Synthetic spectra calculated from model solar atmospheres are central to our understanding of the connection between the observed intensity and the solar atmosphere. To obtain synthetic profiles in the general case, we must solve the 3D non-LTE radiative transfer problem for the model atmospheres. The current method used to solve this problem is extremely time intensive and requires access to a supercomputer. We have developed and implemented a convolutional neural network to learn the full 3D mapping between the LTE and non-LTE populations for a model atom. The network is generalized across atom structure and can be adjusted to fit any model atom. In this thesis, the network is built to fit a 6-level model hydrogen atom. Once we have the non-LTE populations, we can easily calculate the intensity they produce at any viewing angle assuming complete redistribution (CRD). The network architecture we propose successfully learns the population mappings for individual atmosphere simulations, allowing us to calculate non-LTE populations for a range of time-steps for a given simulation. Our network is also rather small, allowing it to easily be trained on a standard consumer GPU and resulting in a run time saving of multiple days. A challenge that remains is to make the network general across different simulations.