Description
To enable efficient shot or controller design, we must be able to estimate plasma parameters, such as the electron density or temperature, from engineering parameters. Even when reduced to 0D aggregate quantities, such as the line-averaged density or the average core temperature, accurate simulation is difficult due to the non-obvious estimation and interplay of various sources and sinks. Additionally, many quantities are affected by hidden parameters dictated by machine conditions, such as unknown levels of recycling linked to plasma-wall interaction (wall fueling).
Addressing these challenges, we propose learning plasma dynamics using neural networks, that is, neural ODEs. Specifically, we present results on 0D modeling of the plasma electron density and temperature on TCV. We model these quantities as a function of engineering parameters connected to shaping, gas fueling, and heating (NBI, ECHR). In attempt to model unknown machine conditions, we propose the use of a hidden state, initialized with a short snapshot of the plasma’s response to the given actuators (e.g., from a past discharge, or an initial warm-up period). We demonstrate how we can understand this state as a ‘wall condition’ variable for a controlled setting, and expand to models that are more broadly applicable. Furthermore, we discuss efforts on automated processing routines to create high quality input datasets and fast inference pipelines, efforts to interpret model failure modes, and early studies for shot and controller design.