Speaker
Description
Modern devices such as KSTAR and ITER are crowded with heating ports, cooling channels, and diagnostics, and rely on auxiliary heating such as neutral beam injection (NBI), so that beam power deposition and the resulting plasma-facing component (PFC) heat loads are governed by real, non-axisymmetric engineering geometry that idealized toroidally symmetric models cannot capture. CAD-to-simulation tools now embed this engineering detail by tracing field lines or ion orbits through 3D CAD geometry to compute power deposition and PFC heat flux (e.g., SMARDDA; the HEAT toolkit [1, 2]), and are being assembled into AI-augmented digital twins (e.g., DIII-D and MAST-U in NVIDIA Omniverse [3, 4]), moving the field toward an AI-native treatment of engineering detail. Despite this progress, experimental validation that such CAD-based predictions reproduce the measured 3D global heat-load footprints — including the localized peaks that arise at leading edges and limiters — remains limited, yet is a prerequisite for trusting these tools in design and between-shot operation.
We report a qualitative validation of Monte Carlo orbit following code NuBDeC [5], a modular and unified CAD-to-simulation framework that performs high-performance fast-ion–wall collision detection on CAD-derived unstructured meshes and maps the per-face heat flux deposited by NBI-induced prompt fast-ion losses onto the 3D PFC surface, with a CAD categorization that links every mesh face to its engineering component (divertor, poloidal limiter, passive stabilizer). Because the pipeline operates on the 3D geometry with engineering details, it predicts non-axisymmetric, locally peaked loss footprints at leading edges and limiters. To test these predictions, we conducted dedicated KSTAR experiments using infrared thermography (IRTV), following the W7-X fast-ion IR validation precedent [6]. Discharges were run in a diverted L-mode (lower single null) held below the L–H transition to suppress ELM- and thermal-ion-driven signals and to steer lost ions into the IRTV field of view; lower Ip and BT were used to enhance the toroidal drift and the prompt-loss signal [5].
Across dedicated shots scanning plasma current (Ip = 400 / 500 / 600 kA), toroidal field (BT = 1.7 / 1.8 T), and beam sources (NB1C, NB2A/B/C), L-mode flat-tops were obtained. Fitted ne and Te profiles from Thomson scattering diagnostics were supplied to NuBDeC, and the simulated prompt-loss heat-flux maps were compared against the beam-on/off IRTV ΔT maps accumulated over 3–7 s window for each shot. For NB1C at fixed BT and power, lowering Ip from 600 to 400 kA enhanced the predicted prompt loss, and the corresponding IRTV ΔT difference shows the localized temperature rise appearing at the predicted leading-edge and poloidal-limiter locations — qualitatively consistent with simulation. The comparison is made under the following assumptions: thermal-ion losses are confined to the divertor region and therefore fall outside the predicted prompt-loss footprints, and the delayed orbit-loss distribution deviates only slightly from the prompt-loss distribution. This first qualitative benchmark confirms that the CAD-based pipeline reproduces where, and how, NBI prompt-loss heat loads concentrate on KSTAR PFCs, establishing a standardized, error-resistant foundation for synthetic diagnostics.
Building on this validated pipeline, we are planning to extend NuBDeC toward an AI-native treatment of engineering detail. The surrogate model can be the future research goal that predicts the prompt-loss heat flux directly on the unstructured PFC mesh, conditioned on the magnetic equilibrium, the beam source, and the as-built geometry. By preserving the full axial asymmetry and engineering-level detail while collapsing per-case cost, the surrogate aims to make exhaustive NBI parameter scans practical during conceptual design of future devices and to deliver between-shot heat-load guidance that helps avoid wasted discharges and fine-tune heating to each experiment’s mission.