From Dark Matter to Galaxies with Convolutional Networks (1902.05965v2)

Abstract: Cosmological surveys aim at answering fundamental questions about our Universe, including the nature of dark matter or the reason of unexpected accelerated expansion of the Universe. In order to answer these questions, two important ingredients are needed: 1) data from observations and 2) a theoretical model that allows fast comparison between observation and theory. Most of the cosmological surveys observe galaxies, which are very difficult to model theoretically due to the complicated physics involved in their formation and evolution; modeling realistic galaxies over cosmological volumes requires running computationally expensive hydrodynamic simulations that can cost millions of CPU hours. In this paper, we propose to use deep learning to establish a mapping between the 3D galaxy distribution in hydrodynamic simulations and its underlying dark matter distribution. One of the major challenges in this pursuit is the very high sparsity in the predicted galaxy distribution. To this end, we develop a two-phase convolutional neural network architecture to generate fast galaxy catalogues, and compare our results against a standard cosmological technique. We find that our proposed approach either outperforms or is competitive with traditional cosmological techniques. Compared to the common methods used in cosmology, our approach also provides a nice trade-off between time-consumption (comparable to fastest benchmark in the literature) and the quality and accuracy of the predicted simulation. In combination with current and upcoming data from cosmological observations, our method has the potential to answer fundamental questions about our Universe with the highest accuracy.

• The paper introduces a novel two-phase CNN architecture that accurately maps dark matter distributions to galaxy formations.
• It employs binary classification to detect galaxies and regression to estimate counts, significantly reducing computational costs compared to traditional methods.
• Quantitative evaluations using power spectrum and bispectrum metrics confirm the model’s superior performance in simulating both large- and small-scale structures.

Overview of "From Dark Matter to Galaxies with Convolutional Networks"

This paper presents a novel approach to simulating the universe's galaxy distribution using deep learning methodologies. Traditional methods to predict such distributions rely heavily on computationally intensive simulations that are both expensive and time-consuming. The authors introduce a two-phase convolutional neural network (CNN) architecture designed to offer accurate predictions of galaxy distributions by learning the mapping from dark matter distributions obtained from N-body simulations.

Key Contributions and Results

The paper addresses a significant challenge in cosmology: predicting galaxy distributions efficiently and accurately. Existing methods, like state-of-the-art hydrodynamic simulations, require enormous computational resources, often consuming millions of CPU hours. To tackle this, the authors explore a machine-learning-based solution using CNNs to relate the 3D dark matter field to galaxy distributions.

The proposed solution involves a two-phase architecture:

1. Binary Classification Phase: This phase employs a classifier to predict the presence or absence of galaxies in a given voxel. The high sparsity of the galaxy data makes this an essential step to guide the subsequent phase.
2. Regression Phase: This phase estimates the number of galaxies in selected voxels, refined by the initial binary classification.

The network architecture incorporates elements from popular models like U-Net and Inception networks. The authors report that using a combination of an Inception module and an R2U-Net variant achieved the optimum balance between accuracy and computational efficiency.

Quantitatively, the method outperforms the Halo Occupation Distribution (HOD) benchmark, particularly in reproducing galaxy distributions on both large and small scales. Metrics such as the power spectrum and bispectrum confirm superior performance, with the proposed CNN-based method capturing non-linear galaxy formation effects better than traditional approaches.

Implications and Future Directions

This model carries significant implications for computational cosmology, offering a promising alternative to traditional, resource-heavy simulation techniques. The ability to quickly generate realistic galaxy distributions from dark matter data can facilitate more extensive and detailed cosmological studies.

Practically, the method could be applied to populate dark matter halos with galaxies in large-scale simulations quickly. Theoretical implications are equally substantial, providing new opportunities to investigate complex astrophysical processes' influence on galaxy formation and evolution. Future research directions include extending the model to predict additional galaxy properties (e.g., star formation rates, metallicity) and exploring varied cosmological conditions to improve model generalization.

The authors suggest that training the model on observations spanning different epochs of the universe could refine predictions at different cosmological scales. Furthermore, integrating this methodology with upcoming cosmological datasets from advanced observatories could profoundly enhance our understanding of universe structure and composition.

In conclusion, the paper presents a significant step forward in cosmological simulations, demonstrating the power and flexibility of deep learning in computational sciences. While the results are promising, ongoing efforts to refine model architecture and expand applications will be critical to realizing the full potential of machine learning in cosmology.