Multi-machine Disruption Databases and Applications of Machine Learning for Fusion

MIT PSFC researchers are applying Machine Learning to a number of fusion-relevant problems, including real time disruption warning on a number of different tokamaks, automated identification of transport confinement mode in Alcator C-Mod, and determination of impurity transport coefficients in C-Mod. A broad selection of Machine Learning methods are being used: Random Forests, recurrent and feedforward neural networks, Gaussian Bayesian, and logistic regression. All AI methods require extensive databases of relevant information in order to train and test the algorithms. For disruption prediction, we would like to know if a universal algorithm can be derived from current tokamaks, so that ITER, SPARC, and future reactors would not have to first generate their own databases of disruptions. To this end, we have established similar databases of real time disruption-relevant plasma parameters from four very dissimilar tokamaks (C-Mod, DIII-D, EAST, KSTAR). We find that the most useful disruption warning parameters are different on each machine, and that the disruption prediction performance of algorithms trained on each machine's database vary considerably.