Inferring Experimental Transport Parameters Through Machine Learning

In this paper, we present a machine learning model for extracting diffusion and convection coefficients at the plasma edge for non-steady-state density profiles. Extracting particle transport coefficients is important for validating both transport and fueling models at the plasma edge, as the operation of a fusion power plant will occur in regimes that are much more opaque to edge fueling than current devices [1]. A convolutional neural network (CNN) is trained on a simple time-dependent convection-diffusion ansatz model over a wide range of plausible diffusion, convection transport coefficients, as well as ionization source [2]. The NN is tested on synthetic ne and ionization profiles, with and without noise, and retrieves the input parameters within a 2% error. The NN is extended to extract transport coefficients for the transition point from L to H-mode operation in the C-Mod experiments, leveraging both time-dependent electron density and source measurements. To limit the size of the NN, a NN filter is employed to accelerate training and self-learning. This NN will enable us to extract the contributions of diffusive and convective particle fluxes, which can be compared to fundamental theoretical turbulence and neoclassical predications.





[1] S. Mordijick 2020 Nucl. Fusion 60 082006

[2] A.M. Rosenthal et al 2024 Nucl. Fusion 64 036006