: Addressing challenges and what’s needed for autonomous science to reach its full potential
ORNL’s Rob Moore is a recognized leader in the development of autonomous science and self-driving laboratories. In this three-part series, he is discussing how artificial intelligence and automation have led to the cutting edge of autonomous scientific research. In part 1, he discussed the origins of AI in science. This time, Moore talks about the challenges of the lab’s INTERSECT initiative and what is necessary for autonomous science to reach its full potential.
Q. Explain the INTERSECT initiative and its objectives.
A. INTERSECT was really started as a way to explore federated instruments. We wanted to see if we could create an ecosystem where we could connect instrumentation across labs so we could have a free flow of information. We wanted to see if we could connect everything together in a way that we could help the scientists regardless of the challenge they were tackling.
In that process, it really became evident that this (ecosystem) is what was needed for developing self-driving labs. So federated instruments kind of morphed into “lets understand how to create self-driving labs, self-driving autonomous workflows.”
Q. What were some of the challenges or things that surprised you with INTERSECT?
A. One of the surprising things was that when we would get the experts in a room — in order to do this you need to have experts who have domain expertise, computing expertise, data expertise — but we found that all of these people speak different languages. It doesn’t take long before people have no idea what each other is saying.
We found you need to have domain scientists that have some working knowledge of what data ontologies are, along with computing experts who have some sense of what they are scientifically trying to achieve. This was kind of a struggle, because we have a deep knowledge in one area or another, but now this is something that requires very diverse skillsets from diverse sets of experts to make this happen. It turns out we all have different languages depending on our expertise, but we have to develop a common language so we can all get on the same page to figure out what we are trying to achieve.
Q. So how do you fix that?
A. The solution was just time — time trying to teach each other what we’re doing and the why behind it in order to come out on the same page and develop an outcome. There is no immediate solution to that. It’s just one of the growing pains of this community. And we’ve talked with academia, and they recognize it as well. There needs to be sort of a rethinking of the workforce that we need to develop these new self-driving processes and self-driving labs.
Q. Do you think new scientists will need this kind of expertise if they’re going to operate in the labs of the future?
A. I think — and a lot of the university people agree — that there is a change in the skillsets needed for people graduating and coming into this field. We still need the deep domain knowledge, but we need something extra. Now with self-driving labs, because of the diversity of knowledge needed to do that, it’s not sufficient to just have the hyper-focused effort scientists have traditionally had.
Now we’re in this mode of operation where we have to think about what we do and how it integrates into the bigger picture. Scientists sometimes have a hard time seeing the forest for the trees because we’re very detail oriented. Now, we need to be able to see the forest, as well, to see how our piece fits into the bigger puzzle.
Q. Are there things a human researcher brings to the table that can’t be replicated by AI?
A. Experience, intuition, creativity, perspective. We can offload a lot of cognitive loads to AI assistants for performing tasks, but the human condition brings experience and a perspective that cannot easily be replicated by AI. We have learned that bringing together people from different backgrounds helps facilitate solutions that help everybody. Diversity of perspectives is important to developing new ideas while maintaining a heathy world view. AI agents will become very “book smart” and can develop hypotheses based on existing information, but creating ideas that can transform thinking arises from humans comparing experiences when thinking about problems.
Q. Beyond speed and correlation, what things does AI excel at above what humans can do?
A. AI and automation can reproduce tasks with incredible precision. AI never gets tired (although some agents seem to be becoming lazy). Most importantly, AI can act as a sounding board for new ideas. Through casual conversations, AI can search existing information, develop calculations, develop code, and quickly generate information around the topic of the conversation. This means that humans can come up with crazy new ideas and quickly test out the viability of the idea with minimal effort. AI’s ability to instantly assemble, compile and assess knowledge acts as a force multiplier for human thought and intuition.
Q. Does this change the old perception of a scientist as a lone genius making breakthroughs isolated in their lab somewhere?
A. It will change that idea, but not by making individual insight less important. If anything, tools like AI can amplify human creativity and intuition in ways that make a single scientist more capable than ever before. What will change is that breakthroughs will be less about isolated events and more about how effectively a person can interact with data, automation, simulation, and autonomous systems. I’ve talked about this with people who do traditional fundamental science. A lot of them work in isolation or small teams. One of the things they always talk about is serendipity; about how many discoveries have been solved by chance. As we start developing self-driving labs, they wonder if we’re going to lose that happenstance and serendipitous chance of discovery.
On the flip side, one of the things these AI agents are really good at is finding correlations in data that we have not noticed. So, are we eliminating serendipity or developing a serendipity engine? That’s an interesting question.
UT-Battelle manages ORNL for DOE’s Office of Science, the single largest supporter of basic research in the physical sciences in the United States. The Office of Science is working to address some of the most pressing challenges of our time. For more information, please visit energy.gov/science. – Greg Cunningham
