Pre-trained on cheap FastPM simulations and then refined on a small batch of high fidelity CAMELS simulations, a network needs only 5 to 15% of them to achieve the same accuracy.
To exploit the large surveys of the sky to come, cosmologists rely on sets of very expensive simulations of the Universe, whether they are purely gravitational many-body simulations or complete hydrodynamic simulations. A team on arXiv shows that transfer learning saves a large part of it.
Learning on the "low-end", refining on the high-end
Their model is first trained on fast FastPM-type simulations to estimate two key cosmological parameters, Ωm and σ8, from the power spectrum of matter. It is then refined on a small number of high-fidelity simulations in the CAMELS suite. For a given accuracy, it only needs 5 to 15% of the high-fidelity simulations needed without pre-training, i.e. a total cost in computational hours reduced by up to an order of magnitude, including pre-training.
The gain is even in the face of hydrodynamic simulations: while FastPM completely ignores baryon physics, a pre-training whose cumulative cost remains lower than that of a single hydrodynamic simulation still divides the necessary calculation by about five. The authors see this as an argument in favor of foundation models trained on cosmological simulations.
Read the full article on the original source
Go to the source →