Will the next discovery in physics come from AI? How artificial intelligence is also beginning to understand the deepest reality

In 1980, at the University of Cambridge, physicist Stephen Hawking talks during a lecture about the theory of everything, the ultimate overarching law of nature that will link the theory of relativity, quantum physics and all other corners of physics into one fundamental insight. To develop this, he believes, humanity will have to get a little help from machines in the future: ‘The end is not in sight for theoretical physics, but perhaps for theoretical physicists.’

Fast forward 44 years and physicists are starting to come close to that oft-cited vision of the future. Not that the ultimate theory of everything is yet within reach, which is proving to be a bit unruly for the time being, but artificial intelligence is certainly starting to get better at understanding physics.

“I had ChatGPT4 create my physics exams for bachelor students,” says Sascha Caron, researcher in the field of high-energy physics at Radboud University and co-organizer of EuCAIFCon 2024, a major international conference in Amsterdam where physicists last week presented the latest applications of discussed artificial intelligence in physics. ‘GPT now gets an 8 or 9 for those exams.’

Good sparring partner

Caron and many of his colleagues therefore see the chat program as a good sparring partner. “I sometimes ask it to look up formulas for me, or to suggest bits of programming code,” he says. Of course: AI sometimes gives completely reliable-looking answers that, on closer inspection, turn out to be nonsense. “But people also make mistakes,” says Caron. ‘It’s up to me to double check everything. Whether something comes from a master’s student or from an AI.’

At the Atlas experiment, one of the two large detectors at the Large Hadron Collider, the particle accelerator in Geneva, people are working on ‘Chatlas’, an AI that is specifically fed with knowledge about particle physics and databases full of information about the experiment. The chat AI can easily save a researcher a month of research with the answer to a question, the initiators wrote in a presentation they gave in April at the Cern research institute.

Professor of radio astronomy Anna Scaife (University of Manchester), who is also an AI expert at the British Alan Turing Institute, sees a lot of future in AI as a tool for scientists. She herself works with colleagues on self-learning systems that can tame the flow of measurement data that comes from modern, large-scale experiments.

“In astronomy, the volume of data collected is unimaginably large,” says Scaife. For example, the Lofar radio telescope, whose receivers are located in the Netherlands, collects about 50 petabytes per year. “And the Ska telescope in South Africa and Australia will collect 600 petabytes per year when it is completed,” she says, the amount of information on a stack of CDs – without boxes – 1,000 kilometers high, more than twice as high as the orbit. of the International Space Station.

Huge acceleration

Behind all the analysis techniques that Scaife develops are methods that are comparable to well-known AI programs such as ChatGPT and Stable Diffusion. “Language models such as ChatGPT are foundation models, self-learning systems that can, for example, label large amounts of data,” she says. And that is very useful in astronomy.

‘We have thus developed a system that selects ring-shaped galaxies from a collection of millions of photos of the sky. This has produced the first large database of those systems, on which we can now conduct research. A few years ago, something like this was absolutely impossible,” she says – simply because manually sorting through such a volume of photos would have been impossible.

Foundation models are also gaining popularity in particle physics, for example in the analysis of particle collisions. When two particles collide with each other at almost the speed of light in the interior of particle accelerators, many debris – other particles – are released. ‘In the LHC experiments you sometimes have as many as a hundred thousand sensors, causing ten thousand particles to fly around at the same time after a collision,’ says Caron.

Physicists then want to quickly reconstruct all the paths of those particles. ‘These types of models run tens of times faster than all the analysis methods we use. Five years ago this was really unthinkable.’

While the outside world is amazed by increasingly beautiful AI images and videos, scientists such as Caron and Scaife see a parallel, but no less impressive development. ‘This field is accelerating enormously. The number of people in astronomy working with AI, the number of publications about AI in astronomical journals: everything is increasing exponentially,” says Scaife.

