Nuclear fusion power, considered the ultimate energy source, operates on the same principle as how the Taeyang produces energy, and is commonly referred to as the "artificial sun" project. Domestic researchers have developed artificial intelligence (AI) that can simulate the plasma state within a fusion reactor 1,000 times faster than before.
Professors Lee Ji-min and Yoon Ui-sung from the Ulsan National Institute of Science and Technology noted on the 17th that they developed a deep learning-based AI model called "FPL-net" that can quickly find solutions to mathematical equations describing the plasma state. The research was published in the international journal "Journal of Computational Physics" on the 15th.
In nuclear fusion power technology, known as the "artificial sun," it is essential to maintain the interior of the generator in a high-temperature plasma state similar to that of the Taeyang. Plasma is a state in which matter is separated into negatively charged electrons and positively charged ion particles, and accurately predicting the collisions between particles in this state is key to maintaining stable fusion reactions.
The plasma state is represented by mathematical models, one of which is the "Fokker-Planck-Landau equation (FPL)." The "Fokker-Planck-Landau equation" predicts the collisions of charged particles with both positive and negative charges, known as "Coulomb collisions." Until now, iterative methods that gradually find solutions have been used to solve this equation, but there has been a limitation due to high computational requirements and time consumption.
The FPL-net developed by the researchers can find solutions to the equation all at once, unlike the iterative methods previously used. It can obtain solutions 1,000 times faster than before, showing high accuracy with a prediction error of 1/100,000. The process of Fokker-Planck-Landau collisions has characteristics where density, momentum, and energy are conserved, and the researchers defined functions during the AI model training process to ensure the conservation of these physical quantities, enhancing accuracy.
The researchers said, "While maintaining accuracy, we shortened computational time by 1,000 times compared to existing codes that used a central processing unit (CPU) by utilizing deep learning with a graphics processing unit (GPU)," adding that it would serve as the cornerstone for turbulence analysis codes that simulate the entire fusion reactor area, or for digital twin technology that implements real tokamaks in virtual computer space.
A tokamak is a special structure that confines plasma. The researchers added, "This study is limited to electron plasma; however, research is needed to extend to complex plasma environments with multicomponent particles containing impurities for applications."
References
Journal of Computational Physics (2025), DOI: https://doi.org/10.1016/j.jcp.2024.113665