GANSky -- fast curved sky weak lensing simulations using Generative Adversarial Networks

TL;DR

GANSky introduces a fast, accurate, and interpretable GAN-based method to generate full-sky weak lensing maps from lognormal inputs, enabling efficient non-Gaussian statistical analysis for cosmology.

Contribution

This work presents GANSky, a novel GAN approach that produces high-fidelity weak lensing maps with minimal network complexity, improving simulation speed and accuracy over traditional methods.

Findings

GANSky accurately reproduces key statistical measures of weak lensing maps.

The method requires only about 1,000 network parameters.

GANSky effectively captures non-Gaussian features for cosmological analysis.

Abstract

Extracting non-Gaussian information from the next generation weak lensing surveys will require fast and accurate full-sky simulations. This is difficult to achieve in practice with existing simulation methods: ray-traced $N$-body simulations are computationally expensive, and approximate simulation methods (such as lognormal mocks) are not accurate enough. Here, we present GANSky, an interpretable machine learning method that uses Generative Adversarial Networks (GANs) to produce fast and accurate full-sky tomographic weak lensing maps. The input to our GAN are lognormal maps that approximately describe the late-time convergence field of the Universe. Starting from these lognormal maps, we use GANs to learn how to locally redistribute mass to achieve simulation-quality maps. This can be achieved using remarkably small networks ($\approx 10^3$ parameters). We validate the GAN maps by computing a number of summary statistics in both simulated and GANSky maps. We show that GANSky maps correctly reproduce both the mean and $\chi^2$ distribution for several statistics, specifically: the 2-pt function, 1-pt PDF, peak and void counts, and the equilateral, folded and squeezed bispectra. These successes makes GANSky an attractive tool to compute the covariances of these statistics. In addition to being useful for rapidly generating large ensembles of artificial data sets, our method can be used to extract non-Gaussian information from weak lensing data with field-level or simulation-based inference.