Cosmology is the scientific discipline that tries to understand the physics of the universe as a global object. To achieve this, we are facing the challenges of requiring to handle complex dataset, with badly characterized systematic effects, the impossibility of repeating the experiment and having an unclear physical model. Within that context, we have attempted to build a Bayesian Hierarchical Model that is capable of overcoming a lot of these issues. One of the step in the model involves the capability to run fast forward simulation of classical objects found in cosmology, which galaxies are supposed to inhabit, dark matter halos. I will present and discuss new methods based on Deep Learning that we have developed to achieve this goal. I will show the architecture, some of the advantages, and the current drawbacks of the strategy, alongside the design choices that were made.