A new study shows that transfer learning can reduce the computational costs of searching for physics beyond the standard cosmological model by more than a factor of ten. The approach trains AI first on simpler simulations before moving to complex ones. However, it can lead to negative transfer that hinders detection of genuinely novel effects.
Researchers at Princeton University and the Flatiron Institute tested transfer learning in cosmology simulations. They pretrained neural networks on standard ΛCDM models before applying them to scenarios involving massive neutrinos or modified gravity.
Adrian Bayer, a cosmologist at the Flatiron Institute and Princeton University, described the method as a shortcut that avoids training directly on the most expensive simulations. Veena Krishnaraj, the study's first author and a Princeton undergraduate, noted that the strategy prevents the AI from digesting everything at once.
The team identified cases of negative transfer where pretrained models struggled to distinguish new physics signals from familiar patterns, such as those linked to the σ8 parameter. Krishnaraj said this issue stems from underlying physical degeneracies and requires mitigation.
The findings, published in the Journal of Cosmology and Astroparticle Physics, suggest transfer learning could support future surveys but has so far been tested only in simulations.