An active learning approach for determining MHD edge stability boundaries

Assessing the stability of MHD equilibria is important for understanding and mitigating Edge-Localized-Modes (ELMs) in tokamaks. However, performing a detailed simulation of every possible machine configuration and scenario is extremely time consuming. Surrogate models trained using machine learning (ML) can potentially mitigate this. An accurate ML model trained on a dataset that correlates configurations with stability properties can be useful for predicting the stability properties at new machine settings. In our PEDestal Active Learning (PEDAL) approach, a dataset for training such a model is grown iteratively. We consider a configuration space spanned by modifying pressure gradients and current densities in the vicinity of an existing equilibrium. The ML model takes as input discretized pressure and current profiles and predicts the growth rates across different modes. During training, new points for simulation are selected based on a uncertainty measure computed from an ensemble of models trained on the data. Specifically, by scaling the pressure gradient and current densities associated with existing equilibria, new equilibria can be obtained, and MHD stability analysis can be performed for these new equilibria. The stability analysis produces a full mode structure at each simulation point. This new information is added into the dataset, and the model is retrained after each new point is added. This helps us refine an estimate of the stability boundary in the space of profiles.