A Bayesian Framework for Multi-Diagnostic Inference of Tokamak Divertor States

A critical challenge for future nuclear fusion tokamak power plants is the controlled management of heat and particle exhaust. The divertor, a specialised region within the tokamak, is responsible for managing this exhaust. However, the divertor plasma exhibits complex, multi-dimensional dynamics involving various species. No single diagnostic can provide the required comprehensive experimental insight.

This thesis outlines and assesses developments to a divertor multi-instrument Bayesian analysis system (D-MIBAS) designed to infer, given available measured experimental data, which plasma states were plausible. These plasma states (two-dimensional profiles of electron temperature, electron density, and neutral hydrogen density) offer a deep insight into divertor plasma dynamics and enable direct comparison between interpretive simulations and experiments. D-MIBAS uses a mesh-based approach, spatially aligned to the magnetic equilibrium, to provide a natural way to describe the plasma state and bring together different diagnostics and prior physics knowledge within a single, consistent framework.

Through synthetic and experimental validation using the MAST-U tokamak, this work demonstrates the benefit of combining multiple diagnostics and embedding physics information within the D-MIBAS mesh-based framework. This represents a significant advance: a physics-informed machine learning framework to reliably infer divertor plasma states in a non-machine, non-scenario specific manner. Ultimately, this can inform both the design and control of future fusion tokamak exhaust systems.