As cosmologists and astrophysicists move deeper into the darkest corners of the universe, their need for more powerful observation and computational tools has grown exponentially. From facilities such as the Dark Energy Spectroscopic Instrument to supercomputers such as the Lawrence Berkeley National Laboratory's Cori system at National Energy Research Scientific Computing (NERSC), they seek to collect, simulate and analyze more and more data that can help explain the nature of the disease Things we can not see, as good as those we can.
To this end, the gravitational lens is one of the most promising tools scientists can use to extract this information by studying both the geometry of the universe and the growth of the cosmic structure. Gravitational lenses distort images of distant galaxies in a way that depends on the amount of matter in the line of sight in a particular direction, and provide a way to view a two-dimensional map of dark matter, according to Deborah Bard. Group leader for the Data Science Engagement Group at Berkeley Lab's National Energy Research Scientific Computing Center (NERSC).
"Gravitational lenses are one of the best methods to study dark matter, which is important because it says a lot about the structure of the universe," she said. "Most of the matter in the Universe is dark matter, which we can not directly see, so we have to use indirect methods to study their distribution."
With the increase in experimental and theoretical datasets, however, the number of simulations is also increasing. To map and analyze this data, a new challenge has arisen: these simulations are increasingly ̵
A multitude of deep generative models are currently under investigation for scientific applications The team led by Berkeley Lab is following a unique path: Generative Adversarial Networks (GANs). In an article published on May 6, 2019 in Computational Astrophysics and Cosmology they discuss their new deep-learning network called CosmoGAN and its ability to create highly accurate, weak convergence maps for gravitational lensing.
A convergence map is virtually a 2-D map of the gravitational lens that we see in the sky along the line of sight, "said Bard, co-author of Article Computational Astrophysics and Cosmology Convergence map have a peak that corresponds to a peak in a large amount of matter along the line of sight, it means that there is a large amount of dark matter in that direction. "
The Benefits of GANs
Why choose GANs instead of other generative models? Performance and precision, according to Mustafa.
"From a deep learning perspective, there are other ways to learn how to make convergence maps from images, but when we started this project, GANs seemed to be more competitive Producing very high resolution images "We were looking for two things: to be accurate and fast," added co-author Zaria Lukic, a co-scientist mputational Cosmology Center in the Berkeley Lab. "GANs offer the hope of being nearly as accurate as complete physics simulations."
The research team is particularly interested in constructing a replacement model that can reduce the computational cost of performing these simulations. In the work Computational Astrophysics and Cosmology they describe a number of advantages of GANs in investigating large physical simulations at the end of training and the images are starting to look good – then network updates can be very chaotic, "said Mustafa. "However, since we have the summary statistics we use in cosmology, we were able to evaluate the GANs at each step of the training, which helped us determine the generator we thought was the best. This technique is typically not used to train GANs. "
Using the CosmoGAN generator network, the team was able to create convergence maps that are described with high statistical reliability using the same summary statistics as the fully simulated maps. This very high level of match between convergence maps, which are statistically indistinguishable from maps created with physics-based generative models, is an important step in building emulators from deep neural networks.
"The tremendous advantage was that we were dealing with the problem was a physical problem that was related to metrics," said Bard. "But with our approach, you can use actual metrics to quantify how accurate your GAN is, which is really exciting to me – how these kinds of physical issues can affect machine learning methods.
Ultimately, such approaches could be Changing the science that is currently based on detailed physical simulations that require billions of computation hours and petabytes of storage space – but there is still much to do Cosmological data (and scientific data in general) may require very high-resolution measurements, such as full-sky telescope images.  "The 2-D images considered for this project are valuable, but the actual physical simulations are 3-D images and can vary over time – and be irregular and create a rich, web-like image structure of features," said Wahid Bhmiji, a big data architect in Data and NERSC's Analytics Services group and co-author of the paper Computational Astrophysics and Cosmology . " The approach needs to be expanded to explore new virtual universes, rather than those that have already been simulated – and ultimately to create a controllable CosmoGAN. "
" The idea of creating controllable GANs is essentially the Holy Grail of the entire problem we are working on: to truly emulate the physical simulators, we need to create replacement models based on controllable GANs, "added Mustafa "At the moment we are trying to understand how training dynamics can be stabilized in the light of the progress made in the field in recent years. Stabilizing training is extremely important to actually do what we want to do next. "
A new filter for better illustration of the dark universe
Mustafa Mustafa et al., CosmoGAN: Creating Low Fidelity High Fidelity Convergence Maps Using Generative Contradictory Networks, Computational Astrophysics and Cosmology (2019). DOI: 10.1186 / s40668-019-0029-9
CosmoGAN: Training a Neural Network to Study Dark Matter (2019, May 16)
retrieved on May 17, 2019
This document is subject to copyright. Apart from any fair dealings for the purposes of private study or research, no
Part may be reproduced without written permission. The content is provided for informational purposes only.