AI for Science: How Can AI Be Used for Discovery

Professor James Fergusson:

Welcome to this talk on AI for science and how it can be used for scientific discovery. I'm Professor James Fergusson. I'm the executive director of the Data Intensive Science Group here at Cambridge. I'm also the director of the Infosys-Cambridge AI Center. So if we think about science, really, when you go to school, what you hear is these stories of talented individuals making intellectual leaps. Exemplars, Newton, obviously, went to an orchard, sat down and thought heavily about the world and came up with this theory of gravity. Darwin went off in his boat and observed all of the finches in the Galapagos, came up with this idea of evolution. Einstein sat in his patent office in Bern, thought again about trains and light rays, and came up with the theory of general relativity. And this is really where the scientific method comes from. It's this brilliant individual sitting there coming up with theories. We then design experiments on how to test those theories. Those experiments gain data. We analyze that data. And then we either validate the theory or we reject it and go around and create a new one. I think with AI becoming more and more powerful, this paradigm is beginning to change. And so we're starting to see is that the new story of science is really being defined by these three key strengths of AI. The first one is that AI is perfectly happy learning very complex systems and predicting how they will evolve. So this is the key breakthrough of alpha fold. It's an area where it’s just computationally very difficult. You can't write down simple equations that explain where it's going to go. You can't write down high-level descriptions that allow you to see where it's going to go. You just have to grind through very difficult calculations to get the answer. And AlphaFold can automate all of that in a way that's better than all of the other models we can do, because it's really leading to the key strength of AI, which is that it's an automatic way to approximate any system, regardless of its complexity. The next real power of AI is its speed. It allows us to explore data and theory very, very rapidly. If I want to think about coming up with a new theory, I can come up with one or two or three, and then it takes quite a long time for me to check if they're right. But AI allows me to say, generate 1,000 theories and check all of them at the same time. Or it allows me to say, here is a particular theory. Please explore all the consequences of this. Solve the equations every way you can and give me all the answers. And this allows us to go much, much faster in the way that we explore science and narrow down to what the real true descriptions are. And the last one is the one that's really impacting every area of human knowledge, which is that these large language models can efficiently encode vast sums of human knowledge and make them easy to search and also can use this knowledge to automate tasks. And so this really has been a huge problem in science. If you really look at what stops a lot of scientists making progress or solving problems, is that they don't know everything. And so people working in biology may not know the relevant field of maths that they need to solve their problems. But AI allows you to automate that because you can say, this is my problem. What mathematical technique should I use? And it can also code it up and implement it for you in an automatic way. So it's really democratizing the whole sum of human scientific knowledge across all fields and breaking down barriers between fields in the way that it does this. I think this sort of gives us a new way of doing science, right? Where now we can actually just get an experiment and collect a lot of data, which is kind of what we're doing in some areas. We just build very, very large data collectors. We get all that data in, then we can use AI to explore that data to automatically generate and test theories and to build sort of science in an automated way. We're not having to do it in that theory first approach and then building things specifically. And I think there are some really exciting things that we can look towards the future that we might be able to do with these neural networks. We are hardwired to find some kind of calculations more easy than others. Babies have some concept of addition or multiplication. We have some basic understanding of classical mechanics. We think objects are permanent. We think of objects moving to the left or keep going to the left. If it's moving to the right, it'll keep going to the right. And sort of built into our brains. And so when we learn these in science, they're reasonably easy for us. We also really like linear systems, where if I do some of something, I get some back. But if I do more, I get more back, right? The things that our brains find really hard are things that are different to the kind of things we experience in everyday life. So when you first teach students quantum mechanics, they find it really difficult because it's very weird. Things can pop in and out of existence. You know, single things can go through two doors at the same time in a slit experiment. We also really struggle with turbulence and chaos. So this is systems where very small changes in the input can lead to very different outputs. And also multi-scale systems where lots of different scales are important at the same time. Very small scales and very large scales can interact. And the real power of neural networks or AI is it doesn't care about any of this. All it cares about is data and it will model things regardless of how complex they are. And so there's an exciting sort of hint that maybe we can use AI to understand systems that our brains don't like or that we find hard to model. Turbulence is very hard to model from a first principles thing because it's so unstable.

But we have seen that AI emulators can model turbulence better than we can. If we give it the data and say what's going to happen next, it will do better than numerical codes. Because it's learned some kind of new way of thinking about turbulence that we don't know that allows it to evolve it more reliably. And they said then If the exponential computing that powers it keeps going, what will science look like in the far future? Far on the scale of AI, maybe in 10 years time. Well, if it's finding these new ways of thinking about systems, could it actually discover new branches of mathematics or new physical laws in an automated way? Like if we give it enough data and say, please explain the system, will it come back with something that we haven't thought of or a new way of thinking about things that's not natural for people to do? And then with the automation side, can we actually get AI to do science all by itself? Can we give it data and just say, generate all the theories, test all the theories, and then just come back to me and explain how this data works? What is the mathematical description of this data? And I think there's tantalizing hints that this may be in some way possible. And certainly thinking five to 10 years away, I think we'll actually start to see a lot of this coming into normal science where we can just give it data and it will actually do a lot of the science on its own.