A study published in the Journal of Cosmology and Astroparticle Physics examines how transfer learning could lower the computing demands of testing cosmological theories beyond the standard model. Researchers first trained a neural network on simpler simulations based on the ΛCDM framework, then adapted it to more complex scenarios involving massive neutrinos or modified gravity. This method cut the need for expensive simulations by over ten times in some tests. However, the work also identified negative transfer, where prior training caused the system to misinterpret new effects that resemble known patterns, such as those linked to the parameter σ8. The findings show both the efficiency gains and the risks of applying pretrained models to fundamental physics research.
