The Future of AI in Physics

Roger Melko is an associate faculty member at Perimeter Institute and professor of physics at the University of Waterloo. His primary area of research is quantum matter: he and his students study how large groups of particles interact to create surprising new behaviours, such as superconductivity. 

But as the number of particles in a quantum system grows, the physics becomes more complex and harder to predict. 

That is why, long before the world at large became aware of artificial intelligence with the release of ChatGPT in 2022, Melko and colleagues from the Vector Institute for Artificial Intelligence in Toronto were already studying how “neural networks” could be helpful in studying phases of matter.

In 2018, Roger Melko launched the Perimeter Institute Quantum Intelligence Lab (PIQuIL) to explore how Artificial Intelligence (AI) could be applied to problems in quantum physics. 

Today, PIQuIL is going strong, and has forged partnerships with startups at the forefront of research in quantum computing and artificial intelligence, including 1QBit and Haiqu Inc, both quantum computing software companies, as well as chip developer Irréversible. PIQuIL researchers have even launched their own startups, like Canada-based yiyaniQ, which provides quantum solutions to financial firms.

The most recent startup, co-led by Melko alongside Crystal Senko and Rajibul Islam at the University of Waterloo’s Institute for Quantum Computing, is called Open Quantum Design. It is a non-profit organization committed to developing an open source quantum computer, to foster collaboration between academia, industry, and government and generate breakthroughs in quantum computing using AI.

In 2024, Melko and his Vector colleague Juan Carrasquilla (now at ETH Zurich) published a perspective paper in Nature Computational Science describing how the algorithmic structure used by large language models (LLMs) can also be used to advance quantum computing.

Melko has become a resident expert in applying artificial intelligence to the field of theoretical physics. We sat down with him to talk about how artificial intelligence is changing the way we do physics.

“The sky's the limit for applications in physics. It is only limited by our imagination.” 
— Roger Melko

The following conversation has been edited for clarity and length:

Q: When did physicists begin to use AI, and how is it being used today in physics?

More than a decade ago, the power of what we then called deep learning was becoming apparent (AlexNet, a machine learning program created by Toronto-based researcher Geoffrey Hinton, was released in 2012).

Physicists began to see the power of these deep learning algorithms and started to adopt them into their workflow. That’s how we got into it, and we launched the Perimeter Institute Quantum Intelligence Lab (PIQuIL) in 2018.

Another big breakthrough happened in 2020, when OpenAI researchers led by Jared Kaplan published a paper that said large language models could make predictions based on scaling laws. They showed that certain types of what we now think of as large language models could be made to perform better and better in a predictable way, by scaling them bigger and bigger. . And scaling meant three things: scaling the size of the model, scaling the amount of data that you train it on, and scaling the amount of compute that you use to do the training.

Previous breakthroughs such as deep learning and the discovery of the scaling laws precipitated more watershed moments such as the subsequent launch of ChatGPT models.

Today, I would say the sky's the limit for applications in physics. It is only limited by our imagination. We see it in every field. If you look at high energy physics, if you look at astronomical surveys, if you look at telescope data, or condensed matter physics and materials physics and quantum computing, there isn’t any field of physics, I will argue, that's left untouched by AI now.

Q: What types of AI are being used in physics today?

There are broad categories. Ten years ago we used to think in terms of supervised learning, unsupervised learning, reinforcement learning. Canadians played a large part in developing all these paradigms.

But with large language models (LLMs), the taxonomy got blurred — that paradigm combines these different concepts together. The idea underlying modern LLMs is that if you have sequences of words (or what they call “tokens”), the trained model can predict the next token of a sequence.

There are different types of AI, but I would say LLMs have come to dominate the discourse in physics. 

Q: What role has Perimeter played in the development of AI for physics?

Perimeter has played a very important role in many areas. One example I can point to is what we call normalizing flow models, which are a type of unsupervised generative model used for simulations. These have been valuable, for example, in simulating lattice field theories like Quantum Chromodynamics, a fundamental theory in physics that describes how quarks and gluons interact.

There is an interesting story about how Michael Albergo (from New York University), a visiting fellow at PIQuIL, and Phiala Shanahan, a former Simons Emily Noether Fellow at Perimeter  (now at MIT), serendipitously met at Perimeter and launched an entire field of physics based on using AI generative flows for high energy simulation.

They just crossed paths in the Bistro at Perimeter, started talking and did a couple of seminars, and suddenly, a whole field was launched. I have been on research trips recently, to MIT and the Kavli Institute for Theoretical Physics for example, and I was surprised by how many people are working on normalizing flows for lattice field theory.

Q: What is the latest at the Perimeter Institute Quantum Intelligence Lab (PIQuIL)?

PIQuIL is still going strong. We have been running for eight years now and we are still at the forefront of artificial intelligence, quantum computing, and research in quantum condensed matter.

Startup companies have picked up residency at PIQuIL and have launched out of PIQuIL. It has been a great training ground. Our masters’ students, even PhD and postdocs have gone on to industry through us.

Our biggest development in the past two years has been the launch of Open Quantum Design. I think it has been an incredibly successful startup. It is an open source foundation and is structured as a non-profit. It is dedicated to delivering a quantum computer, where the design is democratically developed by a whole community of open source enthusiasts. So it’s been a busy two years for Open Quantum Design since being launched out of PIQuIL. 

