Simeon Bird

Algorithm helps speed up simulation of vast, complex universes

Anonymous — Thu, 13 May 2021 21:28:52 +0000

Algorithm helps speed up simulation of vast, complex universes Anonymous (not verified) Thu, 05/13/2021 - 14:28 More News Iqbal Pittalwala | Inside UCR May 11, 2021

Simeon Bird, an assistant professor of physics and astronomy at UC Riverside, is a member of a team of astrophysicists that has used machine learning to simulate the universe with high resolution in a thousandth of the time conventional methods would take.

Simeon Bird

The researchers uploaded models of a small region of space at both low and high resolutions into a machine learning algorithm that is trained to upscale the low-resolution models to match the detail of the high-resolution versions. Such training allows the code, which uses “neural networks,” to generate super-resolution simulations containing up to 512 times as many particles as the low-resolution models.

The work, published in the Proceedings of the National Academy of Sciences, was led by Yin Li at the Simons Center in New York and Yueying Ni at Carnegie Mellon University. The research paper is titled “AI-assisted superresolution cosmological simulations.”

Bird, who joined UCR in 2018, studies machine learning, black holes, neutrinos, and dark matter. He said it was a privilege to collaborate on the project. He maintained the simulation code used to generate the training data. In this Q&A, he answers a few questions about the project:

Q. What is a neural network in artificial intelligence and how does it work?

A neural network is a very flexible model to fit any kind of data. You can think of it as a series of filters that show different interesting features of the input. Neural networks are trained to pick out specific interesting parts of the simulation and reproduce them. For our work this is done by training one network which tries to reproduce the simulation and one network which tries to find differences between the reproduction and the original. By playing these two networks against each other we end up with something very hard to distinguish from the original simulation.

Q. How did you get involved in this project?

I run a lot of very large computer simulations. This takes a lot of time. Nowadays it is very hard to make the simulation substantially faster. This type of machine learning offers the possibility of scaling simulations to ten times their current size, which would be very hard in any other way.

Q. What do you think this technology will make possible for astronomy research?

This technology will ultimately enable much larger simulations. These simulations are necessary in the near future to make sure we have models to compare to upcoming much larger astronomical surveys.

Q. How does the algorithm learn how to upscale the low-resolution models to match the detail found in the high-resolution versions?

It is quite difficult to explain this simply. We think it works because on large scales one part of the cosmic web looks very like another: once dark matter starts to collapse it forgets where it comes from.

Q. The code can take full-scale low-resolution models and generate super-resolution simulations containing up to 512 times as many particles. How can you be sure the upscaling isn’t generating unobservable “nonsense”?

For this model we were able to run a simulation directly with 512 times as many particles and check that it looks similar to the output of the upscaling. When we start using this on larger problems, where we can't do that — which is the point! If we can just run the larger simulation, there is no need for machine learning — we will run smaller simulations of parts of the upscaled simulation and check them.

Q. The team couldn’t get the simulation generator to work for two years. What were the obstacles?

Machine learning is a rapidly progressing field. In the last two years there have been advances in how to train these models. We applied these advances and the previously impossible training problem suddenly became possible.

Q. Where else can this new technology be potentially used?

This new technology can hopefully dramatically increase the dynamic range of our simulations, allowing us to model individual galaxies at the same time as the large-scale distributions of galaxies on the sky.

Q. How can someone use this technology?

This technology is freely available, built off open source technology. Anyone with a decently powerful computer and a graphics card could do something similar with enough patience.

^{Thumbnail photo: Milky Way Galaxy photographed by Spitzer Telescope. (NASA/JPL-Caltech/S. Stolovy; Spitzer Science Center/Caltech)}

New Application of Artificial Intelligence Just Removed One of the Biggest Roadblocks in Astrophysics

Anonymous — Tue, 04 May 2021 22:33:01 +0000

New Application of Artificial Intelligence Just Removed One of the Biggest Roadblocks in Astrophysics Anonymous (not verified) Tue, 05/04/2021 - 15:33 More News Thomas Sumner | Simons Foundation May 04, 2021

Simulations of a region of space 100 million light-years square. The leftmost simulation ran at low resolution. Using machine learning, researchers upscaled the low-res model to create a high-resolution simulation (right). That simulation captures the same details as a conventional high-res model (middle) while requiring significantly fewer computational resources. Y. Li et al./Proceedings of the National Academy of Sciences 2021

Using a bit of machine learning magic, astrophysicists can now simulate vast, complex universes in a thousandth of the time it takes with conventional methods. The new approach will help usher in a new era in high-resolution cosmological simulations, its creators report in a study published online May 4 in Proceedings of the National Academy of Sciences.

