Classification and model discovery of edge localized modes on DIII-DZ. Daniel¹, C. Hansen¹, D. Smith², S. Joung²1- Columbia University2- University of Wisconsin-Madison

Edge Localized Modes (ELMs) introduce high heat and damage plasma-facing components. Therefore, ELM free regimes that can deliver performance comparable to H-mode without damaging the wall are of great interest. Understanding the dynamics of Edge Localized Modes (ELMs) is crucial for understanding the transition between ELM-ing and ELM free regimes. Robust classification of ELM events into clusters that represent distinct behavior can serve as a means to validate and develop models of ELM dynamics. In this work, we apply hierarchical clustering to a database of 3651 labeled ELM events from DIII-D. For this analysis, the multichannel Beam Emission Spectroscopy diagnostic is projected into a one dimensional time series using principal component analysis as in [1]. Three similarity metrics are applied to the projected data, with each seeding a different hierarchical clustering analysis. The learned clusters are then compared across different plasma parameters and figures of merit to determine if the classifications yield meaningful insights. Additional machine learning techniques that respect the two dimensional nature of the full BES data for model identification, and clustering analysis are also presented.

1 - D. R. Smith et al. PPCF 58, 045003 (2016)

Work supported by US DOE under awards DE-AC02-09CH11466, DE-AC02-06CH11357, DE-FC02-04ER54698, DE-FG02-08ER54999, and DE-SC0021157.