Researchers present an AI method that turns low-resolution cosmological simulations into higher-resolution particle realizations while taking cosmology into account. The model produces both particle displacements (how far particles move) and velocities, giving a full 6‑dimensional phase space (positions and speeds) conditioned on the input low-resolution fields and the values of five standard cosmological parameters: Ω_m (matter density), Ω_b (baryon density), h (Hubble parameter), n_s (primordial spectrum tilt) and σ_8 (clustering amplitude). The approach is built around a generative adversarial network (GAN), a pair of neural networks that train against each other: one generates candidate high-resolution outputs and the other tries to tell real from generated data, pushing the generator to improve.

To train and test the method the team used the Quijote Latin‑hypercube suite, a set of matched low- and high-resolution N‑body simulations. They trained a single, cosmology-aware model on 100 matched pairs at redshift z = 0. Each training pair maps 512^3 particles to 1024^3 particles inside boxes 1 h^-1 gigaparsec on a side. The model is therefore learning how to add the small-scale detail missing from the low-resolution runs while respecting the overall large-scale evolution encoded in the inputs and the chosen cosmological parameters.

When applied to ten cosmologies that were held out of training, the model reproduced the broad matter distribution and captured how clustering and halo abundances change with cosmology. Quantitatively, over the wave-number range 0.1 < k < 3 h Mpc^-1 the mean of the per-cosmology maximum absolute errors in the matter power spectrum was 3.6%. Counts of halos found by a friends-of-friends (FoF) algorithm matched the high‑resolution reference to within about 15% in the analyzed mass bins below 10^14.8 h^-1 solar masses.