Data efficiency and long-term prediction capabilities for neural operator surrogate models of edge plasma simulations

Abstract: Modelling of plasma dynamics is fundamental to ensure appropriate diverter and core performance, and is desirable for both interpreting the current generation of experiments and informing the next generation devices like ITER \cite{Loarte2007Chapter4P,Eich2013ScalingOT}. Yet the computational expense of many plasma simulations makes them unsuitable for real-time applications or iterative design workflows. Neural operator surrogate models of JOREK \cite{Hoelzl_2021} and STORM \cite{Walkden2016-ys} are evaluated, investigating their capability to replicate plasma dynamics accurately whilst reducing computational cost. It is found that the accuracy of the surrogate models will degrade for long term predictions, and that physics considerations are important in assessing the performance of the surrogates. Surrogates trained on one dataset can be effectively fine tuned with only a few simulations from a target domain. This is particularly effective where the source domain is a low fidelity physics model and the target domain is a high fidelity model, with an order of magnitude improvement in performance for a small dataset and a short rollout.