Flow regime identification for plasma turbulence with machine learning

Toward the access to purely sustainable energy, one of the promising enablers is fusion energy, which leverages heat production from nuclear fusion. One way to harness this phenomenon is to use a reactor where magnetic field is applied to confine plasma as fuel for a sufficiently long time, and heat is continuously provided to the inside for maintaining the plasma state. Today, a major obstacle to steady-state operation of such fusion reactors and net power generation is the significant energy loss due to plasma leakage to the outside of the magnetic confinement. This is known as anomalous transport due to plasma turbulence driven by plasma density gradient, which is needed to be suppressed. A key feature for reducing the anomalous transport is self-organized and shear-dominant flow structure inherent in plasma turbulence: zonal flow. Such a physical background is a strong motivation to develop control strategies to utilize zonal flow for suppressing plasma turbulence, while it is still difficult to derive control laws for those phenomena in the physical space due to their high-dimensionality. In this study, we leverage a nonlinear autoencoder to compress flow field data of plasma turbulence. Here, the modified Hasegawa-Wakatani equation is considered as a toy model including a variety of flow regimes between turbulent and zonal flows. In the talk, we will show that the present approach results in a zonal flow-aware representation in a low-dimensional space.