Artificial intelligence could dramatically speed up the search for new laws of physics, but researchers have discovered an unexpected problem: AI trained on current models can become so dependent on that knowledge that it misses entirely new phenomena hiding in plain sight. A study published in the Journal of Cosmology and Astroparticle Physics (JCAP) shows that while transfer learning, a technique that applies knowledge from one task to another, can reduce the computational burden of exploring physics beyond the standard model by more than a factor of ten, it can also create blind spots that prevent AI from recognizing truly novel discoveries.

What Is Transfer Learning and Why Does It Matter for Cosmology?

Transfer learning is a machine learning strategy that allows AI systems to build on knowledge gained from one task to learn another task more efficiently, rather than starting from scratch. For cosmologists, this approach offers a practical solution to an expensive problem. Investigating physics beyond the Lambda-CDM model, the current standard framework that explains cosmic expansion and galaxy distribution, requires generating vast numbers of detailed simulations of virtual universes. Each simulation demands enormous computing power and time.

Researchers at Princeton University and the Flatiron Institute tested whether transfer learning could reduce that computational burden. Their strategy was straightforward: first train a neural network using simulations based on the well-understood Lambda-CDM model, then expose it to more complex cosmological models that include possible new physics.

"It's basically a shortcut. Usually people train the AI directly on the most computationally expensive simulations. What we do instead is first use simpler and less expensive Lambda-CDM simulations to give the AI an idea of what's happening, and only afterward move to the more complex models," explained Adrian Bayer, a cosmologist at the Flatiron Institute and Princeton University and co-author of the study.

Adrian Bayer, Cosmologist at the Flatiron Institute and Princeton University

The results were impressive. In some cases, transfer learning reduced the number of costly simulations required by more than a factor of ten, making it far more feasible to explore new physics theories.

How Can AI's Knowledge Become a Liability?

The study revealed a less obvious challenge called negative transfer, a phenomenon where prior knowledge actually hinders learning. Bayer uses a medical analogy to explain the problem: imagine a medical student learning from introductory materials who later encounters a rare disease that resembles a common illness. Existing knowledge is usually helpful, but it can sometimes lead to the wrong conclusion.

The researchers observed this effect while studying simulations that included massive neutrinos. Some observable consequences of neutrino mass closely resemble changes associated with an existing Lambda-CDM parameter called sigma-8, which measures how strongly matter clusters throughout the universe. Because the two effects can appear so similar, the pretrained neural network initially had trouble telling them apart.

"The negative transfer is not random. It is driven by underlying physical degeneracies in the model. In other words, different physical parameters can create nearly identical observable signatures, making it difficult for the AI to correctly separate them. So this is something we need to be aware of and try to mitigate," noted Veena Krishnaraj, an undergraduate student at Princeton University and the paper's first author.

Veena Krishnaraj, Undergraduate Student at Princeton University

This discovery highlights a fundamental tension in AI research: the same training that makes AI efficient at recognizing familiar patterns can blind it to genuinely novel phenomena. The problem is not random error but rather stems from the underlying physics itself, where different parameters can produce nearly identical observable signatures.

How to Mitigate AI Blind Spots in Physics Research

• Awareness of Physical Degeneracies: Researchers must understand which physical parameters can produce similar observable signatures and account for these degeneracies when designing AI systems.
• Careful Pretraining Design: Rather than relying solely on transfer learning from existing models, scientists should consider hybrid approaches that balance computational efficiency with the ability to recognize novel phenomena.
• Validation Against Real Data: Testing AI systems on actual astronomical observations, not just simulations, can help identify cases where the model has learned patterns that don't reflect genuine physics.
• Iterative Refinement: As next-generation cosmological surveys produce unprecedented volumes of data, researchers should continuously update and refine AI systems to account for new discoveries.

The findings illustrate both the benefits and potential pitfalls of applying foundation model strategies, the conceptually similar techniques used in modern generative AI systems and large language models, to physics research. As the authors note, pretraining can speed up inference but may also hinder learning new physics.

So far, the method has only been tested using simulations. However, the researchers believe it provides an important foundation for future applications involving real astronomical observations. That could become increasingly valuable as next-generation cosmological surveys begin producing unprecedented volumes of high-precision data about the universe. If used carefully, transfer learning could help scientists analyze that information far more efficiently while continuing the search for physics beyond the standard model.

The research, titled "Transfer Learning Beyond the Standard Model," was conducted by Veena Krishnaraj, Adrian E. Bayer, Christian Kragh Jespersen, and Peter Melchior, and is now available in JSTAT. The study underscores a broader lesson for AI in science: efficiency and discovery are not always aligned, and the most powerful tools require the most careful handling.