Q: How can Perimeter help in the development of AI?

Perimeter is one of the only places in the world where it is possible to think deep thoughts about AI and why it works the way it does.

There are theories that we can develop that underlie artificial intelligence. The breakthroughs that happened in 2020 which told us about the scaling laws that enabled LLMs are still not completely understood. Yet those scaling laws, which give you a predictable way of growing LLMs so that their performance continuously improves, are very crucial for industry. They have changed the world. At Perimeter Institute and at PIQuIL we have the luxury of being able to think about how theoretical concepts that might explain these scaling laws, which might lead to ways to improve them. It might lead to completely new paradigms in artificial intelligence in the future.

In fact, I suspect that as Perimeter expands going forward, developing the underlying theory for both today's artificial intelligence and the AI models of tomorrow will become more important.

Q: Most people think of large language models like ChatGPT in terms of writing essays or poetry. But how can that be applied to physics?

In large language models, the AI is predicting the next word (or “token”) in a sequence. So, for example if you start with ‘Hickory Dickory Dock,’ what are the next words? If it has consumed enough data during training, it will come back with “the mouse ran up the clock.’ This is what LLMs are exquisite at. They are, in a sense, learning a probability distribution that underlies sequences of words. They are good at completing sentences using probabilistic rules.

So how do you apply that to physics? We can do it by mapping probability distributions in sequences. So, for example, I work on sequences of qubits (quantum bits) and their measurement outcomes. At the University of Waterloo we have a quantum computer that involves a row of atoms, and each atom is qubit, When I measure that, what will the sequence look like in terms of qubit outcomes? It might come out as down, down, up, down, and so on … or some other sequence. An AI based on an LLM would be able to then finish the ‘sentence.’ It'll give me the most probable outcome based on the rest of the qubits in the sequence.

Q: But can AI actually “reason” and come up with solutions to questions that humans haven't been able to resolve? Or is it just mimicking what humans feed it?

At the base level, it’s just probabilistic. AIs are very good at using probabilities to mimic human sentences. But then again, before LLMs, I don’t think everyone believed that language could be probabilistic either.

But as you build AI agents on top of these base probability models, you do get into the question of reasoning. You might sample one of these base models multiple times and get different examples of sentences, and then at some higher level, you might have an AI agent look at those sentences and compare them, and it will try to select the best one. One can think of it as an onion with shells. At the base it is probabilistic, but then you build above that. So then, you very much get into the philosophical question of whether you consider that reasoning. We can wax philosophical if you like, but that is above my pay grade.

Q: Can AI give us a new answer to a problem, such as quantum gravity (how gravity fits with quantum theory)? Could it give us an answer that we haven't thought of before?

I think it's possible, but also, it might come up with an answer we don't fully understand, or that we ourselves don't have full intuition for understanding. An AI involves a lot of microscopic interactions, neural networks interacting with each other, a lot of weights and biases. One of my students pointed out that AI itself harnesses complexity and emergence as a resource. So it is possible that for some unsolved problems in physics there are complex emergent explanations that an AI might be perfectly suited for, but it might not be an intuitive answer that we as humans can relate to.

Q: What are the limitations of using AI for physics? For example, there is the “black box” problem in AI, whereby the AI can spew out an answer, but it may be impossible to know exactly how it got to that answer.

That’s a big question. How do you interpret what goes on in AI? It’s difficult and maybe somehow it is fundamentally intractable. We have all these connections in the human brain that we don’t really understand, and in AI, there are all these little connections that we might not be able to understand either.

But the question is really whether that will hamper our use of AI in physics. That requires a different way of thinking about how we do physics and how we use AI in physics. Just as we can do calculus, but a dog cannot, it is possible that there are some things that artificial intelligence is capable of, or better at, of which we are not as capable.

Q: How will AI and quantum computing feed into each other?

I've thought about this a lot. I think that almost out of necessity, quantum computers will need to be integrated with artificial intelligence in their control stacks because they are the most exquisitely complicated devices we've ever built.

Just as large experiments, like the particle accelerators at CERN, use AI to process data and model particle interactions, I believe quantum computers will need to use AI just to scale up into the useful regime.

But also, quantum computers themselves produce a plethora of interesting data. Going back to how we use LLM-like AI models to predict sequences of qubits, you can imagine training LLMs to be what we might call digital twins of the quantum computer.

So we're in this very interesting cycle that we can call a co-design cycle: AI helps to build quantum computers, and quantum computers produce data that can be used to train artificial intelligence systems. From there, these AI models can predict and characterize the behaviour of the device, and they can basically predict the behaviour of future quantum computers.

It is almost inevitable that this co-design cycle will keep going, and quantum computers and AI will become completely integrated together.

Q: Looking to the future, how will AI change the way physics is done?

It already has. I was recently on an academic visit to Oxford and here, every lunch, physicists talk about AI, how it affects the way we write papers, the way we do calculations, the way we train graduate students. We talk about how we are using ChatGPT, how we are building our own AI models using data from the laboratories. So it has already changed every facet of what we do, and there is no doubt it will change how we do physics in the future.

This is the first of an ongoing series about the Future of Physics. Stay tuned for more!