Scientists Make It One Billion Times Faster to Simulate Fusion Reactors

Researchers at the Oden Institute for Computational Engineering and Sciences at The University of Texas at Austin have unlocked the door to quantifying uncertainty in real-world problems.

Published in the journal Communications Engineering from Nature, their findings are applicable to fields ranging from plasma physics to weather prediction.

Uncertainty Quantification (UQ) is the work to understand and reduce uncertainties in numerical simulations. UQ often requires thousands of highly accurate simulations, taking hundreds of thousands of hours on a supercomputer to perform. 

Dr. Ionut Farcas and his collaborators, Drs. Gabriele Merlo and Frank Jenko, set out to trim the least necessary parts of these simulations through sensitivity analysis (SA), to make them just as accurate but far less expensive to run. 

Their approach was to exploit the fact that in real-world problems there are often many uncertain inputs, but usually only a small subset of them is important. 

SA determines which inputs can be blamed for the uncertainty in a simulation’s output. This can guide the development of a simplified model or help scientists explain the results of an experiment. The same principle is what makes it possible to watch a movie on your phone: decreasing pixel count in imperceptible ways by eliminating tiny details like shadows around an actor’s face. 

Another successful application is weather prediction. UQ is the difference between a possibility of rain sometime this weekend and a light drizzle beginning at 2:32 PM and ending at 8:47 PM tonight.

A computational scientist like Farcas would call this the transition from interpretable numerical simulation to predictive numerical simulation. 

“To get from an interpretable simulation to a predictive simulation, you have to quantify uncertainty,” Farcas explained. “Uncertainty is present in most situations due to things like measurement errors.”

To demonstrate that their approach enables both UQ and SA, the three researchers employed it in the context of plasma physics. The overarching goal of plasma physics is to replicate what happens in stars with the end goal of producing energy. The tools of this trade are fusion power devices like tokamaks and stellarators. 

 

“This is where computational mathematics shines, because we can probe experimental settings that cannot be tested otherwise,” Farcas said.

The researchers set out to show that their approach to uncertainty can enable UQ and SA even in complex problems. In doing so, they hope to rid their peers of an all-too-common stumbling block. 

"Problems like plasma physics that involve quantifying uncertainty in computationally expensive simulations are very challenging,” Farcas said.

In his research, Farcas sat down with plasma physicists to find a specific uncertainty-rife scenario relevant to their work that is expensive to simulate: turbulence in nuclear fusion reactors. Farcas’ final product is an accurate ‘surrogate’ model that is one billion times faster than a high-fidelity model.

It took the researchers 8,173 core-hours on a supercomputer to run a traditional computational model of fusion reactor turbulence. By using only the inputs they deemed necessary, their simulation took 9.4 milliseconds to run on a laptop computer. This is a time difference of nine orders of magnitude: one billion. 

“This will help scientists and engineers find answers faster,” Farcas said. 

These simulations were performed on the Frontera supercomputer at the Texas Advanced Computing Center (TACC). Frontera is one of the most powerful supercomputers in the world, and the fastest supercomputer on a college campus. All three researchers were supported by the Exascale Computing Project, a collaborative effort of the U.S. Department of Energy Office of Science, and the National Nuclear Security Administration.