“At the moment, constraints on computation time usually mean we cannot simulate the universe at both high resolution and large volume,” says study lead author Yin Li, an astrophysicist at the Flatiron Institute in New York City. “With our new technique, it’s possible to have both efficiently. In the future, these AI-based methods will become the norm for certain applications.”

The new method developed by Li and his colleagues feeds a machine learning algorithm with models of a small region of space at both low and high resolutions. The algorithm learns how to upscale the low-res models to match the detail found in the high-res versions. Once trained, the code can take full-scale low-res models and generate ‘super-resolution’ simulations containing up to 512 times as many particles.

The process is akin to taking a blurry photograph and adding the missing details back in, making it sharp and clear.

This upscaling brings significant time savings. For a region in the universe roughly 500 million light-years across containing 134 million particles, existing methods would require 560 hours to churn out a high-res simulation using a single processing core. With the new approach, the researchers need only 36 minutes.

The results were even more dramatic when more particles were added to the simulation. For a universe 1,000 times as large with 134 billion particles, the researchers’ new method took 16 hours on a single graphics processing unit. Existing methods would take so long that they wouldn’t even be worth running without dedicated supercomputing resources, Li says.

Li is a joint research fellow at the Flatiron Institute’s Center for Computational Astrophysics and the Center for Computational Mathematics. He co-authored the study with Yueying Ni, Rupert Croft and Tiziana Di Matteo of Carnegie Mellon University; Simeon Bird of the University of California, Riverside; and Yu Feng of the University of California, Berkeley.

Cosmological simulations are indispensable for astrophysics. Scientists use the simulations to predict how the universe would look in various scenarios, such as if the dark energy pulling the universe apart varied over time. Telescope observations may then confirm whether the simulations’ predictions match reality. Creating testable predictions requires running simulations thousands of times, so faster modeling would be a big boon for the field.

Reducing the time it takes to run cosmological simulations “holds the potential of providing major advances in numerical cosmology and astrophysics,” says Di Matteo. “Cosmological simulations follow the history and fate of the universe, all the way to the formation of all galaxies and their black holes.”

So far, the new simulations only consider dark matter and the force of gravity. While this may seem like an oversimplification, gravity is by far the universe’s dominant force at large scales, and dark matter makes up 85 percent of all the ‘stuff’ in the cosmos. The particles in the simulation aren’t literal dark matter particles but are instead used as trackers to show how bits of dark matter move through the universe.

The team’s code used neural networks to predict how gravity would move dark matter around over time. Such networks ingest training data and run calculations using the information. The results are then compared to the expected outcome. With further training, the networks adapt and become more accurate.

The specific approach used by the researchers, called a generative adversarial network, pits two neural networks against each other. One network takes low-resolution simulations of the universe and uses them to generate high-resolution models. The other network tries to tell those simulations apart from ones made by conventional methods. Over time, both neural networks get better and better until, ultimately, the simulation generator wins out and creates fast simulations that look just like the slow conventional ones.

“We couldn’t get it to work for two years,” Li says, “and suddenly it started working. We got beautiful results that matched what we expected. We even did some blind tests ourselves, and most of us couldn’t tell which one was ‘real’ and which one was ‘fake.’”

Despite only being trained using small areas of space, the neural networks accurately replicated the large-scale structures that only appear in enormous simulations.

The simulations don’t capture everything, though. Because they focus only on dark matter and gravity, smaller-scale phenomena — such as star formation, supernovae and the effects of black holes — are left out. The researchers plan to extend their methods to include the forces responsible for such phenomena, and to run their neural networks ‘on the fly’ alongside conventional simulations to improve accuracy. “We don’t know exactly how to do that yet, but we’re making progress,” Li says.

Information for Press

For more information, please contact Stacey Greenebaum at press@simonsfoundation.org.

Read the original article here:

view article

Evolving the Early Universe in 24 Hours on Frontera

Anonymous — Fri, 15 Jan 2021 16:36:32 +0000

Evolving the Early Universe in 24 Hours on Frontera Anonymous (not verified) Fri, 01/15/2021 - 08:36 More News Simeon Bird January 13, 2021

A 50 Mpc/h square box centered on the largest galactic halo showing in turn the dark matter, temperature, metallicity and neutral hydrogen fraction, and including a zoomed-in image of the host galaxy. [Credit: Yueying Ni, Carnegie Mellon University]

The Texascale Days event in December 2020 provided an opportunity for nine research groups to use large sections of the National Science Foundation-funded Frontera supercomputer at the Texas Advanced Computing Center (TACC) to solve problems that in many cases have never been attempted.

