Identifying physical disruption boundaries in Alcator C-Mod using linear Support Vector Machines

The threat of disruptions in next-generation tokamaks and future fusion power plants has motivated a broad investigation into disruption avoidance and prediction. Among the various analytical tools applied to the disruption problem, machine learning (ML) has garnered interest for its ability to learn from empirical data and evaluate predictions in real-time. Unfortunately, the black-box nature of many ML algorithms creates uncertainty about their reliability, physicality, and generalizability. Here, we present two approaches to learning physically interpretable disruption boundaries using linear Support Vector Machines (lSVMs): a sum-of-polynomials fit and a power law fit. We demonstrate these approaches on data from C-Mod, examine their physical significance, and discuss their utility as disruption predictors. Future work will apply this approach to data from DIII-D.