Abstract Details
Abstracts
Author: Ben Zhu
Requested Type: Poster
Submitted: 2026-04-14 17:29:19
Co-authors: L.Sun, J.Guo, J.X.Wang
Contact Info:
Columbia University
500 W. 120th St., #200
New York, New York 10027
USA
Abstract Text:
The modified Hasegawa-Wakatani (MHW) model, which captures the essential physics of drift-wave turbulence, spontaneous zonal flow generation, and their complex nonlinear interactions within a local, slab approximation, is often considered a minimal or toy model for understanding magnetized plasma turbulent dynamics. The MHW model, alongside the original model, has been extensively studied for decades and is now emerging as an ideal testbed for AI/ML applications in plasma physics before moving to more complex systems. Recent advancements have successfully utilized the HW systems to validate Physics-Informed Neural Networks (PINNs), develop novel acceleration algorithms for numerical solvers, rapidly predict macroscopic mean quantities, and forecast critical transition states.
Here, we present two additional exploratory studies designed to expand the frontier of ML-driven plasma physics. The first study is the development of a conditional generative foundation model tailored for plasma turbulent dynamics. This model is capable of predicting highly nonlinear turbulence quantities and flow structures, enabling advanced applications such as physical state super-resolution and missing data inpainting. In the second study, we introduce external electrodes as actuators and apply Reinforcement Learning (RL) to active turbulence control. By treating the highly nonlinear MHW system as an environment, the RL agent successfully discovers non-trivial optimal trajectories to suppress turbulent transport more effectively. Additionally, this RL-discovered control policy is robust to various initial conditions despite the chaotic nature of the plasma turbulence. Together, these approaches demonstrate the potential of generative AI and reinforcement learning in diagnosing, predicting, and ultimately controlling complex plasma dynamics.
Characterization: 6.0
Comments: