Identifying sustained reverse magnetic shear (RMS) in NSTX discharges with unsupervised learning

The majority of NSTX discharges start with an early phase of reverse magnetic shear (RMS), often followed by some MHD events leading to current redistribution and monotonic safety factor, q(r), profile. A database consisting q and magnetic shear profiles is constructed of NSTX data with LRDFIT equilibrium reconstructions in order to identify and study long-lived reversed magnetic shear (RMS) discharges, with the aim to achieve an understanding of MHD events for future predictions. LRDFIT utilizes the radial electric field (E_r) corrected pitch angle measurements obtained using the multi-channel motional Stark effect (MSE) diagnostic. Based on the derived q-profile, an unsupervised k-means clustering of the data is developed to study RMS formation and its duration as a function of time. Starting with a small group of shots comprised of short-lived and sustained RMS cases, the method is tested. The cumulative proportion variance explained (PVE) analysis of the clustering results indicated about 60% accuracy of the model. By increasing the number of parameters and shots included in the analysis, the cumulative PVE is increased to about 80%. A supervised algorithm utilizing these results as a preprocessing step will be developed in the near future.