AI Learns Cosmic Rules, Then Struggles to See New Physics
Researchers used transfer learning to train an AI on standard ΛCDM simulations before exposing it to models with new physics, dramatically speeding up cosmological simulations. The study found that while pretraining can reduce computational costs by over tenfold, it can also cause negative transfer, making the AI misinterpret genuine new signals as familiar patterns, especially when neutrino mass effects mimic changes in σ8. The work suggests transfer learning can accelerate future data analysis, but awareness of parameter degeneracies is essential to avoid missing new physics.
