This talk will outline prospects for conducting an unbiased census of relatively nearby dwarf galaxies using convolutional neural networks (CNN) to both identify the galaxies and estimate their distances in Roman Observatory images.

The predicted space density of low-mass dark-matter halos in standard Lambda-CDM cosmology greatly exceeds the density of dwarf galaxies that might have been expected to form in such dark-matter concentrations. From a theoretical standpoint, this “small-scale crisis” for LCDM could signal either a flaw in the dark-matter physics that predicts the power spectrum, or incomplete understanding of the baryonic physics that regulates the conversion of gas into stars within dark-matter halos.

Observationally, this discrepancy is best quantified in groups and clusters; the space-density of gas-poor dwarfs in the field is particularly uncertain. It is hard to detect these low-surface brightness galaxies and hard to obtain redshifts. Roman observations can have a big impact here. For galaxies within ~50 Mpc, WFI images will have a slightly mottled appearance due to stochastic fluctuations in the number of stars per pixel. CNNs offer the prospect using this texture and color information to automate the census.