Machine Learning-Based Real-time Plasma Control of DIII-D

Future tokamaks such as ITER will require high-reliability systems for prediction and avoidance of plasma disruptions. Two machine learning algorithms (MLA) have been developed for disruption prediction and avoidance and were tested in real-time. The first MLA is designed to predict the onset of the 2/1 neoclassical tearing mode (NTM), generates a ‘tearability’ that quantifies the likelihood that an NTM will occur. This algorithm was used to command the neutral beam (NB) and electron cyclotron heating (ECH) in real-time on DIII-D in two scenarios: In the first scenario, a ‘tearability’ threshold was chosen above which an NTM was expected to occur. The algorithm was then used to modulate the NB power with the aim of preventing the ‘tearability’ from exceeding its threshold. In the second scenario, the ‘tearability’ was monitored and used to trigger ECH for preemptive NTM suppression when the ‘tearability’ exceeded the threshold. The second MLA generates a ‘disruptivity’ to predict the occurrence of disruptions. During several experiments, this MLA was used to trigger a controlled ramp-down of the plasm current when the ‘disruptivity’ exceeded a user-defined threshold.