Non-linear interpolation

We have trained a neural network to find the mapping between the value of the cosmological (Omega_m & sigma_8) and astrophysical (ASN1, ASN2, AAGN1 & AAGN2) parameters and the star-formation rate density (SFRD). We have used 850 IllustrisTNG simulations of the LH set (each simulation has a different cosmology, astrophysics and value of the initial random seed) to train the network. The movie below shows the SFRD from 150 IllustrisTNG simulations never seen by the network (test set) and the predictions of the neural network.