Machine Learning Algorithms for Profile Control and Prediction in Tokomaks

Future tokamaks such as ITER will require precise systems for equilibrium prediction to implement control mechanisms. Machine learning algorithms (MLA) have been developed to predict future internal profiles. The algorithms employ neural networks and deep learning methods, and were trained using databases containing several hundred tokomak shots. The algorithms take the internal profiles at time t and predict the profiles at t+Δt using the actuator inputs such as heating power, gas puff, etc. One major area of exploration is to compare the accuracy of predictions between classic deep neural nets and the newer recurrent neural nets. Another major goal is to create a neural net that is accurate, but also efficient enough to provide predictions in real time. With this neural net, it is possible to create controllers to improve tokomak efficiency and optimize plasma performance. It can also be used to gain a deeper understanding of the evolution of a plasma discharge in time.