Computer simulations help design new nuclear reactors
When it was founded in 1946, Argonne was charged with developing the
technology to enable peacetime uses of nuclear energy. Now, more than
60 years later, Argonne again stands at the forefront of nuclear research
as it brings its new high-performance computing facilities to bear
on reactor design, enabling safer, cheaper and more efficient generation
of electricity.
A fast-neutron nuclear reactor is called “fast” for a reason. The
free neutrons in the reactor chamber travel at average speeds of more
than 5,000 kilometers per second until they slam into uranium or other
heavy nuclei, triggering the fission reaction that creates nuclear
energy. But with an enormous number of neutrons zooming around and
colliding with other particles, scientists have, until now, lacked
the ability to tackle reactor dynamics from the ground up.
“At the fundamental level, we have a very good understanding of the
neutron physics of a nuclear reactor, but we've never been able to
represent all of its details at once,” said Hussein Khalil, director
of Argonne's Nuclear
Engineering Division. “We need to resort to some models or
equations that simplify the situation, but when you do that, you introduce
errors and can't be completely confident that your model represents
reality.”
Computer models of reactors help to bridge the gap between conception
and operation of new nuclear facilities. By using mathematical algorithms
to process information—the
reactor's geometry, composition, measured physical properties and targeted operating
conditions—these models can predict how different reactors will function.
Because scientists cannot forecast exactly how a reactor will respond to variations
in its configuration or the conditions under which it is operated, every nuclear
facility has to be designed and operated with large safety margins, which reduces
the efficiency of the plant. “The traditional tools have large uncertainties,” said
Andrew Siegel, an Argonne computational scientist and physicist who works at
the nexus of the two disciplines. “These early models were coarse, and they represented
only certain fundamental physics with a lot of assumptions and idealizations.”
These
old models were developed during the second heyday of nuclear research in
the late 1970s and early 1980s. Back then, the stud of the supercomputing industry
was the CRAY-1, a five-and-a-half-ton behemoth that could perform 100 million
calculations (known to computer scientists as floating-point operations,
or FLOPS) per second.
That kind of computing power far surpassed anything else available at the
time, and nuclear scientists were among the first to requisition the CRAY -1's
muscle for simulation work. As nuclear research fell out of favor with government
and academic institutions that rationed supercomputing resources, however,
other fields of study—such as astronomy and meteorology—replaced nuclear research
as the preeminent uses of the available hardware. As a result, Siegel said,
nuclear engineers today still have to rely on many of the design tools that
resulted from work done on the CRAY-1 and machines with similar capabilities.
Although
the CRAY -1's 100 megaflops capacity represented a remarkable accomplishment
for the time, the nearly 30 years that have passed have seen the development
of supercomputers that have made the CRAY-1 as scientifically relevant as
an abacus. Last fall, Argonne became the new home of the IBM
Blue Gene® /P
super-computer, which has the potential to perform at speeds of up to one
petaflop—or
one quadrillion calculations per second—representing a ten-million-fold improvement
over the CRAY-1.
Argonne's newly acquired access to petascale-capable hardware, combined
with three decades of accumulated scientific expertise, will revolutionize
how scientists and engineers model nuclear reactors, Siegel said.
“Now, petascale computing allows us to create models that can explicitly represent
a reactor's geometry,” he said. “For the first time, we can resolve a great deal
of the detail of what's happening in a reactor core—it's a true paradigm shift.”
The core of a fast reactor consists of an assembly of stainless steel pins,
which contain thin cylindrical rods of fuel composed of uranium or plutonium
metallic alloys or oxides. Once the reactor gets going, these pins become extremely
hot. In order to convert this heat to electricity and to prevent melting of
the stainless steel, thousands of gallons of coolant—typically sodium— are
washed over the pin bundles.
Altering the geometry of the pins even slightly,
however, affects the neutron chain reaction and changes the dynamics of coolant
flow and heat transfer. Because of the expense and potential safety concerns
of testing different pin configurations experimentally, Siegel and his colleagues
use computer simulations to model how the coolant flows around and between
individual pins.
