Simulations of dwarf galaxies and intensive data analysis resolve a major problem with the reigning theory of how the cosmos evolves

Diminutive in size, with about one-percent the number of stars in the Milky Way, dwarf galaxies have posed a big problem for cosmological theory. Astronomers who look at them — many are known to orbit the Milky Way — know they are “bulgeless,” with a distribution of stars that’s more or less flat on edge, slightly humped like a frisbee. Astrophysicists who run computational simulations that test the reigning “cold dark matter” (CDM) model of how galaxies form, however, have been seeing bulges.

PHOTO: Fabio Governato
Fabio Governato
University of Washington
Photo Credit: Sunny Facer

PHOTO: Tom Quinn
Tom Quinn
University of Washington

A spherical central bulge of stars is common in larger galaxies — the Milky Way has one, but when CDM simulations of dwarf galaxies have them, as they have for about 15 years, it suggests a problem with the model, which has in nearly every other way shown excellent agreement with observations. Similarly, dwarf galaxies from CDM simulations have a centrally concentrated distribution of dark matter, the invisible matter that comprises the largest part of the universe's mass, again inconsistent with observed dwarfs. “Basically we have a model that's really good at explaining a lot of what’s going on in the universe,” says University of Washington astrophysicist Fabio Governato , “and then there’s these two sore points — bulges and dense dark-matter halos in the dwarf galaxies.”

“This failure is potentially catastrophic for the CDM model,” wrote Governato and his University of Washington colleague Tom Quinn and an international team of collaborators in their paper — Nature (January 2010) — reporting simulations that convincingly resolve this dwarf galaxy problem. A big part of the solution was GASOLINE, astrophysics simulation software developed over a 15-year period at PSC by Quinn, James Wadsley of McMaster University, and Joachim Stadel of the University of Zurich. “We made very good use of BigBen [PSC’s Cray XT3 system, decommissioned this year],” says Quinn, “both in developing this code and in earlier simulations.”

By improvements to the way GASOLINE represented some of the physics involved and with higher resolution than had before been possible, the researchers more realistically captured the processes of star formation and evolution, including the violent star death and spectacular gas outflow phenomena of supernovae. Totaling more than a million hours of TeraGrid computing (mainly at NICS and TACC), their simulations showed that not only dark matter, but also stars and gas influence the structure of dwarf galaxies. The result is bulgeless dwarf galaxies — with density properties of dark matter and stars that agree well with observed dwarf galaxies.

“Realistic dwarf galaxies are thus shown to be a natural outcome of galaxy formation in the CDM scenario,” wrote the researchers. Or as Governato says, “CDM lives to fight another day.”

Supernova Outflows

“It was a massive computational project,” says Governato. “This kind of research wasn’t possible just three years ago. We took advantage of the fact that computers are getting faster and faster.”

To keep the amount of computing within limits, the researchers simulated a cubic volume of space 25 megaparsecs on edge (about 475 million-trillion miles). With the typical distance between galaxies being one megaparsec (19 trillion miles), this volume gives a representative sample of galaxy formation. GASOLINE has the ability to adapt resolution according to regions where stars and galaxies form, with a “tree code” — which automatically sprouts more “branches” in active areas of the cube.

Properties of a Simulated Galaxy
These two frames from Governato and colleagues’ simulation of a dwarf galaxy at high resolution show light distribution face-on (left) and edge-on.

To look at the details of galaxy formation, the team focused on some of the sites within the cube where mass congregated. “We picked galaxy formation sites from a lower resolution simulation,” says Quinn, “and then went back and did a higher resolution simulation where we selectively sampled the region where chosen galaxies formed.” In these regions, they simulated with resolution up to 100 parsecs — the most detailed picture of galaxy formation to date.

By implementing star birth and supernova explosions, the simulations produced bulgeless dwarf galaxies while adhering to the physics of the CDM model. “The massive stars explode and form supernovae,” says Governato, “and these giant stellar explosions have a dramatic effect on the distribution of hydrogen and helium in the galaxy. The star explosions produce what we call ‘outflows,’ which blow most of the gas away from the center of galaxies. There’s not much left to form stars, so the bulge doesn’t form. It’s not like you remove it — it doesn’t form in the first place. This is the key result of our simulations.”

Along with bulges, previous models of dwarf galaxies also suffered from what’s called the“cusp” problem — high density of dark matter in the central region. The researchers had no reason to expect their simulation would resolve this, but a colleague analyzing their data found that the dark-matter density profile was flat throughout the central volume of the dwarf galaxies. This was initially a puzzle, since dark matter, which responds only to gravity, is unaffected by “wind” from a supernova explosion, and wouldn’t blow away in the outflow.

The new SGI Altix UV system at PSC will help with large datasets produced by this group and others.

“At that point we started scratching our heads,” says Governato. “This is where it gets fun, where the computer gives you a hint but it doesn’t come with an explanation.” What they eventually realized is that the luminous matter acts as a binding agent for dark matter. “When most of the gas in the central region of the galaxy is blown away, there’s nothing left to hold the dark matter; it expands, and the dark matter density near the center decreases.”

As a result, the simulations solved two problems of the CDM model: the observed lack of bulges and the flat distribution of dark matter in real dwarf galaxies. “Usually this is a good sign for science,” says Governato, “when one result explains more than one thing; a complex system resolves to a simple process.”

Artificial Observations via Shared Memory

To further verify their findings, Quinn, Governato and their colleagues have turned to PSC’s shared-memory system, Pople, for “artificial observations” of the dwarf galaxies formed in their simulations. “After we run the simulations, we want to look at them as an observer would,” says Quinn. “What would this galaxy look like if you looked with the Hubble space telescope at five different wavelengths, five different colors? To do that we have to trace how the starlight propagates through the dust and the gas of the galaxy.”

For this analysis, they use a program called SUNRISE (a Monte Carlo ray-tracing code developed by Patrik Jonsson), which takes the data generated by the simulations and produces images of the light distribution based on the masses and ages of the stars, their chemical compositions, and how they are distributed in the galaxy. The galaxy’s light distribution might make it look blue, or red, or ultraviolet.

Which Galaxy is Real?
A galaxy from Governato and colleagues’ simulation (left) appears in all respects identical to a real galaxy (right) and background image from the Sloan Digital Sky Survey Collaboration.

Image courtesy of Chris Brook (The Jeremiah Horrocks Institute at the University of Central Lancashire) and Patrik Jonsson (Center for Astrophysics, Harvard).

This kind of data analysis is best suited to a shared-memory computing system. “Photons are flying all over the place,” says Quinn. “You can’t easily domain decompose and expect reasonable communications between domains. To handle the photons from a single dwarf galaxy simulation requires about 50 gigabytes of memory, and you really need to have the whole simulation in shared memory.” The improved processor speed and larger shared memory of the new SGI Altix UV system at PSC will help with the larger datasets and artificial observations produced by this group and others.

“This is what’s really new in our field,” says Governato. “You don’t just run the simulation, you want to analyze the data so that they look like data out of a telescope, because when you compare data with observers you want to compare apples with apples. We find that our artificial galaxies look remarkably similar to those observed. They have the same light distribution, they form the same amount of stars, the same light profile without the bulge, and based on how fast the stars and the gas move, they have the same amount of dark matter as the real galaxies. And this confirms the results of our Nature paper.”

“We’ve learned two things,” he continues. “One is that you can create realistic galaxies within what we think is the correct model for structure formation — the CDM model. So it removes a major problem for this model. And second it shows that we’ve started to understand what physical processes shape the distribution of mass at the center of galaxies, namely these large gas outflows caused by supernova explosions.”