The search for habitable worlds keeps running into the same wall. We can see the planets. We can see, in ultraviolet light, how violent their host stars are. What we cannot afford to do is simulate, from first principles and at scale, the plasma physics that decides whether a flare or a coronal mass ejection strips a close-in rocky planet of its atmosphere. Chuanfei Dong's work is about closing that gap, with AI inside the physics and a UV telescope outside it.
The space problem is concrete. In a 2018 paper, Dong argued that close-in rocky planets around active M-dwarf stars, including the worlds in the TRAPPIST-1 system, would struggle to hold onto their atmospheres in the face of strong stellar wind and ionizing radiation. Later James Webb Space Telescope observations of TRAPPIST-1 b and c are consistent with the parts of that concern that have been tested so far, while planets further out in the same system remain under study. The same plasma physics shows up much closer to home, in the space weather that drives geomagnetic storms, threatens satellites, and stresses power grids. In both cases, a serious answer needs kinetic plasma behavior in domains where running a kinetic simulation directly is out of reach.
That is the obstacle Dong's recent paper, published in the Proceedings of the National Academy of Sciences with Ziyu Huang as lead author and Liang Wang as collaborator, sets out to address. The team trains a Fourier Neural Operator on first-principles Vlasov simulations of a kinetic plasma, then plugs the trained operator into a much cheaper fluid model as a heat-flux closure. The resulting hybrid reproduces nonlinear Landau damping, a textbook kinetic phenomenon, at a small fraction of the cost of the Vlasov reference. Dong, who advised the work as principal investigator and held a 2024 Alfred P. Sloan Research Fellowship, frames the design choice cleanly. The FNO does not replace the physics solver. It learns one operator the solver needs and otherwise cannot get right, then hands the rest of the simulation back to the equations.
For space science, that is the point. The same kinetic-inside-a- fluid-model bottleneck is what makes high-fidelity space weather prediction around Earth, modeling of the solar wind's interaction with planetary atmospheres, and any honest attempt to forecast atmospheric loss at an exoplanet so expensive today. A closure that carries kinetic physics into a fluid run is, in principle, the seam where AI changes what the field can afford to ask. The honest framing of the PNAS result is that it demonstrates the trick in one regime against the harder reference, and opens a clear question about how far the same closure idea travels into magnetospheres, stellar winds, and exoplanet environments.
The observational half of the same program is Mauve, where Dong is a principal investigator. Mauve is a small ultraviolet telescope that launched in late 2025 for a three-year survey of stellar magnetic activity, flares, and high-energy radiation, the very inputs that any habitability claim about a close-in rocky planet has to respect. The modeling work and the mission are not two stories. They are the two halves of one question: measure how loud the star actually is, and have a physics model fast enough to ask what that loudness does to the planet next to it. That is what AI for space looks like in Dong's hands, and it is the through-line that earth-space-ai.org is interested in. The goal is not a faster plot. It is a working answer to whether a given world is a barren rock or still a candidate for an atmosphere.