Speaker
Description
Surrogate enable the rapid evaluation of plasma quantities across different divertor operating conditions. Building on previous work on surrogate modelling for geometry, fuelling and impurity seeding studies, this work focuses on the continued development and evaluation of deep learning-based surrogates trained to predict 2D plasma variables from simulation inputs (both 1d and 2D).
We present an investigation into the impact of model design choices, including input feature selection, multi-output versus single-output training, model size, embedding dimension and dataset scaling. We will also demonstrate the surrogate's performance across conventional, Super-X and expanded divertor configurations, highlighting its ability to generalise across magnetic geometries and outlining downstream applications where rapid surrogate predictions can accelerate analysis and parameter studies.