Machine Learning Acceleration of Tokamak Scrape-Off Layer Modelling
George Holt
UK Atomic Energy Authority
Effective heat and particle exhaust management is critical for tokamak fusion reactors, requiring accurate scrape-off layer (SOL) modelling for experimental interpretation, scenario optimisation and reactor design. Current approaches face a fundamental trade-off: fast analytical models enable rapid design iteration but omit essential physics, while comprehensive simulations provide high physics fidelity but require prohibitive computational resources, limiting their use in time-critical applications. In this work, a framework is presented that mitigates this trade-off by accelerating high-fidelity SOL simulations with machine learning. An engine for automation and scaling of the EDGE2D-EIRENE code was developed and used to generate large-scale data sets comprised of tens of thousands of simulations, systematically varying key parameters including input power, fuelling and impurity seeding rates and locations, magnetic geometry, and transport coefficients. Multiple machine learning model architectures were trained, including neural networks, ensemble methods, and neural operators, tailored for distinct applications in plasma state reconstruction (e.g., 2-dimensional electron temperature maps), derived quantity prediction (e.g., divertor target peak heat flux), and boundary condition generation for core-edge integrated modelling. The trained models deliver millisecond inference while maintaining the physics fidelity of the underlying fluid plasma and kinetic neutrals simulations, demonstrating how data-driven approaches can be used to allow high-fidelity SOL models to be deployed for experimental support and comprehensive design space exploration.