Pedestal Structure Prediction using Experimental Data with Machine Learning

Predicting pedestal structure is an essential piece of modeling and controlling the plasma behavior in tokamaks. In the high confinement mode (H-mode), increased edge pressure, the so-called pedestal, leads to better plasma performance, correlating to greater generated fusion power. It has also shown that pedestal electron density is an important factor for the achievable pedestal pressure and pedestal stability with respect to Edge Localized Modes. Machine learning models, in particular neural networks (NN), provide a data-driven way for pedestal predictions and can work quickly during plasma operation, making them useful for real time control. This contribution presents multiple structures of NNs to predict the pedestal shape. They are trained on features drawn from extracted experimental data, including external actuators such as plasma current, toroidal magnetic field, heating power, and gas puff. Each of the NN structures is built on from a fully connected network, with various techniques employed to minimize prediction error. The optimized structures can be incorporated into advanced tokamak control schemes for better stability and higher plasma performance.