Caron even dreams of AI that can do more than make simulations and do data processing. “How we use ChatGPT now, as a research partner, almost as a kind of artificial muse, I also see the cautious rise of the artificial scientist,” he says.

He wants to take the first step by encouraging the physics community to work together on larger, more ambitious AIs. ‘The bottleneck now is mainly time and money. I often work together with a master’s or PhD student and with such a small group it is impossible to develop a physics AI that can be compared to something like GPT4.’

Discovering new particles

Yet even such ‘small’ projects are already achieving impressive results. Physicists are even trying to discover new particles using AI. Since the Higgs boson in 2012 – whose existence was predicted decades earlier by the recently deceased physicist Peter Higgs – particle accelerators such as the LHC have not discovered any new particles. Has humanity found all the basic building blocks of the world around us, or do we not know where to look?

Particle physicists associated with the CMS experiment at the LHC had an AI look completely openly at old measurement data from their experiment in the hunt for an answer. This did not yet produce any new particles, but the analysis itself was promising, as preliminary results showed last March. “This was the first time that an AI looked at measurement data without physicists telling it which particles to look for,” one of the researchers said on Cern’s website.

It raises the question of whether AI will still need people to gather new insights. In 2019, a computer independently discovered that the Earth revolves around the sun. It was the first time that AI started to come close to Hawking’s 44-year-old prediction.

The same year, physicists used a hundred equations taken from the famous Feynman Lectures of Physics, a kind of basic college-level physics course to produce data. They then fed that data to an AI, which promptly managed to reconstruct all hundred equations, as was written in the trade journal at the time Science Advances.

“That was all very rudimentary,” says Caron. After all, it was relatively simple physics, laws of nature that physicists such as Kepler and Newton discovered about four hundred years ago. ‘Not at all comparable to the complexity of what Higgs did, for example,’ says Caron. “But I do think we will get there eventually.”

Developments have been rapid since 2019. Artificial intelligences, among other things, deduced the masses of black holes from measurement data of gravitational waves and described more than a billion voids in the cosmos, areas with virtually no galaxies, which together cover a volume larger than the visible part of the universe.

AIs simulated complex particle collisions and performed fresh calculations within highly specialized areas of physics such as quantum chromodynamics, the theory that describes how quarks – the smallest known building blocks of all matter around us – clump together into larger particles such as protons.

Even the behavior of the as yet unexplained dark matter – stuff that astronomers can only see indirectly because it gravitationally pulls on ‘normal’ matter, but which is otherwise completely invisible – can now be subjected to simulations by artificial intelligences.

Learning new language

It is therefore logical that ChatGPT creator OpenAI is thinking out loud about the next step: developing an artificial brain that is capable of making scientific discoveries independently, says Caron. And that company is not the only one. For example, last year an international team of academic researchers announced the start of Polymathic AI, a ‘multidisciplinary’ program that aims to connect knowledge from different fields, from astronomy to climate science. In this way, the system should be able to understand reality more easily, “similar to the way it is easier to learn a new language when you already know five languages,” the researchers involved said in a press release.

“Scientific AI is learning to process data and run simulations better and better,” says Caron. The next step is to use those simulations to develop new theories. ‘And you can then test those theories in experiments where AI can interpret the results,’ he says. Tie it all together and you no longer need a human to gain new insights into the deepest workings of nature. Although, according to Caron, it is more likely that in the future human physicists will hunt together with AI for the deepest possible level of knowledge about reality.

“Making predictions is always dangerous, so I have to be careful because I’m an optimist on this subject,” he says. ‘But I wouldn’t be surprised if in the next five to ten years we see the first discoveries in physics that are not made by humans, but automated by computers.’ People will always have to verify and interpret such AI findings.

Will the Nobel Prize committee soon only have to decide who will receive the coveted Nobel medal and the associated eternal fame for truly groundbreaking discoveries of artificial brains: the human behind the machine, or the machine itself?