Machine Learned Emulator of  the Extragalactic Background Light

I worked at NASA Jet Propulsion Lab creating a way to speed up models of the Extragalactic Background Light (EBL). The EBL is light between resolved objects like galaxies and stars. The EBL Power Spectrum reveals this information by tracking features in the EBL as a function of the angular separation between them. Unresolved galaxies from the Epoch of Reionization, diffuse halos of stars (Intrahalo Light or IHL) and satellite galaxies around galaxies, and the diffuse galactic light (DGL) of the Milky Way all contribute to the shape of EBL power spectrum. 

NASA JPL will send up a satellite named SPHEREx, which maps the sky in  6x6 arcsecond pixel infrared spectra.  This information will be used to create power spectra at 10 wavelength bands, effectively creating power spectra images with one axis as separation distance, another as wavelength, and the pixel value as the power spectra

To understand the cosmology of the universe from these power spectra, we will need to run millions of models. These models take 5-10 minutes each, making these runs computationally intensive. 

Emulators are a method of fitting a cheap function to a much more difficult run to speed up run time. I fit an emulator to a few thousand iterations of the model showing that we are able to speed up our models with an emulator while still gaining accurate results. (from ~10 minute to 7.4 millisecond) The emulator is a deconvolutional neural network which takes in a list of parameters and builds features of an image based on them. To get an uncertainty at each pixel and prevent overfitting we used a method called Monte Carlo Dropout.  (paper coming)