Artificial intelligence requires unlearning to discover new physics laws
Artificial intelligence requires unlearning to discover new physics laws
Publish Date: 2026-06-11 04:12:00
Source Domain: www.openaccessgovernment.org
Two images from the Quijote simulations used in this study. The panels show the same region of the Universe, but in different cosmological models. The top image corresponds to the standard ΛCDM model, while the bottom image shows a universe with massive neutrinos and modified gravity. The differences are subtle, but they reveal how changes in the underlying physics can affect the formation and distribution of cosmic structures.
Credit
Francisco Villaescusa-Navarro
A new machine learning study demonstrates that while adaptive AI techniques can drastically lower the computational costs of simulating the universe, they carry a hidden risk
Prior training can cause artificial intelligence to misinterpret signs of new physics beyond the standard cosmological model.
The computational burden of new physics theories
Cosmologists rely heavily on the standard cosmological model, known as Lambda Cold Dark Matter, to describe properties like cosmic expansion and galaxy distribution. However, physicists recognise this model is likely incomplete.
Recent data suggests that phenomena such as massive neutrinos, modified gravity, or changing dark energy could point toward entirely new physical laws.
Testing these alternative theories requires running massive, high-precision computer simulations of virtual universes under varying physical assumptions. Because these simulations demand immense computational processing resources, researchers look for technical shortcuts to streamline the work.
Transfer learning as a processing shortcut
A study published in the Journal of Cosmology and Astroparticle Physics evaluated the use of transfer learning to solve this processing bottleneck. Conducted by researchers from Princeton University and the Flatiron Institute, the project tested whether an AI could reuse knowledge gained from one task to speed up learning in a more complex one.
Instead of training a neural network directly on highly expensive data, the team…