Reactors convert the uranium and plutonium inside each pin into energy through
nuclear fission, which, when viewed at the atomic level, looks very much like
a game of high-speed billiards. The “cue balls”—free neutrons—zoom around at
mind-boggling speeds until they smack into uranium or other heavy nuclei. Instead
of caroming off in a new direction, however, the nuclei absorb the free neutrons
before splitting into a couple of lighter isotopes and more free neutrons,
releasing energy. The newly freed neutrons then smash into other nuclei, continuing
the chain reaction.
Millions of these collisions happen every microsecond in
a fast-neutron reactor, and even with petascale-computing capabilities,
nuclear scientists could not hope to represent every single one. Older models
dealt with this problem not by analyzing discrete atomic events, but by treating
them as continuous processes; this approach involved many assumptions and
idealizations that simplified the physical realities, according to Khalil.
“What we hope to do with the more powerful computers,” he said, “is actually
begin the simulation at a very fundamental scale by building the model from the
atomic level where the interactions are taking place. Obviously, we can simulate
this degree of detail only for a very small portion of the system, but the hope
is that we can then make use of this information to create a less- detailed—yet
valid—model for the entire system.” This research is funded by the U.S. Department
of Energy's Office of Nuclear
Energy to promote secure, competitive and environmentally
responsible nuclear technologies to serve the present and future energy needs
of the nation and the world.
To carry out this vision, Argonne's scientists
base their neutron physics models on one of two types of mathematical algorithms:
the Monte Carlo method or the deterministic method, according to Argonne
nuclear engineer Won-Sik Yang. The Monte Carlo method, in its simplest form,
involves the use of random mathematical snapshots of neutrons in the reactor,
which are then integrated into a larger model of the system's behavior, a process
somewhat akin to stop-motion animation.
While the deterministic method also attempts to integrate the neutron-nuclide
interactions within a reactor, it does not use random numbers and will always
produce the same output from given inputs. In order to remove the complexities
of measured physical properties in the energy domain, deterministic models
manipulate “bins” of
neutrons that are grouped by energy level. Fast-neutron energies range from
1 electronvolt (eV) to 20 million electronvolts (20 MeV). Current deterministic
models partition neutrons into many energy groups, the behaviors of which are
calculated in several steps with different levels of spatial detail. This technique
dramatically relieves the computational load while still closely describing
the true character of the fission reactions, according to Yang.
“We can accurately determine the integrated qualities of the system with the
Monte Carlo method,” he said, “but it is still exceptionally difficult to get
the local quantities with sufficient accuracy because of the many statistical
fluctuations.”
Because the Monte Carlo method relies on a stochastic approach, the move
to the petascale will only slightly improve its performance in determining
small localized quantities. However, the quantum leap in computing power will
enable deterministic models to process a greater number of narrower energy
bins and improve spatial resolution. “The goal is to eliminate approximation
in proportion to advances in computing,” Yang said.
With better models, scientists like Yang
and Siegel hope to reduce modeling uncertainties and operate reactors just
as safely while improving efficiency and driving down cost. “The new models provide answers in which we can have much more confidence,” Siegel
said. “The solutions produced by the Blue Gene/P can be directly translated
into dollars saved in the operation of new reactors because we no longer have
to put an enormous amount of space between what you thought the answer was
and how wrong you might be.”
Khalil added, “We hope that access to this technology will allow us to design
higher-performance reactors and to operate them closer to their true performance
capabilities.”
In addition to shrinking needlessly large design and operation safety margins,
these new, more flexible models could cut the costs of constructing and operating
a nuclear facility even more by allowing researchers to substitute simulation
for the once-obligatory experiments they performed to validate the results
that the older, highly idealized models produced. “The problem is, when you
do experiments, the parameters that you use are not universal,” Siegel said. “So
if you want to change any facet of your reactor design, you have to ask yourself, ‘will
my code be able to handle that, or do I have to do a new set of experiments?'
With the introduction of advanced computing, for the first time the answer
is likely to be that it can.”
For more information, please contact Dave Baurac (630/252-5584
or media@anl.gov) at Argonne.
|