Princeton University Thunch Talk

In the era of JWST, observations are yielding more detailed information about the distribution and structure of star-formation regions in distant galaxies than ever before. To interpret these observations, galaxy simulations need to use increasingly sophisticated sub-grid models of star formation and feedback. One feedback mode is direct heating of gas by incident photons. Gas often needs to cool radiatively to reach the low temperatures and high densities required for star formation. Consequently, a key component of galaxy simulations is the gas cooling and heating functions, which control gas thermodynamics. These functions depend on both gas properties (temperature, density, metallicity) and the incident radiation field, which generally includes both a spatially constant extragalactic background and contributions from local light sources. Although cooling and heating functions can be computed exactly with photoionization codes, that is computationally expensive and impractical to do on-the-fly in hydrodynamic simulations. We train machine learning models to predict cooling and heating functions calculated with the photoionization code Cloudy at fixed metallicity. We find that models trained with as few as 3 averaged radiation field intensities can outperform a traditional interpolation approach at each fixed metallicity.