ELM Regime Classification by Conformal Prediction on an Information Manifold

Characterization and control of plasma instabilities known as edge-localized modes (ELMs) is crucial for the operation of fusion reactors. Recently, machine learning methods have demonstrated good potential in making useful inferences from stochastic fusion data sets. However, traditional classification methods do not offer an inherent estimate of the goodness of their prediction. In this paper, a distance-based conformal predictor classifier integrated with a geometric-probabilistic framework is presented. The first benefit of the approach lies in its comprehensive treatment of highly stochastic fusion data sets, by modeling the measurements with probability distributions in a metric space. This enables calculation of a natural distance measure between probability distributions: the Rao geodesic distance. Second, the predictions are accompanied by estimates of their accuracy and reliability. The method is applied to the classification of regimes characterized by different types of ELMs based on the measurements of global parameters and their error bars. This yields promising success rates and outperforms state-of-the-art automatic techniques for recognizing ELM signatures. The estimates of goodness of the predictions increase the confidence of classification by ELM experts, while allowing more reliable decisions regarding plasma control and at the same time increasing the robustness of the control system.