Description
Reactor-scale divertor design requires rapid yet physics-consistent exploration of exhaust operational spaces. Fusion power plants beyond ITER demand integrated scenario development that maintains core-plasma performance while meeting engineering limits on plasma-facing components, particularly heat loads and erosion. Accelerating the evaluation of key divertor metrics - target fluxes, detachment behaviour, neutral dynamics, and pumping efficiency - remains challenging. In addition, predictive capability comparable to the full SOLPS-ITER higher fidelity model is needed to capture upstream conditions such as separatrix density, impurity concentration, and fuelling at the same time. We present an AI-enabled workflow that addresses this need by combining SOLPS-ITER databases with data-driven surrogate modelling.
The workflow employs the SOLPS-NN surrogate model, trained on wide-grid SOLPS-ITER simulations using the advanced fluid neutral (AFN) model. These surrogates reproduce the dominant plasma–neutral physics governing the transition from attached to detached regimes while providing orders-of-magnitude computational speedup relative to full SOLPS-ITER simulations with EIRENE. To improve fidelity, transfer learning retrains AFN-based surrogates on a smaller set of high-fidelity kinetic-neutral simulations, enabling capture of non-Maxwellian transport, molecular processes, and detailed ionisation–recombination dynamics without exhaustive kinetic modelling. Active-learning strategies can further refine the training set by targeting nonlinear or high-uncertainty regions.
Employing two test configurations, (a) the EU-DEMO low-aspect-ratio and (b) the European volumetric neutron source (EU-VNS), it is demonstrated how the AI based workflow efficiently maps exhaust operational spaces and identifies regimes consistent with power-handling constraints. Comparisons with reduced models highlight where simplified approaches remain valid and where kinetic-neutral effects are essential. Post-processing of 2D surrogate outputs enables rapid generation of heat and particle flux maps for plasma-facing components and supports fast synthetic diagnostics for digital-twin applications. Integration into the JINTRAC and ASTRA frameworks is underway, enabling computationally efficient core–edge coupling and dynamic boundary-condition exchange for future pulse design tools.