Uncertainty Quantification for Machine Learning Models in Plasma Physics

While several machine learning models exist for predicting future states of plasma, little work has been done to evaluate how reliable these models are, and how their accuracy varies with different inputs and plasma types. As most physics-based models of plasmas include many approximations, it is difficult to discern whether machines are learning the correct patterns, or if they are simply memorizing data. The ability to accurately predict plasma states is integral to achieving the control needed for a stable, high performance fusion reaction. We demonstrate methods to quantify the uncertainty of machine learning models, as well as methods to improve the predictive accuracy of existing models. We further describe methods to generate predictions and accurate uncertainty bounds for both existing models and newly developed ones.