Speaker
Description
Accurate scrape-off layer (SOL) modelling is essential for tokamak heat and particle exhaust studies, supporting experiment interpretation, scenario optimisation, and reactor design. Existing methods require a compromise between speed and fidelity: analytical models are fast but simplified, while comprehensive simulations capture plasma and neutral physics but are too costly for rapid-turnaround applications.
This work presents a machine-learning framework for accelerating high-fidelity SOL modelling. A scalable automation workflow for EDGE2D-EIRENE was developed to generate tens of thousands of simulations spanning input power, fuelling and impurity seeding rates and locations, magnetic geometry, and transport coefficients. Neural networks, ensemble methods, tabular foundation models and neural operators were trained for tasks including two-dimensional plasma state reconstruction, derived quantity prediction (e.g., divertor peak heat flux), and boundary condition generation for core-edge integrated modelling.
The resulting surrogate models provide millisecond inference while retaining the fidelity of the underlying simulations, enabling high-fidelity SOL modelling for experimental support and broad design-space exploration.