Data augmentation for disruption prediction via robust surrogate models

This work presents methodological developments regarding disruption prediction using machine learning techniques. When working with neural networks, a comprehensive training data base is important to achieve satisfying and reliable results. Here, we aim for a robust augmentation of the training data base using surrogate models. 

Gaussian processes (GPs) can act as surrogate models to enlarge the training data base giving the covariance structure in addition. However, the computational complexity of standard GP regression increases with the third power of training data points and outliers are punished very severely, which results in unreliable uncertainty estimates. These drawbacks complicate the application of standard GP regression to noisy high-resolution time series data. 

Here, these difficulties are addressed using Student-t processes in combination with a state space representation allowing for inference via Bayesian filtering. While the Student-t process allows a heavy tailed noise distribution and is more robust against outliers, the computational complexity of Bayesian filtering is linear in time and thus can also be used if the time resolution is high. Results based on data from recent tokamak experiments are presented.