Avoiding m/n=2/1 neoclassical tearing modes via machine learning-based gas flow controller

In this work, we present the offline development and performance of a closed-loop, real-time controller designed to avoid onset of the disruptive m/n = 2/1 neoclassical tearing mode (NTM) in low-torque DIII-D plasmas characterized by the ITER normalized parameter set and shape. In these plasmas, onset of the 2/1 NTM typically occurs when the difference between toroidal rotation frequency of the q=1 and q=2 surfaces (△f_1,2) falls below 0.8 kHz. The controller maintains △f_1,2 above this threshold by modulating gas flow into the plasma and feeding back on △f_1,2 using simple proportional-integral-derivative (PID) logic. To enable robust and accurate real-time determination of △f_1,2, an artificial neural network (NN) was trained to predict △f_1,2 as determined by post-shot analysis using real-time charge exchange recombination measurements of the rotation profile. Offline PID gain tuning was conducted using an NN-based surrogate model of the plasma response. The controller produces gas-flow commands similar to those used in prior DIII-D shots where gas was used to maintain stability against 2/1 NTMs [1], demonstrating the controller’s readiness for online implementation and testing.



[1] L. Bardóczi et al 2025 Nucl. Fusion 65 026049