AI Learned Universe Rules. Then Things Got Tricky.
Artificial intelligence can speed up the search for new physics. However, it has a hidden flaw. Sometimes AI knows too much. That becomes a real problem.A new study in JCAP explains why. Researchers used transfer learning to train AI. This method starts with simple simulations. Then it moves to complex ones. As a result, the AI learns faster. For example, it reduced costly simulations by over ten times.
The Shortcut That Works
The team first taught AI using the standard model of the universe. That model is called ΛCDM. After that, they introduced exotic ideas. These include massive neutrinos and modified gravity. “It’s basically a shortcut,” says Adrian Bayer, a co-author. The AI does not need to digest everything at once. Therefore, the process becomes cheaper and faster.So far, so good. But then came a twist.
The Unexpected Trap
The problem is called negative transfer. Imagine a medical student who learns common diseases first. Later, a rare disease looks similar. The student might misdiagnose it. AI does the same thing.For instance, massive neutrinos create effects that look like a known parameter called σ8. Because AI learned universe rules so well, it confuses the two. The AI interprets new physics through old lenses. Consequently, it struggles to recognize truly new phenomena.“The negative transfer is not random,” says Veena Krishnaraj, lead author. It comes from real physical degeneracies. Therefore, scientists must be careful. This does not mean AI is useless. It means we need smarter training methods.In short, AI learned universe rules beautifully. But that very knowledge can become a blind spot. The hunt for new physics continues. And now we know one more hurdle to overcome.
