Discovering new physics may require AI to “unlearn” old physics

Research published in the Journal of Cosmology and Astroparticle Physics (JCAP) investigates how a machine learning strategy known as transfer learning can dramatically reduce the computational costs of exploring new physics beyond standard cosmological models. On the other hand, it also reveals unforeseen risks. That means AI systems can become too dependent on known knowledge.

Artificial intelligence is widely used in cosmology to analyze the universe. However, testing theories that go beyond the standard cosmological model known as ΛCDM remains computationally very demanding.

Although ΛCDM can successfully describe many properties of the universe, from its expansion to the distribution of galaxies, physicists know that it is probably incomplete. Recent observations suggest that phenomena such as massive neutrinos, modified gravity, and evolving dark energy may point to new physics beyond current models.
Testing these alternatives requires running a large number of high-precision simulations of virtual worlds under various physical assumptions, often requiring significant computational resources.

Transfer learning, basically shortcuts

A new paper investigates whether transfer learning, a technique in which an AI system reuses knowledge from one task to accelerate learning for another, could make this process much more efficient.

In this case, the researchers first trained a neural network in a ΛCDM-based simulation (this is known as pre-training) and then adapted it to a more complex cosmological model that included new physics possibilities.

“It’s basically a shortcut,” explains Adrian Beyer, a cosmologist at the Flatiron Institute and Princeton University and co-author of the study. “Typically, people train AI directly on the most computationally expensive simulations. What we do instead is first use simple, low-cost ΛCDM simulations to make the AI ​​aware of what’s going on, and then move on to more complex models.”

This idea is similar to studying a difficult topic by first reading an introductory book. “We start by reading basic books to understand the knowledge, and then we move on to very complex books,” Beyer says.

This strategy avoids forcing the AI ​​to “digest everything at once,” said Princeton University undergraduate Veena Krishnaraj, lead author of the paper.
Results show that this approach works very well. In some cases, transfer learning has reduced the number of expensive simulations required by more than 10 times.

negative transfer

But the study also revealed a more subtle phenomenon known as negative transference.

Returning to the Beyer textbook analogy, this is like studying medicine with an introductory textbook and then encountering a rare disease whose symptoms are similar to a common disease. Prior knowledge is almost always helpful, but it can also mislead the reader.

Something similar can happen with AI systems. The effects produced by new physics can be very similar to patterns already associated with standard cosmological models. In such cases, the AI ​​tends to interpret new information using the categories it learned during pre-training, making it more difficult to recognize truly new effects.

The researchers observed this behavior in simulations involving giant neutrinos. The specific effects caused by the neutrino’s mass are very similar to the variations associated with the existing ΛCDM parameter known as σ8, which describes how strongly matter clusters throughout the universe. As a result, the pretrained network initially had difficulty distinguishing between the two effects.

“Negative transitions are not random; they are caused by the underlying physical degeneracy of the model,” says Krishnaraj. In other words, different physical parameters can produce very similar observable effects, making it difficult for AI to disentangle them correctly. “So this is something that we need to be aware of and try to mitigate,” she concluded.

This study highlights both the potential and the risks of applying a “foundational model” strategy (conceptually similar to the strategy behind modern generative AI and large-scale language models) to fundamental physics. As the authors note in their paper, pre-training can speed up inference, “but it can also impede learning new physics.”

So far, the method has been tested in simulations, laying the groundwork for its application to real observational data. Researchers believe this is a powerful tool for future cosmological investigations, which will generate an unprecedented amount of high-precision data about the universe in the coming years.

The paper “Transfer learning beyond the standard model” by Veena Krishnaraj, Adrian E. Bayer, Christian Kragh Jespersen, and Peter Melchior has been published in JSTAT.