Nuclear predictiveSeptember 20, 2010
Imagine a new-generation nuclear plant – inherently safer, more efficient and able to refine petroleum or manufacture plastics while producing energy.
Such power plants would reduce America’s reliance not just on foreign oil but also on fossil fuels in general; while heating and lighting millions of homes, they would simultaneously replace some of the most energy-intensive processes in American manufacturing.
That’s the potential of technology that would improve the United States’ water-cooled reactors. The technology is promising as America moves toward more advanced gas-cooled reactors.
One hurdle is to win the hearts of Americans who remember the Three Mile Island accident and Chernobyl nightmare.
But another huge hurdle is cost – how to justify investing billions of dollars in constructing a new-generation plant with no guarantee that the BTUs produced would make that investment a sound one in the long run.
The construction costs of a nuclear power plant are enormous, but so are the costs of research – the painstaking hours, months and years invested in analyzing the interactions of neutronics, fluid mechanics, and structural mechanics in order to predict the behavior of the reactor throughout its lifetime.
Potential investors in next-generation reactors and the U.S. Department of Energy are counting on the synergistic efforts of reactor designers, computational scientists and applied mathematicians to find ways to analyze reactor flow through simulation – capitalizing on the power of high-performance computers.
One of the people leading the way is Paul Fischer, an applied mathematician and mechanical engineer who works in the Mathematics and Computer Science Division at DOE’s Argonne National Laboratory near Chicago.
Fischer uses millions of hours of computer processing time on the IBM Blue Gene/P and a unique code to make detailed simulations of coolant flow in a reactor. His modeled device is about the size of a long, narrow mailbox, packed with 217 fuel pins and about 1 billion data points.
Any one of a number of properties can be calculated at each data point – temperature, pressure, turbulence and velocity.
It takes 65,000 processors working eight hours a day for 16 days, crunching numbers and information, to understand what the entire mailbox-sized device is experiencing at those pressures and temperatures.
And when Argonne gets its next-generation IBM Blue Gene computer, expected to be among the world’s fastest, Fischer’s nuclear reactor flow simulation will be one of the first applications to run on it.
Fischer’s large-eddy simulations of flow in sodium-cooled reactors are 100 times more ambitious than any done before. The work is designed to run at petascale speeds – more than 1 quadrillion floating-point operations a second – and to provide detailed information about heat transfer within the core.
The aim is to demonstrate that the temperature inside a helium-cooled or sodium-cooled reactor can be reliably predicted. That’s crucial, because if plant operators have confidence in the precise temperature, they can run nuclear reactors at higher power levels without compromising safety – resulting in reduced energy costs.
In Fischer’s simulation, coolant passes through interior flow channels between the 217 pins – each of which has a single wire spiraling around it – and through corner channels between the pins and the walls.
By exploiting symmetries, and by virtue of the relatively short entrance length for the flows, Fischer can simplify the calculations so that only a single wire pitch must be simulated.
An aging fleet
America’s operating nuclear power plants were built in the 1960s and 1970s, but they are using technology that is even older – circa 1940s and 1950s.
“These plants were constructed when we didn’t even have desktop computers,” says Tim Tautges, a computer scientist and head of Argonne’s SHARP project, which is developing high-accuracy simulation tools for reactors. “The basis for their designs was largely experimental – they performed a great deal of experiments to characterize the behavior of nuclear reactors.”
Innovative at the time, the experiments nonetheless could only crudely approximate what would happen with temperature, pressure and vibrations inside the core. The scientists therefore built in a large safety margin, a fudge factor that cut the risk of meltdowns by lowering the maximum temperatures for reactor operation. That fudge factor greatly compromised the efficiency of the reactors. The fact they weren’t burning anywhere near the safest peak temperatures also meant they weren’t producing as much energy as was feasible.
Today’s scientists know a lot more about the physics of nuclear reactors and about the physics of what happens to materials inside them. “We’ve demonstrated with nuclear weapons that we can certify and characterize their behavior using high-performance parallel computers,” Tautges says. “It’s a natural progression to apply that same technology to nuclear reactors.”
Two realities haven’t changed: safety and cost.
Utility companies don’t want to build a reactor that has been tested for safety using only the coarse experiments of the past. And they can’t afford to build a reactor that is tested to the limits of science – to the billions and billions of data points calculating the temperature, pressure and velocity through the enormous plant.
That’s where Fischer and the SHARP team come in, with simulations that can provide data and insight previously accessible only via expensive and extremely time-consuming experiments.
They’re known as large-eddy simulations because they try to capture the most significant parts of turbulent motion, the large ones, without resolving the smaller movements. “You use a model to capture the smaller scales of motion,” Fischer says. “Otherwise you would need a lot more points than what we have.”
Ultimately, the simulations can assess the uncertainty of the data points, a key factor in determining plant safety. “We can tell whether we are on an equilibrium point, whether we are stable or unstable,” Tautges says.
To know the flow
Fischer’s research looks at the thermal hydraulics of reactors, whether the flow is water in the case of light-water reactors, liquid sodium in the case of fast reactors or helium in the case of very fast gas-cooled reactors.
“We want to answer several questions: What is the peak temperature in the reactor, where does it occur and under what circumstances?” Fischer says.
Fischer’s team has built a computational model with a mesh of grid points at which temperature, speed of the flow and fluid properties are known. The SHARP project also has nuclear engineers working on a computational model to simulate neutron transport, or what happens with fissionable material, the source of heat in a nuclear reactor. Fischer’s computational model aims to determine how the heat moves and is conducted when the fuel pin heats up.
“We solve a partial differential equation governing the heat transport,” Fischer says. “It also requires computing the flow of the coolant, which is the principal challenge.” The simulation involves 1 billion grid points for one fuel assembly. “We need that many grid points to capture the small features of the flow,” Fischer says, “since each point can vary in velocity and pressure because the flow of the fuel is so turbulent.”
His 217-pin rod-bundle simulation, the one that involves 65,000 processors, requires 30 million hours of computer time on the IBM Blue Gene/P through DOE’s INCITE program (for Innovative and Novel Computational Impact on Theory and Experiment).
“It’s a big calculation,” Fischer says, “20 times larger than we were doing two years ago.”
Fischer’s video depiction of the simulation presents slices through a multi-pin mesh, divided into squares that look like plaid shirts, each containing an 8 by 8 grid of points. There are 3 million such little bricks, with the points spread over a space 10 centimeters on a side and 50 centimeters long.
Other scientific teams have done seven-pin calculations, but no one else has attempted the 217-pin simulation.
“Even this calculation doesn’t capture everything,” Fischer says. “But it’s certainly the most detailed to date.”
The next step is to compare the 217-pin data with the seven-pin data. The researchers aim to validate less expensive models and the experiments done in the 1970s.
“Once we’re all on the same page,” Fischer says, “people will be comfortable that we have validated the process to predict the behavior of these devices.”