Developing a Database for a Machine Learning Algorithm utilizing Multi-Point Motional Stark Effect (MSE) Diagnostic

To improve the efficiency of developing a target plasma in NSTX-U and make better use of valuable machine time, a machine learning (ML) algorithm collecting features of the multi-point MSE data, and its associated q-profiles is underway. We aim to identify the patterns leading to MHD events that can cause changes in the current distributions and q-profiles. We started constructing a database based primarily on LRDFIT equilibrium reconstructions utilizing MSE, consisting of pitch angle profiles and their residuals, q-profiles and the magnetic shear profiles. As we increase the total number of runs stored in the database to 140+, we are also investigating features in the data by utilizing unsupervised learning techniques for clustering. For this purpose, a subset of 24 shots, showing signatures of early reversed magnetic shear was created. Each array of q₀-q_min as a function of time per shot was then fed into a k-means clustering code using 2D principal component analysis (PCA) to compute the cluster means. The resulting clustering shows significant differences in both amplitude and the duration of the reversed magnetic shear. As the next step, we plan to use the clustering to detect and distinguish shots with strong magnetic shear as a preprocessing step in a supervised ML algorithm with higher number of components involving various plasma parameters.