Investigation of Multi-Z Impurity Transport in Tokamaks using Neural Networks

Abstract

Achieving clean, sustainable energy at scale is a pressing global challenge. Fusion of light elements holds significant potential to address this critical need. While only experimental fusion reactors are currently operational, significant progress is being made in the research and design of near-future tokamak fusion power plants. Reactor success will depend on a comprehensive understanding of heat and particle transport, including the role of impurities. This thesis focuses on the development of machine-agnostic neural network surrogates for TGLF, designed to predict impurity transport coefficients alongside heat and electron particle fluxes in DD plasmas. Training data are derived from synthetic fluxes generated for L, H, and I confinement modes in Alcator C-Mod, DIII-D, and ASDEX-Upgrade. To reduce training complexity, shot data are discretized by radius, and networks are developed at six ρ coordinates: 0.2, 0.4, 0.6, 0.7, 0.8, and 0.9. Fifteen plasma parameters are selected as inputs to the neural networks after examining TGLF flux sensitivities across all five output channels. Predicted impurity fluxes for arbitrary charge states and masses, ranging from 4He to 184W, are used to derive diffusive and convective transport coefficients. Three types of synthetic TGLF data are created and applied to network training to produce accurate models. The primary synthetic data type approximates experimental data by sampling within a perturbation range of ±10% around a given shot. Supporting data types enhance network performance by improving trends in single-parameter (1D) scans and addressing areas of highest network uncertainty. Hyperparameter optimization and testing resulted in highly accurate networks. Testing set relative errors averaged over ρ = 0.4–0.7 and 0.9 show approximate deviations of 0.12 ± 0.029 for heat flux and 0.42 ± 0.095 for particle flux channels. However, error metrics at ρ = 0.2 and 0.8 require location-specific tuning and potentially more data to match the accuracy achieved at other radii. The networks are used to analyze boron and carbon impurity peaking within machinespecific H-modes. Their predictions are then compared to published results. Qualitative results for boron peaking correlations in ASDEX-Upgrade are clearly reproduced, while carbon peaking trends in DIII-D are weaker. Sparse DIII-D data, which also includes atypical advanced modes, is believed to have contributed to reduced accuracy in these cases. Using H-mode shots spanning low to high local collisionality, impurity diffusion trends with charge state (Z) in ITG and TEM dominated plasmas were examined, showing good agreement with published studies. Additionally, analysis of network-derived convective transport shows that Z-sensitivity increases with collisionality. Network scans of the ion and electron heat flux responses to temperature gradients also reveal the clear presence of a critical gradient at all radii. These results demonstrate that the neural networks developed in this work can reliably reproduce TGLF results and deliver fast predictions of heat, electron particle, and impurity transport in tokamaks.