# Deep Generative Models for Galaxy Image Simulations

**Authors:** Francois Lanusse, Rachel Mandelbaum, Siamak Ravanbakhsh, Chun-Liang Li, Peter Freeman, Barnabas Poczos  
**arXiv:** [2008.03833](https://arxiv.org/abs/2008.03833)  
**Date:** 2020-08-09  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2008.03833v1/

## Key points

- Image simulations are important in astronomy for analyzing wide-field optical surveys
- Existing galaxy models used for simulations are limited and rely on limited data
- A new methodology based on Deep Generative Models (DGMs) is proposed to create complex models of galaxy morphologies
- The hybrid approach combines Deep Learning techniques with physical Bayesian hierarchical modeling
- The generative model is conditional on physical galaxy parameters, allowing for flexibility in generating realistic morphologies
- Training and sampling from the model using galaxy postage stamps from the HST/ACS COSMOS survey demonstrates its effectiveness
- Deep generative models produce more realistic morphologies compared to conventional parametric models
- GalSim-Hub is introduced as a community-driven repository for generative models and a framework for incorporating them within GalSim software
- These generative models offer improved accuracy and broader applicability in future wide-field optical surveys.

## Summary

, , , ,  In the field of astronomy, image simulations play a crucial role in preparing and validating the analysis of wide-field optical surveys. However, the existing galaxy models used for these simulations are often limited to simple parametric light profiles or rely on a limited amount of available space-based data. To address this limitation, a new methodology based on Deep Generative Models (DGMs) is proposed in this study. The aim of this research is to create complex models of galaxy morphologies that can meet the image simulation needs of upcoming surveys. The authors tackle the technical challenges associated with learning morphology models from noisy and Point Spread Function (PSF)-convolved images by developing a hybrid approach that combines Deep Learning techniques with physical Bayesian hierarchical modeling. This hybrid model explicitly takes into account the properties of PSF and noise in observed images. To make the generative model more versatile, it is made conditional on physical galaxy parameters. This allows for sampling new light profiles from specific galaxy populations, providing flexibility in generating realistic galaxy morphologies. The effectiveness of this approach is demonstrated by training and sampling from the model using galaxy postage stamps obtained from the HST/ACS COSMOS survey. To validate the quality of the generated models, a range of second- and higher-order morphology statistics are employed. The results show that these deep generative models produce significantly more realistic morphologies compared to conventional parametric models. To facilitate wider usage within the scientific community, GalSim-Hub is introduced as a community-driven repository for generative models. It also serves as a framework for incorporating these generative models within GalSim, an image simulation software widely used in astronomical research. Overall, this study presents an innovative methodology based on Deep Generative Models for creating complex and realistic galaxy morphologies suitable for image simulations. By addressing technical challenges and incorporating physical constraints, these generative models offer improved accuracy and broader applicability in future wide-field optical surveys.