The Faculty of Information Technology at the Czech Technical University in Prague (FIT CTU) has launched a new research collaboration with the University of California, Berkeley—one of the world’s most prestigious universities and a long-standing leader in computational science, physics, and artificial intelligence. With a rich history of Nobel laureates and pioneering discoveries ranging from the cyclotron to modern high-performance computing, Berkeley sets the highest standard of scientific excellence.
The goal of this collaboration is to significantly accelerate one of the most demanding simulations in physics—the behavior of plasma, the charged and turbulent “fourth state of matter” that fills stars, fusion reactors, and the space around our planet—by allowing neural networks to learn parts of the physics. Traditional particle-in-cell (PIC) simulations consume vast amounts of computational time by repeatedly solving the same equations (Poisson, Maxwell, Vlasov–Poisson), which describe how moving charges generate electric and magnetic fields. The joint FIT CTU–Berkeley team is training machine learning models to estimate these calculations directly, preserving accuracy where it matters while reducing computation time from hours to minutes for problems that would otherwise be out of reach.
A second line of research addresses a long-standing issue in PIC codes—particle noise. Because simulated plasma is represented by a finite number of macroparticles, the fields computed from their positions are affected by statistical noise. A simple solution—using more particles—is computationally expensive. The team is therefore exploring machine learning methods, specifically denoisers and physics-informed neural networks (PINNs), which can reconstruct fields from noisy data while preserving physical laws. The aim is to separate the true signal from noise using knowledge of the governing equations.
The third direction goes even further—training emulators capable of “skipping” multiple time steps at once. This would make it possible to more efficiently simulate phenomena occurring across vastly different temporal and spatial scales.
While Berkeley’s reputation in physics and computational science is unparalleled, FIT CTU brings its own distinctive strengths. The faculty has built a strong tradition in machine learning and applied mathematics, and its students are highly talented, selected through a competitive process and trained to combine strong theoretical foundations with real-world capability.
PhD student Ing. Matěj Jech, supervised by Mgr. Alexander Kovalenko, Ph.D., is contributing to WarpX, an open-source exascale PIC framework developed by the Berkeley team and used by plasma physicists worldwide.
“Our work aims to integrate machine learning components directly into the WarpX codebase, so that advances resulting from this collaboration do not remain confined to papers, but become tools that the broader community can download, inspect, and build upon,” says Matěj Jech.
