Using Graph Neural Networks, CMU Researchers Trained Generative Adversarial Networks to Correctly Predict the Coherent Orientations of Galaxies in a State-of-the-Art Cosmological Simulation

Physics has a branch called cosmology that examines the entire cosmos. It seeks to answer fundamental questions about nature by investigating various things like the matter/energy content of the cosmos, the evolution of the cosmos, etc.  

Future wide-field cosmological surveys, like the High Latitude Survey (HLS) of the Roman Space Telescope, will be able to measure cosmological parameters with unprecedented precision, including dark energy. These surveys will use techniques like weak gravitational lensing. 

The appearance of adjacent galaxy images is impacted by the cosmic matter distribution’s coherent spacetime distortion, which is how weak lensing is detected by searching for coherent patterns in galaxy forms. Intrinsic alignments (IA) create coherent form distortions that anti-correlate with lensing shape distortions. If IA is not considered, the cosmological analysis will be biased. This is one of the main astrophysical contaminants when quantifying this signal. 

High-resolution hydrodynamical simulations are useful tools for studying these alignments and can simulate the birth and evolution of individual galaxies. But they can only examine small cosmic volumes due to their high processing costs.

Many studies have adopted artificial intelligence (AI) advancements in astrophysical and cosmological frameworks. A class of unsupervised deep learning techniques known as deep generative models (DGM) is gaining popularity in many industries. The DGMs can be used to create new sample data after being trained to determine the likelihood of a given dataset.

To paint accurate galaxy alignments in considerably bigger N-body simulations at a low cost, researchers at Carnegie Mellon University create a deep generative model of 2D and 3D galaxy orientations that can capture the right correlations of galaxy alignments. The research was carried out in collaboration with McGill University, Mila, Quebec AI Institute, AIM, CEA, CNRS, and Epistemix Inc.

To capture the pertinent scalar and vector properties for IA, they trained a DGM using a cutting-edge simulation (the TNG100 hydrodynamical simulation from the IllustrisTNG simulation suite) with galaxy formation/evolution.

The researchers highlight the importance of accurate and exact models of survey data to derive robust cosmological information from weak lensing studies. With no fixed pattern or regular geometry, galaxies in the universe are sparsely dispersed throughout space, making them incompatible with traditional approaches representing data as vectors, grids (tensors), or sequences (ordered sets). As a result, they believe that graphs are best suited to naturally model galaxies in the universe.

They explain that the basic notion in their paper is that they have a list of characteristics that are important for capturing the reliance of intrinsic alignments within a halo. Further, tidal fields are important for capturing the dependence of IA for galaxies on the matter outside their halo. The GAN-Generator is then fed these inputs and tries to learn the desired output labels.

To evaluate the performance of the GAN-Generator, the input and output from the GAN-Generator are fed into the GAN-Critic (blue box). The distributions of their actual measured counterparts from the simulation were in good quantitative agreement with the GAN model’s values for both scalar and vector quantities.

The team is also investigating how the model works with low-resolution big volume gravity alone simulations. In the future, they aim to develop a comparable neural network with SO(3) equivariance.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'Galaxies and Halos on Graph Neural Networks: Deep Generative Modeling Scalar and Vector Quantities for Intrinsic Alignment'. All Credit For This Research Goes To Researchers on This Project. Check out the paper and reference article.

Please Don't Forget To Join Our ML Subreddit