Simeon Bird, professor of physics and astronomy at the University of California, Riverside, used his one-day access to the full Frontera system to run the largest cosmological simulation his team — or any team — have ever performed at this resolution. What follows is a first-person account of the experience.

During the Texascale days event we made progress on our "big run," now dubbed the "Asterix" simulation (named after the cartoon Gaul!). This is the largest cosmological simulation yet performed at this resolution, with 5,500³ dark matter and gas particles, and a plethora of stars and black holes, adding up to almost 400 billion elements.

The calculation was carried out under an Large Resource Allocation (LRAC) grant of supercomputer time to support "Super-resolution cosmological simulations of quasars and black holes." The team includes Tiziana Di Matteo (principal investigator), Yueying Ni and Rupert Croft at Carnegie Mellon University (CMU), Yu Feng at the University of California, Berkeley and Yin Li at the Flatiron Institute.

A histogram of the the number of galaxies in the Universe by stellar mass
at z=6.1. The results achieved by Bird's group demonstrate reasonable agreement
with the real Universe. [Credit: Yueying Ni, Carnegie Mellon University]

The simulation volume — 250 Mpc/h [megaparsec/hubble constant; an alternative measurement to light-years] across — includes for the first time a statistical sample of high redshift supermassive black holes at sufficient resolution to follow their host galaxies. We include a variety of physical models, including black hole growth, star formation, enrichment of intergalactic gas with metals, and gas physics.The black hole modeling in particular was based on a new treatment developed by graduate students Nianyi Chen and Yueying NI (both at CMU), tested just in time for this run. The treatment will allow us to make predictions for black hole mergers across cosmic history that the next-generation, space-based gravitational wave detector space, LISA (Laser Interferometer Space Antenna), will detect.

Our simulation ran from the beginning until redshift 6, a time when the Universe has just become ionized and the era targeted by the upcoming James Webb Space Telescope. This is already a longer run than was achieved in our previous simulation on the Blue Waters machine, which reached z=7 over the course of a year (the previous simulation was lower resolution but with more elements, so it is not fully comparable, but this illustrates that the improvement is an order of magnitude). The largest previous simulation of this class to reach z < 6 contains about an order of magnitude fewer resolution elements.

We have obtained our first science results (right) which shows a histogram of the number of galaxies in the Universe by stellar mass at z=6.1 and demonstrates reasonable agreement with the real Universe.

We used the full Frontera supercomputer for 24 hours, peaking at 7,911 nodes (or more than 400,000 processors) in a single job. This was our first time attempting usage at this scale. I fixed a couple of bugs in the code over the course of the day that only manifest at this scale. We require at least 2,048 nodes to start our job due to the memory requirements of the simulation and we scaled pretty close to linearly up to the full machine.

Our code is efficient, but without access to a machine of this size, it would not matter: memory requirements are driven by the number of resolution elements and cannot easily be reduced. With a machine of Frontera's size, we are able to really push the boundaries of cosmological simulations in a way that is not possible otherwise.

A 3 Mpc/h region around the largest galactic halo in the simulation,
showing gas temperature, gas metallicity, neutral hydrogen fraction and star
formation rate. [Credit: Yueying Ni, Carnegie Mellon University]

We have also used the GPU cluster on Frontera to good effect, and have trained a machine learning-based model on small simulation runs. Yueying Ni and Yin Li have led efforts that resulted in a paper which demonstrated that we could combine high resolution simulations with large volume simulations, approximating a single simulation with a volume one thousand times greater than the individual runs. In the future, we plan to apply this technique to our big run and further increase its dynamic range.

Frontera has really opened up new scientific opportunities. At these high redshifts, the Universe is quite smooth and galaxies are rare. In order to be able to generate predictions for upcoming high redshift telescopes (James Webb Space Telescope and Nancy Grace Roman Space Telescope), we are forced to look at very large volumes and thus very large simulations.

Without the continuing advance of computational capacity this would just not be possible at this resolution and so we would have no way to interpret the upcoming flood of data from new missions.

Read the original article here:

view article