In the hunt for physics beyond our current understanding, researchers at the Flatiron Institute and Princeton University have discovered that teaching AI systems to forget might be just as important as teaching them to learn.
A new study published in the Journal of Cosmology and Astroparticle Physics reveals how transfer learning—a machine-learning shortcut that lets AI reuse knowledge from one task to master another—can slash the computational cost of testing revolutionary cosmological theories by more than a factor of 10. The discovery matters because physicists suspect our standard model of the universe, called ΛCDM, is incomplete. Recent observations hint that massive neutrinos, modified gravity, or evolving dark energy might point toward new physics, but testing these alternatives requires running countless high-precision simulations of virtual universes—a process that demands staggering computing power.
The researchers' solution borrows from how humans learn difficult subjects. Rather than forcing an AI to tackle the most computationally expensive simulations all at once, they first trained neural networks on simpler ΛCDM simulations, then adapted them to more complex models incorporating possible new physics. "It's basically a shortcut," explains Adrian Bayer, a cosmologist at the Flatiron Institute and Princeton University. "Usually, people train the AI directly on the most computationally expensive simulations. What we do instead is first use simpler and less expensive ΛCDM simulations to give the AI an idea of what's happening, and only afterward move to the more complex models." The approach proved remarkably effective, sometimes reducing the number of expensive simulations needed by more than a factor of 10.
But the research uncovered an unexpected trap. Veena Krishnaraj, an undergraduate student at Princeton and first author of the paper, and her colleagues discovered what physicists call "negative transfer"—when prior knowledge actually hinders discovery. The AI, having learned to recognize patterns in standard cosmology, sometimes struggled to spot genuinely new effects that superficially resembled old ones.
This happened strikingly with simulations involving massive neutrinos. Certain effects produced by neutrino mass closely resembled variations associated with σ8, a known ΛCDM parameter that describes how strongly matter clusters across the universe. The pretrained network initially confused the two, interpreting new physics through the lens of familiar categories. "The negative transfer is not random. It is driven by underlying physical degeneracies in the model," Krishnaraj notes. In other words, different physical parameters can produce observationally similar effects, making it genuinely difficult for AI to disentangle them.
The finding mirrors a timeless human struggle: sometimes what we already know gets in the way of what we're trying to discover. Yet it also points toward solutions. By understanding when and why negative transfer occurs, physicists can design better strategies to help AI systems overcome their preconceptions.
The work highlights both the promise and peril of applying "foundation model" strategies—similar to those behind modern generative AI and large language models—to fundamental physics. Transfer learning can accelerate discovery, but as the authors write, pretraining "may also hinder learning new physics." For now, the method has proven powerful on simulations. As researchers continue refining these techniques, they're learning that sometimes breaking free from what you know is the key to understanding what you don't.