AI Partnership: Infosys - Cambridge AI Center

Professor James Fergusson:

So welcome to this talk on the AI partnership between Infosys and Cambridge. I am Professor James Fergusson. the executive director of the Data Intensive Science Group here in Cambridge. I'm also the director of the Infosys-Cambridge AI Center. And so the question is, why does business like Infosys really want to partner with a university and cosmologists like myself on AI? And it's because we have the same challenge. And the challenge is that we are both facing the challenge of how do you work with exponentially growing data? Around when I started as a cosmologist, the high Z supernova experiment was the cutting edge. It had 40 data points, and it demonstrated that dark energy was real. And so this is the kind of experiment you could do very easily on your laptop. You can be friends with all 40 of those data points. You can know them very, very well. You can name them. But we're moving into an era more like with the square kilometer array observatory, which is going to put a million radio antenna across Australia and turn Australia into a telescope essentially. And that's generating about 710 petabytes a year, or around 173 terabytes of data a second. And this is really forcing us to transform in quite a radical way the way we think about data analysis and the way we think about doing science. And this really is something that extends to everyone, right? So we started a PhD program, which gave training in programming statistics and an AI. We connected with industry because we realized very early on that industry had the same problem and we could learn and work with each other, right? And so Infosys was one of our very early partners, right? Back in 2019, they came and they gave a talk to us and we discussed our problems and saw that there was lots of overlap. And so this led to creating other programs. On top of the PhD program, we created two training programs, one for our students, which is our Masters in Data Intensive Science, really taking people with physics backgrounds, preparing them for data-driven research in fundamental physics. And in parallel, we created an industrial AI program with Infosys, where we take Infosys employees and we prepare them for doing data-driven work in industry. And then on the top of both of these, we realized that actually it made sense if we collaborated on research as well. And so we created the Cambridge, Infosys-Cambridge AI Centre in 2025, really looking at key areas that we think will fundamentally change the way we do science, but will also change the world, right? And the three things we identified was using AI to enhance simulations, to look at the mathematical side of AI, and then also looking at agentic systems. So going through them, simulations. Our interest here is really that a lot of science relies on simulations, as does a lot of industry, industrial design, manufacturing, things like that. And so one of our key projects is building a foundation model for data, the same way we built a foundation model for language with ChatGPT. And hopefully this brings some of the same benefits, where we can go into low data areas or into very complex systems and we can simulate them or emulate them very, very easily with AI. So here's the simulation, the example on the right, which is a whole bunch of turbulent flows and we've given it on the bottom row some data and we've asked it to see if it could continue, wind that simulation on. And we see with very very small input, it's able to continue to evolve and understand that system for a very long time. And this is a bit like with ChatGPT. You can give it one sort of sentence prompt and it can generate a very large string of text that makes sense following that. This can be really transformative to the way we look at scientific problems, the way we simulate scientific problems, but also in any other process that relies on simulations. The mathematical AI side has two sort of sides of the coin. Using maths or physics to understand AI, which is where the Nobel Prize for Physics came in 2024, which is really the idea that we weren't able to build good steam engines until we had thermodynamics, which is actually a theory of how that engine works. And the same in the AI, we couldn't build really good AI systems until we have some kind of foundational theory of how AI works. And that's one of the key things we're trying to drive. The other side of the coin is how can we use AI to understand maths or to drive maths? And how can we use it to extract knowledge? And so one of the key things in this side is tools like Pisa, which does symbolic regression. And what this does is it takes data or a neural network and it tries to represent it in terms of an equation in an automatic way. And so you can see the demonstration here of building a tree, which is actually a way of representing an equation and then essentially constantly adapting that by evolutionary methods until we get a really good fit to the data. And so we've used this for things like putting in the orbits of planets and then we've been able to automatically spit out Newton's inverse square law for gravity. And this allows us to go straight from data to knowledge. Sort of reverse engineering the physics, traditional physics approach where you come up with a theory and then you get data to test it. Here we get data and then generate the theory directly. It's also really useful for interpretability in that you can do the same for a neural network. You can take a neural network and represent that in an equation which is much more interpretable and understandable. The final system is really looking at agents and how they can automate the work that we do. So we have the foundation models which have been sort of the revolution with ChatGPT leading the way. And we then put a special hat on them which allows us to tune them for specific tasks. And that allows us to make those very flexible models do very specific things in a reliable way. The analogy I think is most useful here is to think of these foundation models really as the engine, the other thing driving the AI revolution, but we need to harness them. And the way we do that is by using these agents to build a complex system that fits around that engine and helps narrow it down to do very specific things. So here, we have to build things like the wheels and the steering, but we also have to put in brakes and we have to put in bumpers so that those foundation models can be used in a very robust and interpretive way to do science. And I think the beauty of this partnership between Cambridge and business coming through this center has a few really key benefits, right? The first is that it allows researchers to learn about the key challenges in business, but also allows the businesses to learn about the cutting-edge research and AI. And this interplay and this sharing between the two worlds, I think, allows us to really think about things in a different way and will drive us to do more exciting work and tackle real world challenges. The other thing is something that we've seen a lot in the research side, which is the best work always comes from domain experts on the particular problem you're trying to solve, working with the main experts in the field of AI or methods. It's very hard to do really good work with only one side of that. You need to have these two people working together to solve things. And that's really the brilliance of this partnership with Infosys. Infosys via all its clients is a real expert on the kind of challenges businesses are facing. And then our key researchers can be the experts on how to implement on the cutting edge of AI. And working together, we think we'll get some really exciting solutions coming up.