Abstract
In recent years, deep learning approaches have achieved state-of-the-art results in the analysis of point cloud data. In cosmology, galaxy redshift surveys resemble such a permutation invariant collection of positions in space. These surveys have so far mostly been analysed with two-point statistics, such as power spectra and correlation functions. The usage of these summary statistics is best justified on large scales, where the density field is linear and Gaussian. However, in light of the increased precision expected from upcoming surveys, the analysis of -- intrinsically non-Gaussian -- small angular separations represents an appealing avenue to better constrain cosmological parameters. In this work, we aim to improve upon two-point statistics by employing a PointNet-like neural network to regress the values of the cosmological parameters directly from point cloud data. Our implementation of PointNets can analyse inputs of $\mathcal{O}(10^4) - \mathcal{O}(10^5)$ galaxies at a time, which improves upon earlier work for this application by roughly two orders of magnitude. Additionally, we demonstrate the ability to analyse galaxy redshift survey data on the lightcone, as opposed to previously static simulation boxes at a given fixed redshift.
Introduction
Estimating the fundamental parameters of our universe—such as the matter density ($\Omega_m$) and the amplitude of matter fluctuations ($\sigma_8$)—is a central goal of modern cosmology. Traditionally, researchers have relied on "two-point statistics" like power spectra to extract this information from galaxy surveys. While these methods are robust on large scales where matter distribution is relatively simple and Gaussian, they struggle to capture the complex, non-Gaussian structures—filaments, clusters, and voids—that emerge on smaller scales due to gravitational collapse.
This paper presents a novel approach using PointNet, a deep learning architecture specifically designed for 3D point clouds. By treating galaxies as a collection of points rather than a grid or a set of pre-defined summary statistics, the model can automatically learn to extract informative features from the spatial distribution of galaxies, potentially unlocking a wealth of information previously hidden from traditional analysis.
Deep Learning on Point Clouds
The researchers utilized a PointNeXt architecture, an evolution of the original PointNet. Unlike standard convolutional neural networks that require data on a rigid grid (like pixels), PointNet-based models process sets of coordinates $(x, y, z)$ directly.
Key technical advantages of this approach include:
- Permutation Invariance: The network's output is independent of the order in which galaxies are listed, reflecting the physical reality of a survey.
- Hierarchical Learning: PointNeXt partitions galaxies into overlapping local regions, learning small-scale features first and progressively aggregating them to understand global cosmic structure.
- Feature Augmentation: By including halo mass ($M$) as a fourth dimension $(x, y, z, M)$, the network can effectively weight regions of different densities, significantly improving prediction accuracy.
Simulation Framework
To train and validate the model, the study used the COSMOGRIDV1 dataset, a massive suite of dark-matter-only N-body simulations covering 2,500 unique cosmological parameter combinations. A major breakthrough of this work is its scalability; the implementation handles point clouds of $10^4$ to $10^5$ objects—two orders of magnitude larger than previous Graph Neural Network (GNN) approaches. Furthermore, the analysis was conducted on a "lightcone," which accounts for the time-evolution of the universe as we look deeper into space, a critical step toward applying these models to real-world survey data.
Performance and Parameter Constraints
The PointNeXt model consistently outperformed traditional two-point statistics in predicting $\Omega_m$ and $\sigma_8$. For a sample of 8,000 points, the Mean Squared Error (MSE) for PointNeXt was $3.6 \times 10^{-3}$, compared to $8.3 \times 10^{-3}$ for traditional methods. When halo mass was included as an input feature, the MSE dropped further to $1.3 \times 10^{-3}$, representing a nearly six-fold improvement over the baseline.
As shown in the posterior distribution plots, PointNet provides much tighter constraints, particularly for $\sigma_8$. This indicates that the neural network is successfully extracting non-Gaussian information from small-scale clustering that traditional summary statistics miss.
Conclusion and Future Directions
This work demonstrates that hierarchical point cloud networks are a powerful tool for cosmological inference, offering significant improvements in precision and scalability. While the current model handles up to $10^5$ points, future work will aim to reach the $10^6$ scale required for state-of-the-art surveys like DESI. Additionally, incorporating more complex "baryonic" physics and enforcing physical symmetries (like rotational equivariance) directly into the architecture remain important avenues for making these models robust enough for real-world observational data.
Summary updated