Emulating Sunyaev-Zeldovich Images of Galaxy Clusters using Auto-Encoders

Abstract

We develop a machine learning algorithm that generates high-resolution thermal Sunyaev-Zeldovich (SZ) maps of novel galaxy clusters given only halo mass and mass accretion rate. The algorithm uses a conditional variational autoencoder (CVAE) in the form of a convolutional neural network and is trained with SZ maps generated from the IllustrisTNG simulation. Our method can reproduce the details of the aspherical turbulent galaxy clusters with a resolution similar to hydrodynamic simulations while achieving the computational feasibility of semi-analytical maps, allowing us to generate over 10 5 mock SZ maps in 30 seconds on a laptop. We show that the individual images generated by the model are novel clusters (i.e. not found in the training set) and that the model accurately reproduces the effects of mass and mass accretion rate on the SZ images, such as scatter, asymmetry, and concentration, in addition to modeling merging sub-clusters. This work demonstrates the viability of machine-learning-based methods for producing the number of realistic, high-resolution maps of galaxy clusters necessary to achieve statistical constraints from future SZ surveys.

Predicting the thermal Sunyaev–Zel’dovich field using modular and equivariant set-based neural networks

“…This is difficult to reconcile with our argument that the stochasticity is primarily driven by the unresolved AGN activity, since AGN feedback is more efficient in altering the gas distribution in the shallower gravitational potential wells of lower-mass objects. This may be an indication that the primary driver of stochasticity is in fact the mass accretion rate, as was argued in Rothschild et al (2021). Such an interpretation is further supported by the fact that the merger in the rightmost column of figure 3 seems to be the most challenging scenario for the network.…”

“…The network predictions do generally pick up deviations from spherical symmetry in approximately the right direction. However, the network predicts fields that are noticeably smoother than the target; this is a commonly observed problem in similar tasks (Rothschild et al 2021). The Figure 3.…”

“…Since this is a translationally equivariant spatial problem, the seemingly natural approach chosen for similar problems, e.g. by Tröster et al (2019), Yip et al (2019), Zhang et al (2019), Kasmanoff et al (2020), Thiele et al (2020), Rothschild et al (2021), Wadekar et al (2021), is a convolutional neural net (CNN), taking as input the density field of a gravity-only simulation. However, in this work we argue that existing domain knowledge on P e (⃗ x) and similar fields renders the CNN approach inferior to a set-based architecture.…”

The Importance of Being Interpretable: Toward An Understandable Machine Learning Encoder for Galaxy Cluster Cosmology

“…The fundamental component that differentiates deep autoencoders from other deep methods is an explicit encoding or compression of input data. Autoencoders have been applied in astronomy for a range of tasks, including generating realistic galaxy images (Ravanbakhsh et al 2016), denoising gravitational waves (Shen et al 2017), classifying supernovae (Villar et al 2020), and generating realistic SZ mock images of galaxy clusters (Rothschild et al 2021). A traditional application of an autoencoder can be thought of as a flexible version of principle component analysis, summarizing a complicated signal in a few essential values from which an approximation of the original signal can be reconstructed.…”