Development of an automated neural network approach for classifying MHD mode formation using ECEI data for Disruption Event Characterization and Forecasting (DECAF)

Plasma disruptions are often driven by magnetohydrodynamic (MHD) events that lead to the formation of magnetic islands through tearing mode instabilities. Recent studies using the Electron Cyclotron Emission Imaging (ECEI) system have focused on the role of turbulence near magnetic islands in their growth and potential to trigger disruptions [1]. We develop a neural network model to predict whether MHD mode formation is progressing toward disruption events using KSTAR ECEI data. Coherence between vertically adjacent ECEI channels near the q=2 surface is calculated and mapped onto the poloidal plane. Increased coherence levels are often observed around the (2,1) mode prior to disruption. The model is trained on the spatiotemporal patterns of mapped coherence from multiple shots exhibiting global electron temperature collapse events. Leveraging the Disruption Event Characterization and Forecast (DECAF) code [2], this approach aims to uncover the physical mechanisms linking rapid electron temperature changes to mode growth, saturation, and evolution, offering new insights into the event chains that ultimately lead to disruptions. This offline analysis will support the development of disruption avoidance techniques using real-time processing of full 2D ECEI images of MHD modes.

[1] M. J. Choi, et al., Nat. Commun. 12, 375 (2021)

[2] S.A. Sabbagh, et al., Phys. Plasmas 30, 032506 (2023)