AI Trained on Cosmology Simulations Misses New Physics
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- AI neural network was trained on ΛCDM simulations and learned the standard cosmological model using fewer, less costly simulations.
- The study showed that when the AI was applied to extended models (massive neutrinos, evolving dark energy, modified gravity), it suffered negative transfer, causing biases that masked new physics signals.
- Adrian E. Bayer explained that negative transfer indicates the model reinforces degeneracies between overlapping physical effects rather than failing randomly.
- The research team plans to test the AI on realistic survey data that include galaxy formation uncertainties, survey masks, and noise to assess transfer learning’s usefulness.
- Quijote simulations were used to illustrate the AI’s performance across different cosmological models, showing the same region of the universe under varying parameters.
- DESI data preparation involves creating mock universes and running expensive simulations, highlighting why efficient AI methods could be valuable for cosmology.
Why it matters: Physicists could speed up cosmological analysis by using AI, but the study shows that transfer learning can embed existing biases, causing the model to miss signals of new physics such as massive neutrinos or modified gravity, potentially delaying breakthroughs in understanding the universe.
