Can AI Discover New Physics?

The impact of AI on Physics will be considered during the final Panel Discussion and one of the #LINO24 Next Gen Science Sessions. The text should get you in the mood for this topic to follow the discussions in Lindau or later in the Mediatheque.

Artificial intelligence (AI) has permeated every facet of our lives. Even basic tools like docs and spreadsheet applications now come with AI ‘helpers’ that can generate content for you at the tap of a ‘magic wand’. However, scientists – who are always on the cutting edge of technology – have been using machine learning tools for years before Silicon Valley embarked on its AI-first push to bring such tools to the masses.

Physicists in particular have made good use of neural networks to advance their work. For instance, this technology is helping researchers reconstruct particle trajectories in accelerators, find exoplanets many light-years away, and detect whispery gravitational waves. In some cases, AI has accelerated research in physics by years – and we’re just getting started, honestly.

Hold your horses, though. While results so far have been generally great, they haven’t been exactly revolutionary. But what if AI wasn’t just a hyper-productive assistant wearing many hats and occupied the driver’s seat instead? In other words, could AI someday become a theoretical physicist and discover new physics?

The Sum of All Human Knowledge

It’s a fascinating question to ponder. Imagine feeding the sum of all human knowledge to a next-generation AI and asking it to find the bridge between General Relativity and Quantum Field Theory. Or perhaps you can ask it to unravel the nature of dark energy – or at least confirm if it even exists. It’s a seductive scenario for sure, one that may be only a decade away if you ask technology pundits or proponents of artificial general intelligence, (AGI).

I’m not as optimistic as these pundits. These ambitious projections often remind me of how people in the mid-20^th century envisioned the future. When we lived through the introduction of aircraft, combustion engines, the radio and telephone, and the first satellites, the tendency was to extrapolate on the existing curve. But things change, and the world we live in is very different from what most people envisioned in the 20^th century.

Any AI, as they’re designed today, is extremely unlikely to achieve the kind of breakthroughs required to come up with new physics. That would require genuine intelligence and creativity, whereas present-day AI and machine learning technologies operate primarily as optimization tools. They are fed data along with correct responses, and their task is to minimize errors across the dataset. Specifically, neural networks can be understood as advanced forms of curve fitting. Without their activation functions, they essentially become multivariate linear regressions, akin to least squares regression commonly used across various scientific disciplines.

My point is that all AIs so far are pattern recognition machines on steroids. They are computational tools – very powerful tools – but they lack real intelligence, lest of all the kind of genius required to make cognitive leaps that propel physics forward. Those who look at the pace of AI development today expecting it to progress into super-intelligence in no time may be disappointed to find these machines are a one-trick pony.

That being said, such breakthroughs may be possible with another architecture. And although professors’ tenures aren’t threatened any time soon, there are ripples in the field that suggest a machine capable of human-like creativity that may push the envelope of physics is indeed within the realm of possibility. You just need to look closely enough.

AI As an Inventor

One exciting development that was brought to my attention came from a rather unlikely place: the courtroom. In 2023, the U.S. Senate held a hearing on AI and patents. The debate was about whether an AI could be listed as an inventor or co-inventor. While the hearing ultimately settled that only humans can be granted patent rights, at least under current standards, one of the called expert witnesses made an intriguing case.

Dr. Ryan Abbott, a professor at the University of Surrey School of Law in England, who founded the Artificial Inventor Project, a group of intellectual property lawyers and an A.I. scientist, came to hearing with an oddly-shaped drinking container. Abbott proudly displayed the container and passed it around to the audience while explaining how it was created from scratch by an AI system trained on general knowledge. By that he meant that the computer program was never taught specifically anything about container design – it wasn’t even asked to make a container.

The AI’s task was to combine simple ideas and concepts into more complex ones and identify which idea had the greatest chance of making a positive impact. Among the thousands of linear combinations, it proposed a container design that employs fractal geometry to improve heat transfer, the reverse quality of a Thermos. Such a container is useful if you need to quickly cool something down.

The container’s shape with its peculiar pits and bulges makes it look like an alien product. It’s hard for me to imagine how a human could come up with something like this. While it’s difficult to drink from it and perhaps not very practical, the container is certainly novel and purely the creation of an AI without any human control involved.

This system, known as DABUS, for Device for the Autonomous Bootstrapping of Unified Sentience, was made by Stephen Thaler, an AI researcher with decades of experience. Intriguingly, he describes this software as a system that has the machine equivalent of feelings. When it recognizes a useful idea, simulated neurotransmitters excite digital neurons which sets off “a ripening process, and the most salient ideas survive,” he told the NY Times. He even goes as far as describing the way DABUS reacts to stimuli as sentience.

Neural Networks To Simulate Human Emotions

If coming up with new physics is a lot like inventing, you can see how these scientists are on to something. This suggests that perhaps the key to unlocking AGI is using neural networks that simulate human emotions. It’s fascinating – but I’ll leave it to you whether that’s a good idea.

Until then, many physicists are happy to embrace artificial intelligence as a muse. For instance, physicist Mario Krenn from the Max Planck Institute for the Science of Light and colleagues are developing AI algorithms meant to guide them toward new ideas in physics.

When Krenn was struggling to design an experiment that could let him observe a specific type of quantum entanglement, he turned to this AI, known as MELVIN, to design the quantum experiment for him. MELVIN was trained on all known quantum experiments, which it mixed and matched to find solutions to new problems. Guess what: It worked. Using the framework provided by the AI, the physicists built the experimental setup and observed the phenomenon for the first time.

“I let the algorithm run, and within a few hours it found exactly the solution that we as human scientists couldn’t find for many weeks,” he said.

In another situation, MELVIN was able to suggest a long-forgotten technique devised in the 1990s in order to generate highly complex entangled states involving multiple photons. “When we understood what was going on, we were immediately able to generalize [the solution],” Krenn, who is now at the University of Toronto, told Scientific American.

If these developments are any hint, they point towards a future where AI systems could one day autonomously act as scientists in their own right – coming up with a working hypothesis, designing experimental setups, analyzing the data, and providing a conclusion. All that remains now is breaking the huge brick wall that stands between today’s technology and what we aspire to.