Confinement mode identification of fusion plasmas utilizing machine learning methods

Distinguishing features between fusion plasma confinement regimes are explored via machine learning methods to analyze experimental data from the compact, high-field Alcator C-Mod tokamak. Supervised learning techniques with zero-dimensional data and time-independent quantities are employed which increases the generalizability of this approach for instant confinement mode identification and ultimately real-time prediction purposes for fusion devices. Binary classification of L- and H-modes utilizing Gaussian naïve Bayes, logistic regression, multilayer perceptron (i.e. feedforward neural networks), and random forests performed similarly and obtained an average accuracy of 97.2% for L-modes and 86.7% for H-modes using the plasma’s stored energy, volume-averaged density, poloidal beta, ohmic heating power, normalized internal inductance, magnetic axis radial position, and Hα as inputs. Additionally this work investigates I-modes leading to a multi-class classification problem. Development of a new confinement database with over 200 distinct shots consisting of approximately 400 L-, 200 H-, and 100 I-mode periods extends previous databases for large-scale comparative studies.