Helen and Francisco (Paco) tell us about their recent work using neural networks to predict the masses of subhalos within simulations. They find that the neural network trained on a subset of the subhalos is very good at predicting subhalo masses for the rest of the data.

Restricting themselves to just three parameters, the radius, the velocity dispersion and “Vmax” (the maximum circular velocity of the subhalo) the neural network does almost as well. With this as motivation they look for analytic models that might capture what the neural network is seeing. they find a simple three parameter power-law does OK using just radius and velocity dispersion as inputs – but a running power-law with all three (radius, velocity dispersion and Vmax) does extremely well (almost as well as the neural network).

Interestingly, although the neural network starts to perform poorly when extrapolated to masses beyond its training set, the analytical model manages to still perform well. This suggests that they might be on to some genuine new physical insight about how these parameters combine to determine a subhalo mass.

Various subtleties suggest that this new insight might be related to the virial theorem, but more exploration would be needed to be absolutely sure how.

The work here is part of the CAMELS project, which has thousands of N-body and hydrodynamic simulations of the universe from which to train neural networks on.

Helen: https://www.linkedin.com/in/helen-shao-794641169
Paco: https://franciscovillaescusa.github.io/
CAMELS: https://www.camel-simulations.org/

The paper: https://arxiv.org/abs/2109.04484