DeepSphere: Efficient spherical convolutional neural network with HEALPix sampling for cosmological applications

Nathanaël Perraudin, Michaël Defferrard, Tomasz Kacprzak, Raphaël Sgier

Introduction

In modern cosmology, data from the Cosmic Microwave Background (CMB) or weak gravitational lensing surveys are typically represented as spherical sky maps. Analyzing these maps with deep learning presents a unique challenge: standard 2D Convolutional Neural Networks (CNNs) require projecting spherical data onto flat grids, which introduces geometric distortions and violates rotation equivariance—the principle that physical laws remain unchanged under rotation.

Examples of cosmological spherical data: CMB temperature map, galaxy count, and simulated weak lensing mass map.

While spectral domain convolutions using spherical harmonics offer a mathematically sound alternative, they are computationally expensive, scaling at $O(N^{3/2})$, and struggle with "partial-sky" maps where data is missing due to galactic interference or observational limits. This paper introduces DeepSphere, a fast, rotationally equivariant CNN designed to work natively on the sphere.

The DeepSphere Architecture

DeepSphere bypasses the limitations of 2D projections and spectral methods by treating the discretized sphere as a graph. By leveraging the HEALPix pixelization scheme—a standard in astrophysics—the model represents each pixel as a node connected to its neighbors.

The DeepSphere architecture: hierarchical graph convolutions and pooling layers operating on HEALPix-sampled spheres.

Key Methodological Components:

  • Graph Convolutions: Instead of traditional sliding windows, DeepSphere uses filters based on Chebyshev polynomials of the graph Laplacian. These filters are radial (isotropic), ensuring the convolution is equivariant to rotation.
  • Hierarchical Pooling: The network exploits the natural hierarchy of HEALPix, where one parent pixel splits into four children, to perform efficient down-sampling.
  • Global Rotation Invariance: To ensure classification remains consistent regardless of the map's orientation, the network uses a Fully Convolutional Network (FCN) approach, averaging local predictions across the sphere.

Experimental Evaluation

The authors tested DeepSphere on a binary classification task using simulated weak lensing convergence mass maps. The goal was to distinguish between two $\Lambda$CDM cosmological models with nearly identical angular power spectra, making them indistinguishable by standard statistical methods. The maps included varying levels of Gaussian noise and were cropped into partial-sky patches to simulate realistic observational constraints.

Results and Insights

DeepSphere demonstrates significant advantages over classical baselines, including standard 2D CNNs and Support Vector Machines (SVMs) trained on power spectra.

  1. Computational Efficiency: DeepSphere's graph convolutions scale linearly ($O(N_{pix})$), making it 10 to 20 times faster than optimized spherical harmonic transforms for filtering tasks.
  2. Superior Performance: The model achieved over 97% accuracy on large and medium partial maps. Even in high-noise, small-fraction scenarios, it maintained ~90% accuracy, significantly outperforming the 80% achieved by the best baseline.
  3. Data Efficiency: By embedding physical symmetries (homogeneity and isotropy) into the architecture via radial filters and global pooling, DeepSphere requires less training data than networks that must learn these properties from scratch.

Conclusion

DeepSphere provides a scalable and robust framework for spherical data analysis in cosmology. By treating the sphere as a graph, it achieves the efficiency of 2D CNNs while preserving the geometric integrity of the data. Future work aims to refine the graph Laplacian to perfectly match continuous operators and extend the framework to regression tasks for inferring continuous cosmological parameters.

Summary updated

Abstract

Convolutional Neural Networks (CNNs) are a cornerstone of the Deep Learning toolbox and have led to many breakthroughs in Artificial Intelligence. These networks have mostly been developed for regular Euclidean domains such as those supporting images, audio, or video. Because of their success, CNN-based methods are becoming increasingly popular in Cosmology. Cosmological data often comes as spherical maps, which make the use of the traditional CNNs more complicated. The commonly used pixelization scheme for spherical maps is the Hierarchical Equal Area isoLatitude Pixelisation (HEALPix). We present a spherical CNN for analysis of full and partial HEALPix maps, which we call DeepSphere. The spherical CNN is constructed by representing the sphere as a graph. Graphs are versatile data structures that can act as a discrete representation of a continuous manifold. Using the graph-based representation, we define many of the standard CNN operations, such as convolution and pooling. With filters restricted to being radial, our convolutions are equivariant to rotation on the sphere, and DeepSphere can be made invariant or equivariant to rotation. This way, DeepSphere is a special case of a graph CNN, tailored to the HEALPix sampling of the sphere. This approach is computationally more efficient than using spherical harmonics to perform convolutions. We demonstrate the method on a classification problem of weak lensing mass maps from two cosmological models and compare the performance of the CNN with that of two baseline classifiers. The results show that the performance of DeepSphere is always superior or equal to both of these baselines. For high noise levels and for data covering only a smaller fraction of the sphere, DeepSphere achieves typically 10% better classification accuracy than those baselines. Finally, we show how learned filters can be visualized to introspect the neural network.
BibTeX
@article{perraudin2019deepsphere,
  author = {Nathanaël Perraudin and Michaël Defferrard and Tomasz Kacprzak and Raphaël Sgier},
  title = {{DeepSphere: Efficient spherical convolutional neural network with HEALPix sampling for cosmological applications}},
  journal = {Astron. Comput.},
  year = {2019},
  doi = {10.1016/j.ascom.2019.03.004},
}
Figure