Researchers at the US Oak Ridge and Argonne National Laboratories used the Frontier supercomputer to train an artificial intelligence model capable of simulating the magnetic turbulence of plasmas.
A team of researchers, led by Semih Kacmaz and supervised by Eliu Huerta, took advantage of the phenomenal power of Frontier supercomputerlocated atOak Ridge National Laboratory (ORNL), to tackle one of the toughest problems in physics: the magnetohydrodynamic turbulence (MHD).
The goal? Train a artificial intelligence to model with unprecedented fidelity the chaotic behavior of plasmas, these supercharged gases which govern solar flares as well as star explosions or the magnetic environment of our own planet…and soon nuclear fusion reactors.
This method, combining a neural operator and a diffusion model, makes it possible to predict these complex cosmic phenomena with unprecedented precision, perhaps leading to new advances.
Why was magnetic turbulence such an enduring mystery?
Describing the behavior of plasma under the influence of magnetic fields is a real cosmic puzzle. This phenomenon is a chaotic ballet of swirls and energy flows that occurs at all scales in the universe.
Understanding this process is essential for modeling events as titanic as the formation of galaxies or as promising as the production of energy by nuclear fusion. The challenge is immense, because chaos is hard to fake.
Until now, traditional methods, such as the approach RANS (Reynolds-Averaged Navier-Stokes), oversimplified the problem. They smoothed out the finest details to obtain approximations, but in the process lost all the richness and complexity of real interactions.
To unlock the secret, an approach was needed that could capture both the big picture and the tiniest fluctuationsa mission for which classic models were simply not cut out.
How does this new artificial intelligence manage to dissect chaos?
To overcome the complexity of the problem, scientists have developed a strategy by two very distinct stagesforming a sort of shock duo ofartificial intelligence.
The first phase is based on a “ physics-informed neural operator “. This specialized algorithm learns the major laws that govern the behavior of the plasma, tracing the lines of force and the overall evolution of the system. He draws, in a way, the skeleton of the magnetohydrodynamic storm.
Modeling plasma movements using AI and Frontier (credit: Semih Kacmaz)
But the turbulence lives in the details. This is where the second actor comes in: a “ diffusion model “. This generative AI, similar to those that sharpen satellite photos, has been trained to piece together all the missing complexity.
It focuses on what the neural operator cannot see, regenerating the small swirls and rapid fluctuations that characterize chaos. By combining these two approaches, the system respects fundamental physical equations while capturing the smallest details of the phenomenon, reducing prediction errors by more than half compared to old methods.
What role did the computing monster Frontier play in this breakthrough?
Such an approach, as brilliant as it was, had a non-negotiable prerequisite: a absolutely colossal computing power. To train this dual AI model, it was necessary to be able to generate thousands of extremely high-fidelity plasma simulations, covering all possible levels of turbulence.
This was an IT bottleneck that made the project unfeasible until very recently. It took the power of a supercomputer exascale class. The name of this titan? The supercomputer Frontier de l’ORNL.
Capable of performing more than two quintillion calculations per second, it was, in the words of Semih Kacmaz, “ a true savior “. It was thanks to him that the team was able to generate the mountains of data needed to train both the neural operator on the big physics and the diffusion model on the fine details.
Concretely, what will these ultra-precise simulations be used for?
The fallout from this advance opens up dizzying prospects in two major areas. On the one hand, astrophysics. Scientists will finally be able to model cosmic events such as supernova explosions or solar flares with unprecedented precision.
Understanding the dynamics of plasma in these extreme environments is the key to deciphering many of the mysteries of the universe.
On the other hand, and closer to us, this research has a direct impact on the quest for clean and unlimited energy: nuclear fusion. Future reactors, like tokamaks, work by confining extremely hot plasma using magnetic fields.
Turbulence is the number one enemy of this confinement. By faithfully modeling this disruptive elementengineers will be able to design more stable and more efficient reactors, bringing us a little closer to the energy Holy Grail.
The team already aims to extend its model to full 3D simulations, promising even more spectacular discoveries.
