Development of Impurity Transport Coefficient Calculation Algorithm Using Physics-Informed Neural Network

We propose an enhanced algorithm employing physics-informed neural networks (PINNs) to infer impurity transport coefficient profiles in tokamak plasmas. The proposed PINN-based algorithm incorporates the governing equations and related boundary conditions directly into the loss function, eliminating the need for a training process. The algorithm takes the temporal impurity density distribution and seeding rate as input parameters, and outputs the time-independent transport coefficients profiles that reproduce the impurity density distribution. The algorithm was validated through the utilization of three types of argon-seeded H-mode experiments conducted at KSTAR, each with a plasma current of 0.6 MA, a toroidal magnetic field of 2.8 T, and NBI heating power of 4 MW. The calculation algorithm showed rather short computation time, taking less than 10 minutes for each case using a GeForce RTX3090. Performance assessment was conducted using the Pearson correlation coefficient, demonstrating a degree of agreement between the results obtained by the PINN-based algorithm and the reference transport coefficient profiles obtained by the UTC-SANCO code. The results exhibited a correlation coefficient of 98.1% for diffusion coefficient while showed a correlation coefficient of 95.5% for convection velocity. An analysis of impurity transport in krypton seeding experiments with the algorithm will be presented to provide an explanation of the evolution of krypton distribution at KSTAR.