When Atoms Collide

Harnessing fusion power as an endless source of low-cost energy has been  an intriguing challenge for scientists. Just  ask George Vahala, who has relied on  the computing power at NERSC to help  other scientists understand the dynamics  of generating fusion energy. 

Vahala, a physics professor at the  College of William and Mary, has been  one of the most active users of NERSC  resources this year. He has spent most  of the 1.18 million computing hours on  tackling tricky problems for modeling the  flow of hot plasma inside a magnetic  chamber as the atoms collide and fuse, a  leading approach for generating  megawatts of power in a few seconds. For years, Vahala has worked on  improving the lattice-Boltzmann method,  a relatively new approach to simulating  fluid flow that is ideally parallelized and  scales almost perfectly with the number  of processors available. By embedding  the fluid dynamics into a higher dimensional kinetic space, the solution trajectories are easier and more efficiently  solved. The trick is to reduce the required  memory and to ensure numerical stability  at very low transport coefficients.  

Over the years, Vahala and his team  have made tremendous progress in solving one of the toughest problems: how to  ensure the stability of a lattice-Boltzmann  code when a high Reynolds number is  used to better quantify turbulence? His  work also has made it possible for the  leap from two-dimensional to three- dimensional modeling using lattice- Boltzmann methods.  While the initial runs were performed  on the latest Department of Defense  machines (an IBM-P5), it was absolutely  necessary to turn to the old work-horse,  NERSC’s Seaborg machine (an IBM-P3),  to complete the simulation runs in 48-  hours bursts on 4096 processors.  

“Seaborg may be viewed as slow  nowadays – but its many processors and  large memory made it utterly indispensable for our research,” Vahala said.  In many instances, the method has  proven to work better than other computational fluid dynamic codes in modeling  systems with complex physical geometries. In fact, researchers have used lat- tice-Boltzmann to study a wide range of  phenomena, from star formation to the  dynamics of cell proteins. In the commer- cial sector, oil companies have turned to  the lattice-Boltzmann method for figuring  out how to extract more black gold out of  the field.  

For the past seven years, lattice- Boltzmann has been used to characterize  magnetohydrodynamics, the effects of a  strong magnetic field on the turbulent  flow of electrically conducting fluids such  as plasmas. Moreover, the algorithm  automatically permits the divergence of  the magnetic field to be zero to machine  accuracy. Understanding magnetohydro- dynamics, in turn, will help researchers  design better tokamaks, the machines that create the magnetic field.  In 2005, he co-authored a paper showing how well his three-dimensional lattice- Boltzmann code scaled on the Earth  Simulator. The code achieved a sustained  26.25 teraflop/s, registering the best performance by a scientific code on that  supercomputer. Jonathan Carter from  NERSC was the paper’s lead author while  Lenny Oliker from Berkeley Lab’s  Computational Research Division was  one of the co-authors. The research was  published in the Proceedings of SC05.  

Vahala’s other recent publications include “Entropic Lattice Boltzmann  Representation Required to Recover  Navier-Stokes Flows”, Physical Review,  E75, 036712::1-11 (2007), “The Lattice  Boltzmann Representataion for Plasma  Physics” in Physica A362, 48-56 (2006)  and “Quantum Lattice Representation for  Vector Solutions in an External Potential”  in Physica A362, 215-221 (2006).  

Aside from Carter, Vahala’s current  research team includes Jeffrey Yepez  (Hanscom Air Force Base), Linda Vahala  (Old Dominican University) and Min Soe  (Rogers State University).