A year and a half after Oregon State University launched the Center for Exascale Monte Carlo Neutron Transport, or CEMeNT, its researchers have displayed impressive progress in their quest to develop ultra-high-speed computer simulations for predicting the behavior of neutrons.
“Our work is intended to increase the fundamental understanding of neutron transport, which is vital in determining the safety, security, and viability of systems that involve neutron-induced reactions, like fission and fusion,” said Todd Palmer, professor of nuclear science and engineering at Oregon State and CEMeNT’s director.
In 2020, Oregon State was selected by the National Nuclear Security Agency to lead one of nine Predictive Science Academic Alliance Program centers. CEMeNT includes partner institutions North Carolina State University and the University of Notre Dame. Eleven of its 18 members are from Oregon State, including faculty researchers, postdoctoral scholars, graduate students, and undergraduates. NNSA has tasked the center with developing lightning-fast simulations using exascale computing technology.
Exascale computing refers to systems capable of performing at least 1 quintillion operations per second. The country’s first exascale computer, called Frontier, is expected to come on line in 2022 at Oak Ridge National Laboratory.
CEMeNT’s goals include building simulations that run hundreds to 1,000 times faster than is currently possible for NNSA. “Our algorithms exploit the exascale architectures, and we’ll demonstrate that they continue to perform well as we scale up the number of processors,” Palmer said. Until the Oak Ridge system is running, CEMeNT will use existing supercomputers at the Lawrence Livermore and Los Alamos national laboratories, and Oregon State’s NVIDIA DGX-2 systems.
To validate its ability to accurately predict real-world physics, CEMeNT will simulate a series of pulsed-sphere experiments conducted at Lawrence Livermore National Laboratory from the late 1960s until the mid-1980s.
In the experiments — about 70 in all — spheres of different materials and sizes were pulsed with a burst of high-energy neutrons. Detectors at specific distances and orientations from the targets measured neutron arrival times, from which neutron energy can be inferred. Simulating these experiments is nothing new; they were conducted to serve as benchmarks for simulation software — but nothing close to the speed or accuracy that CEMeNT is aiming for.
The plot on the left shows the analytic timeand space-dependent neutron population in a supercritical slab reactor; on the right, simulated results generated by CEMeNT’s software.The striking agreement between the simulated and the reference solutions demonstrates the precision of the group’s work.
“To make things even more challenging, we’ll run the simulations in time-dependent mode — a particularly challenging problem in radiation transport for the last decade,” Palmer said. Historically, because of the nanosecond timescales involved in neutron transport, simulations are performed in a steady state mode where it’s assumed that everything happens instantaneously. A dynamic model, representing a system as it changes, introduces significantly more complexity. The ability to incorporate the element of time into neutron transport simulations would enable NNSA to determine solutions at a level of precision not previously possible. “If our modeling is on target,” he continued, “then we’ve shown that we’re matching reality. It’s a wonderful test of our abilities.”
The researchers face some daunting computing challenges. For one, exascale computer hardware is heterogeneous, meaning it incorporates different types of processors — CPUs and GPUs — an architecture that is not naturally conducive to probability-oriented algorithms like Monte Carlo simulations, which estimate possible outcomes of uncertain events. “The problem is one of scheduling and memory use. You don’t want any individual processor to be waiting for others to finish calculations,” Palmer said. “It can be a bookkeeping nightmare.”
One possible solution is blending Monte Carlo with deterministic algorithms, resulting in large, nonrandom systems of equations. The combination can produce fast, accurate results on heterogeneous machines while reducing the statistical error associated with Monte Carlo simulations.
The group is attacking software development on two fronts. On one side, they’re adding a time dimension to an established Monte Carlo software code called Shift, which was initially designed to solve static physics problems.
On a parallel track, they’re developing original software. “That’s our sandbox where we can try out all sorts of new ideas,” Palmer said.
NNSA recently upped the ante by asking CEMeNT to identify an even more ambitious problem to solve. The researchers chose to simulate a well-documented 1946 incident at Los Alamos in which Canadian physicist Louis Slotin accidentally allowed fissile materials to release a burst of intense neutron radiation. He died from the exposure nine days later. The seven other men in the lab suffered varying degrees of radiation sickness.
“It’s even possible that some of the lab equipment played a role in the multiplication of neutrons,” Palmer explained, “and we need to model the intricate geometry of the situation and determine neutron behavior over time.” Simulating this Gordian knot of neutron transport is certain to establish the center’s prowess beyond any doubt.
CEMeNT was originally funded by a five-year, $4.3 million NNSA grant, but the agency increased funding by nearly $300,000 after appraising the group’s striking progress during its inaugural year.
Competition to become an NNSA lead research institution was fierce, according to Palmer, and Oregon State was selected over other universities with world-class nuclear programs. “This center shines a bright light on Oregon State and the School of Nuclear Science and Engineering,” he said. “We are capable and ready for the challenge.”